pedestrian dead reckoning navigation with the help of -based routing...

11
Research Article Pedestrian Dead Reckoning Navigation with the Help of A -Based Routing Graphs in Large Unconstrained Spaces F. Taia Alaoui, David Betaille, and Valerie Renaudin IFSTTAR, COSYS, GEOLOC, 44344 Bouguenais, France Correspondence should be addressed to F. Taia Alaoui; [email protected] Received 10 March 2017; Revised 8 May 2017; Accepted 5 June 2017; Published 10 July 2017 Academic Editor: Carlo Fischione Copyright © 2017 F. Taia Alaoui et al. is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. An A -based routing graph is proposed to assist PDR indoor and outdoor navigation with handheld devices. Measurements are provided by inertial and magnetic sensors together with a GNSS receiver. e novelty of this work lies in providing a realistic motion support that mitigates the absence of obstacles and enables the calibration of the PDR model even in large spaces where GNSS signal is unavailable. is motion support is exploited for both predicting positions and updating them using a particle filter. e navigation network is used to correct for the gyro driſt, to adjust the step length model and to assess heading misalignment between the pedestrian’s walking direction and the pointing direction of the handheld device. Several datasets have been tested and results show that the proposed model ensures a seamless transition between outdoor and indoor environments and improves the positioning accuracy. e driſt is almost cancelled thanks to heading correction in contrast with a driſt of 8% for the nonaided PDR approach. e mean error of filtered positions ranges from 3 to 5 m. 1. Introduction Pedestrian Dead Reckoning (PDR) is widely adopted in the field of pedestrian navigation with handheld devices. It is particularly adapted to smartphone-based localization as inertial sensors can be designed in a MEMS (Micro- electromechanical Sensors) technology, enabling them to be embedded in lightweight devices. Unlike GNSS receivers, inertial sensors are especially useful indoors as they allow standalone localization without sky visibility. Yet, due to gyro driſt and step detection limitations, additional information is required to assist the PDR positioning process. For foot- mounted sensors, zero velocity update (ZUPT) calibration is exploited to adjust the positioning parameters by detecting stance phases within the gait cycle (static phase), though this calibration is not possible with handheld devices because of free hand motion and an increased difficulty to detect the stance phase. Outdoors, PDR can still be aided by GNSS [1], but this is not feasible indoors because of signal unavailabil- ity and further measurements are needed. ese could be provided by radio beacons or visual information. e first approach requires infrastructure deployment and training [2], while the second necessitates a camera and further image processing for feature recognition [3]. A third possibility is to constrain the pedestrian’s position using map information. Two main paradigms can be retained from previous work. Either walkable space is given by 2D maps delimited by obstacles [4] or it is given by a routing graph network that transforms the positioning process into a piecewise 1D model [5]. In the first case, space is better explored but the map is not exploited further than for detecting static obstacles (e.g., walls). is means that no calibration is performed unless an obstacle is hit. On the contrary, routing graphs are much more constraining because the motion model is directly given by the graph network. Hence, their use has greater impact on the shape and accuracy of the trajectory and they have to be realistic enough to limit positioning errors. is paper focuses on routing graph-assisted PDR. In fact, routing graphs involve a simple motion model that allows both obstacle avoidance and the calibration of walking directions within straight line travels [5]. eir use can even be extended to calibrating the step length model as reported in our previous work [6], though two major drawbacks make their use quite impractical and sometimes ineffective. Hindawi Wireless Communications and Mobile Computing Volume 2017, Article ID 7951346, 10 pages https://doi.org/10.1155/2017/7951346

Upload: others

Post on 03-Jun-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Pedestrian Dead Reckoning Navigation with the Help of -Based Routing …downloads.hindawi.com/journals/wcmc/2017/7951346.pdf · ResearchArticle Pedestrian Dead Reckoning Navigation

Research ArticlePedestrian Dead Reckoning Navigation with the Help ofAlowast-Based Routing Graphs in Large Unconstrained Spaces

F Taia Alaoui David Betaille and Valerie Renaudin

IFSTTAR COSYS GEOLOC 44344 Bouguenais France

Correspondence should be addressed to F Taia Alaoui fadouataia-alaouiifsttarfr

Received 10 March 2017 Revised 8 May 2017 Accepted 5 June 2017 Published 10 July 2017

Academic Editor Carlo Fischione

Copyright copy 2017 F Taia Alaoui et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

An Alowast-based routing graph is proposed to assist PDR indoor and outdoor navigation with handheld devices Measurements areprovided by inertial and magnetic sensors together with a GNSS receiver The novelty of this work lies in providing a realisticmotion support that mitigates the absence of obstacles and enables the calibration of the PDR model even in large spaces whereGNSS signal is unavailableThis motion support is exploited for both predicting positions and updating them using a particle filterThe navigation network is used to correct for the gyro drift to adjust the step length model and to assess heading misalignmentbetween the pedestrianrsquos walking direction and the pointing direction of the handheld device Several datasets have been tested andresults show that the proposed model ensures a seamless transition between outdoor and indoor environments and improves thepositioning accuracyThe drift is almost cancelled thanks to heading correction in contrast with a drift of 8 for the nonaided PDRapproach The mean error of filtered positions ranges from 3 to 5m

1 Introduction

Pedestrian Dead Reckoning (PDR) is widely adopted inthe field of pedestrian navigation with handheld devicesIt is particularly adapted to smartphone-based localizationas inertial sensors can be designed in a MEMS (Micro-electromechanical Sensors) technology enabling them to beembedded in lightweight devices Unlike GNSS receiversinertial sensors are especially useful indoors as they allowstandalone localization without sky visibility Yet due to gyrodrift and step detection limitations additional informationis required to assist the PDR positioning process For foot-mounted sensors zero velocity update (ZUPT) calibration isexploited to adjust the positioning parameters by detectingstance phases within the gait cycle (static phase) though thiscalibration is not possible with handheld devices because offree hand motion and an increased difficulty to detect thestance phase Outdoors PDR can still be aided by GNSS [1]but this is not feasible indoors because of signal unavailabil-ity and further measurements are needed These could beprovided by radio beacons or visual information The firstapproach requires infrastructure deployment and training

[2] while the second necessitates a camera and further imageprocessing for feature recognition [3] A third possibility isto constrain the pedestrianrsquos position using map informationTwo main paradigms can be retained from previous workEither walkable space is given by 2D maps delimited byobstacles [4] or it is given by a routing graph network thattransforms the positioning process into a piecewise 1Dmodel[5] In the first case space is better explored but the map isnot exploited further than for detecting static obstacles (egwalls) This means that no calibration is performed unless anobstacle is hit On the contrary routing graphs aremuchmoreconstraining because the motion model is directly given bythe graph network Hence their use has greater impact onthe shape and accuracy of the trajectory and they have to berealistic enough to limit positioning errors

This paper focuses on routing graph-assisted PDR Infact routing graphs involve a simple motion model thatallows both obstacle avoidance and the calibration of walkingdirections within straight line travels [5] Their use can evenbe extended to calibrating the step length model as reportedin our previous work [6] though two major drawbacksmake their use quite impractical and sometimes ineffective

HindawiWireless Communications and Mobile ComputingVolume 2017 Article ID 7951346 10 pageshttpsdoiorg10115520177951346

2 Wireless Communications and Mobile Computing

First there is no ubiquitous process that generates combinedindoor- outdoor pedestrian graph networks Their construc-tion can be time-consuming and sometimes inadequatewith pedestrian motion While the graph is expected tocounterbalance PDR limitations it may introduce additionalerrors due to inconsistencies with real displacement Secondthey do not handle ldquopseudorandomrdquo trajectories withinobstacle-free areas and during the transition between out-door and indoor spaces as the freedomofmotion increases Inthese cases routing graphs become inefficient as no motionassumptions can be relied on to design a realistic networkGrid-based models are well-suited for exploring large spaces[7] Yet no calibration is possible with this approach Thiseven leads to overestimating the travel distance becauselinearity of movement is lost due to the grid structure As aresult the PDR process is no longer assisted and the basicissues of gyro drift and uncalibrated travel distance are notsolved

A solution to the vulnerability of routing graph-assistedPDR navigation in obstacle-free spaces is investigated inthis study The proposal is to make use of Alowast algorithmwhich is commonly used for path planning within virtualworlds or for vehicle route guidance in order to design arealistic routing graph in obstacle-free spaces The graph isconstructed on the basis of a set of waypoints that are crucialfor pedestrian navigation In fact the latter are expected tobalance the absence of geometrical constraints by providingstrategic locations between which path computation with Alowastalgorithmwould be relevantTherefore the first contributionof this study is to mitigate the drift in the PDR approach evenin large spaces Moreover the issue of accumulating errorsduring the transition between indoor and outdoor spaces isaddressed by improving the routing graph relevance This isaccomplished through the computation of the likely pathsa pedestrian may take from outdoor strategic locations toreach buildings entrance doors In addition the choice of Alowastalgorithm is well-suited for handling different mobility pro-files (eg personal disease that impacts the path of the targetuser) so that the approach can be customized for the specificneeds of each pedestrian Indeed this is ensured thanks to auser-relative weighting of the map which is directly involvedin the graph construction This is because Alowast algorithmcomputes optimal paths according to walkability rates givenby a map of navigation (eg walkability rate according toslopes pavement) For instance one would take the shortestpath to cross an open area whereas a disabled person mayfollow another path such as walking near walls or dedicatedtracks

2 Alowast Pathfinding within Navigation Meshes

21 Overview of Navigation Mesh Generation A navigationmesh (NavMesh) is a set of several 2D or 3D polygonsreachable by some given user [8] This structure can beobtained automatically using GIS software Others make useof RECAST [9] an open source librarywithin the communityof virtual world designers Several studies were conductedon how to construct navigation meshes from raw 3D models[10 11] NavMesh construction allows assigning weights that

Figure 1Walkable zones have beenmeshed (green polygons) Blankspaces are also walkable but the NavMesh is restricted to our zoneof interest The gray blocks are the stores considered as obstacles

translate the polygon walkability rates according to somegiven characteristics (slope type = stairs flat ground )and to the target userrsquos mobility profile (ability to walk toclimb to take the stairs etc)Theweighting of theNavMesh ispretty important for considering the userrsquos mobility profile inorder to compute suitable routing paths DETOUR [12] is anexample of open source software that allows Alowast pathfindingcalculations on the basis of a NavMesh

22 A Shopping Mall NavMesh Generation The proposedmethod was applied to a shopping mall (Figure 1) The latteris mapped in Google Maps with the names of stores themall principal gates and the set of walkable zones Weused QGIS 2121 software to extract the map in the formof a raster with an 11 cm resolution Walkable zones wereextracted by digitalizing the map and eliminating obstaclesDelaunay triangulation was used to create the NavMesh Allpolygons were assigned the same weight as all experimentsare conducted by healthy persons This implies that Alowastcomputed paths will minimize distance and may result in ashape that can be obtained with different approaches [10 13]

23 Alowast Pathfinding Algorithm Alowast is an improved version ofthe Dijkstra algorithm [14] The latter aims at calculating theoptimal path between two points according to one given costfunction Optimization applies to the travel distance the timeof travel the expended energy and so forthThis adaptabilityallows dealing with the issue of reduced mobility due to adisease or handicapTheAlowast optimization process is a discretesearch scheme where space is modelled by a grid composedof cells Each cell is explored according to an adjacency graphthat models connections within the grid Once visited a cellis stored together with its assigned cost until the target cellis reached The optimal path is given by the sequence of cellsfor which the sum of costs is minimal Alowast is an extensionto the principle of Dijkstra that calculates the cost of visitedcells according to their distance from the starting point(which is the same as the Dijkstra cost) but also to theirassumed distance to the target cell in a heuristic approachThe a priori cost-to-go is assumed to be lower than the actual(unknown) cost from the current cell to the target cell Alowast isfaster than Dijkstra algorithm due to the sorted explorationof cells according to their costs-to-go The Alowast pathfinding

Wireless Communications and Mobile Computing 3

scheme can be applied to any rasterized NavMesh wherepolygon weights are inherited by the raster cells

3 Alowast-Based Routing Graph Generation

31 Routing Graph Generation on the Basis of Waypoints Awaypoint is a punctual element that intervenes significantlyin the pathfinding process The significance of a waypoint isrelated to the role it plays in the pedestrianrsquos decision makingduring herhis travel Two situations are worth consideringThe first is when the pedestrian intends to reach a knowndestination within a familiar environment In this case it isobvious that (s)he follows the itinerary best suited to herhismobility profile Generally it is the shortest one but it couldbe any path depending on the cost function that definesthe userrsquos mobility profile This situation is handled by Alowastalgorithm using a weighted NavMesh The target destinationis known if it is visible from the pedestrianrsquos current positionfor example if the pedestrian is located at a graph node that isvisible from herhis target nodeThe second situation is whenthe pedestrianrsquos destination is invisible due to obstacles Inthis case waypoints are built in order to discretize the possibletarget destinations which may be either final or intermediatedestinations

To better understand the implementation of waypointsspace has to be considered in relation to human spatialcognition Indeed walkable areas within a building floorcould be corridors rooms or halls Outdoors walkable areasare the space comprising sidewalks footpaths squares carparks and so forth (roads belong to the drivable areas) Thisintuitive classification of space components allows definingscenes and extends what is called decision scenes to outdoorspace and corridors Decision scenes have been definedpreviously in [15] as the places that ldquocan be entered and leftand [are] physically bounded by buildings and other solidobstacles that prevent movementrdquo The extension of decisionscenes to corridors and footpaths in this paper is motivatedby the fact that these elements have borders (physical bordersfor corridors and geometrical borders for footpaths) andthat they can be accessed and left through specific pointsFor example two corridors are two separate scenes Theycan be accessed and left through their intersection whichis a building corner a corridor and hall are also distinctscenes between which traffic flow is possible through theirintersection Two elements are important for pedestriantravel inside and between scenes The first is related to theisovist which is visible space from the pedestrianrsquos currentposition The isovist is mainly influenced by the presenceof obstacles that have an impact on both visibility andthe pedestrianrsquos trajectory [16] Second element is the setof portals that allow flows of pedestrians from scene toscene [15] such as corridor extremities building gates storeentries or footpath extremities Indeed key elements in thepedestrian wayfinding behavior depend mainly on portalsand obstacle borders (which determine the isovist) Theymaterialize the fact that visibility and purpose (destination)are the parameters that give a shape to onersquos trajectory Inthis study the set of waypoints is composed of portals (cor-ridor extremities footpath extremities and outdoorindoor

Figure 2 Main area where Alowast algorithm has served for the routinggraph construction

doors) and obstacle corners The graph construction basedon waypoints is obviously more realistic than kernel-basedmethods where scenes are modelled by their centers [17]implying erroneously that the pedestrianwalks systematicallythrough the center of the scene Besides the structure of thegraph is not entirely determined by the geometry of spacebut handles behavioral-based paths generated according towaypoints and reflecting pedestrian travel strategies

Once waypoints have been constructed they can beexported in a database that will assist the pathfinding processEach pair of waypoints are related by a set of itinerariesthat are generated with Alowast algorithm and are part of thefinal routing graph Multihypotheses motion is handled asin classical routing graphs by exploring several paths andkeeping the ones that are best adapted to IMUmeasurementsThe graph structure is stored in the form of a GIS databasethat contains the graph segments identifiers their extremitynodes coordinates their length and their connections withinthe graph in a node-connectivity approach Each edge of thegraph is oriented and has an entry node A and an exit nodeB The node-connectivity design is involved in the motionmodel used for filtering

32 Alowast Path Planning for Seamless Transition betweenOutdoorand Indoor Spaces and within Obstacle-Free Areas Previouswork demonstrated that indoor layout-based graphs areefficient for enhancing the PDR localization within geomet-rically constrained areas [5 6] Hence the proposed methodis applied only to the area that precedes the mall principalentrance as well as in the obstacle-free hall (Figure 2) Thelatter represent the critical GNSS-deprived places where mapinformation fails to provide a calibration to the PDR posi-tioning process emphasizing the main issue being addressedin this paper Outdoors the routing graph is constructedaccording to the geometry of sidewalks crosswalks andfootpaths

Figure 3 shows the generated routing graph within thetest area Three waypoints materialize the building entrancegate (drawn in black in Figure 2) The focus is made onthe waypoint where red paths intersect It represents the leftside of the mall entrance Red paths are the Alowast-calculated

4 Wireless Communications and Mobile Computing

Figure 3 The highlighted segments of the graph in red colorshow the Alowast itineraries relating a portal waypoint (left side of thebuilding entrance door) to different destinations The latter areeither extremities of corridors or footpaths obstacle corners or storeentries

itineraries relating the left side of the gate to other places ofinterest such as stores entries or the buildingrsquosmain corridorsThe latter are modelled by a series of straight lines accordingto corridors main directions Three paths relate the outdoorto the indoor space through the left side of the mall entrance

This representation shows that straight line travels areprivileged for practical and fast displacement They are alsomore realistic in regard to pedestrian motion In fact somehypotheses are eliminated such as walking towards walls ormaking a series of turns to attend a place that can be reachedstraight ahead Moreover this provides a measurement ofwalking directions given by the graph segments in obstacle-free zones which would be impossible if a grid-basedapproach had been applied requiring additional informationsuch as radio beacons signal to compensate for the PDRerrors

4 IMU Fusion with GNSS and the RoutingGraph with a Particle Filter

41 Step Detection and Heading Calculation The step detec-tion is realized after motion classification according to [18] bydetecting peaks on the acceleration signal using an adaptivethreshold algorithm [19] According to the same reference ageneric model is used to estimate the step length This modelrelies on a set of three parameters trained on 10 subjects andis given by

119904 = 119896 sdot (ℎ (119886119891 + 119887) + 119888) (1)

where 119904 is the step length ℎ is the userrsquos height 119891 is the stepfrequency 119886 119887 119888 is the generic model parameters and 119896 isa scale factor that is expected to calibrate the model on thepedestrian

Headings have been calculated with MAGYQ attitudeestimation filter [20] that fuses signals from a triaxisaccelerometer a triaxis gyroscope and a triaxis magnetome-ter Heading calculation considers different carrying modessuch as the swinging mode or the texting mode The deviceorientation in 3D-space is then obtained and the yaw angle

deduced giving the orientation of the device relative to thetrue North

42 Calibration of the PDR Parameters Using GNSS Positionsand the Routing Graph ThePDR parameters to be calibratedare the step length and headingsThe step lengthmodel needsto be adjusted to each user as the model parameters aredependent on the pedestrianrsquos physiological features and gaitcycle whereas headings are potentially misaligned with theactual walking direction because of gyro drift and the devicecarrying mode In fact the device may be oriented towards adirection which is different from the walking directionTheseerrors are compensated by fusing the IMU with GNSS andthe routing graph The fusion is realized thanks to a particlefilter that models the state (vector of unknown variables) bya set of particles (a set of sampled state vectors) The statevector contains necessary variables for determining the userrsquosposition and is presented in detail in Section 43

The step length model is adjusted to the pedestrian usingboth the graph and GNSS positions when available In factthe graph allows keeping the positions on plausible path(s)This directly impacts the travel distance Besides GNSSdecreases the particles dispersion by bringing them next tothe GNSS position On the other hand walking directionsare given by path headings They are particularly reliableindoors as the paths are calculated by the Alowast algorithm or aredirectly given by the corridors main directions The routinggraph-derived walking direction is then compared to theIMU pointing direction and the most likely path is selectedThe difference between both headings gives the IMU angularmisalignment with the pedestrianrsquos walking direction

The step length model is calibrated using a scale factor 119896The latter is assumed constant as it depends mainly onthe pedestrianrsquos physiological parameters It is modelled bya Gaussian variable with 1m mean and low variations asthe step length is restricted to realistic values comprising060mndash1m 119896 is expected to converge within straight linetravels where an optimal calibration is guaranteed

Heading misalignment 120573 is the angular differencebetween the actual walking direction and the pointingdirection of the IMU It is time-dependent because it varieswith handmotion Uniformly distributed samples of headingmisalignment values are drawn with variations up to 30∘

43 Model Implementation with a Particle Filter The statevector is given by

IdDp 120575 119896 120573 119908119901119905 (2)

where Id is the edge identifier in the graph database Dp is thecurvilinear abscissa 120575 represents the travel direction whichcan be 1 if the edge is walked forward (from A to B) and minus1otherwise (fromB toA) 119896 is the scale factor that calibrates thestep lengthmodel120573 is themisalignment between thewalkingdirection and the pointing direction of the handheld device119908 is the particle weight and 119901 and 119905 refer respectively to theparticle and time

Each particle is moved along the current path accordingto the travel distance given by the step length The particle

Wireless Communications and Mobile Computing 5

BA connectivity BB connectivity

$5 gt 0

ltL)>minus1gtL)>minus1

)>t = )>tminus1 )>t = )>tminus1 )>t = )>tminus1

$Jt = $5

t = tminus1 t = tminus1

$Jt = $5 minus L)>minus1$Jt = L)>

minus

t = minustminus1

($5 minus L)>minus1)

Figure 4 Dp computation according to DU

filter introduces white noise on each parameter to spread theparticles and optimize the graph exploration The particleposition is determined by the identifier of the graph edge andits curvilinear abscissa If the latter is above the edge lengthone of its connected edges will be explored The transitionmodel is

DU = (119904) sdot 120575 sdot 120572 + 119899119866 + Dp119905minus1 (3)

where DU is compared to the current edge length 119871 Id tocompute the curvilinear abscissa Dp (Figure 4) 119904 is thestep length and 119899119866 is the modelling error represented by aGaussian noise 120572 isin 0 1 2 and determines the numberof steps to be made by one particle It allows detectingovermisdetected steps

Figure 4 gives the transition test process that allowscomputing the curvilinear abscissa Dp when an exit nodeB is exceeded (ie DU gt 0) 119905 is the time index BABBconnectivity translates the search in the graph database inorder to determine if the edge being explored is connectedto the exceeded one via an A or a B node The same logic isadopted if an entry node A is exceeded (ie DU lt 0)

When GNSS positions are accurate enough they areused together with the graph headings in order to calibrateboth the step length model and PDR headings The graphheadings provide ameasurement of walking directions whichare compared to PDR headings taking into account potentialangular misalignment This is performed according to

119908119905= 119908119905minus1sdot exp(minus12 sdot (

1003816100381610038161003816120579PDR minus (120579Id + 120573)10038161003816100381610038162(120590120579Id2 + 120590120579PDR 2) + Δ1015840Σminus1Δ))

Δ = [119864119875 minus 119864GPS119873119875 minus 119873GPS]

Σminus1 = [[[[[

1(120590119864GPS2 + 120590map

119864

2) 00 1

(120590119873GPS2 + 120590map119873

2)]]]]]

(4)

Figure 5 Ubiquitous Localization Unit with Inertial Sensors andSatellites

where 119908119905 is the particle weight and 119864119875 and 119864GPS arerespectively the predicted East coordinate of a particle and itscorresponding GPS observation 119873119875 and 119873GPS are the sameparameters along the North direction Σminus1 is the weightingmatrix considering the graph and GPS position accuracies

Indoors some GNSS positions are still available but thelatter are generally unreliable and are rejected The rejectiontest is made by thresholding the signal to noise ratio (SNR)and the horizontal dilution of precision (HDOP) Only thegraph headings are exploited in the weighting process Theweighting equation indoors is

119908119905 = 119908119905minus1 sdot exp(minus12 times1003816100381610038161003816120579PDR minus (120579Id + 120573)10038161003816100381610038162(120590120579Id2 + 120590120579PDR 2) ) (5)

where 120579PDR is the walking direction estimated by the PDRalgorithms 120579Id is the predicted heading of the current path120573 is the heading misalignment between the path directionand the PDR estimate 120590120579Id is the standard deviation of thepath heading and 120590120579PDR is the standard deviation of the PDRwalking direction estimate

Once the particles have been weighted some of themare assigned low weights and become useless in the processResampling is performed to duplicate the particles with highweights and delete the others The particles amount is keptconstant and their weights equal after each update in order toexplore enough hypotheses of motion

6 Wireless Communications and Mobile Computing

Figure 6 Data collection by a pedestrian with ULISS unit in hand

5 Experiments

51 Data Collection with a HSGNSS and IMU in Hand Threehealthy volunteers (2 men 1 woman) collected data withULISS (Ubiquitous Localization Unit with Inertial Sensorsand Satellites) device (Figure 5) held in hand (Figure 6) Datawere collected in both outdoor and indoor environments fortwo different device carrying modes Two acquisitions weremade in a textingmode and one in both swinging and textingmodes The average duration of acquisitions was about 14minutes and the walking distance for each trial almost 15 km

Different scenarios were chosen to perform the experi-ments These will be explained in detail in Section 61 wherethe text is enhanced by figures

52 Input Data ULISS device [21] comprises 9 degreesof freedom inertial mobile unit a high sensitivity GNSS(HSGNSS) receiver and antenna a memory card and abattery It delivers measurements that are time-stampedin GPS time Inertial sensors and magnetometers providemeasurements at a 200Hz frequency

TheHSGNSS receiver operates in a standalone mode anddelivers positions in real time at a 5Hz frequency Deliveredpositions are time-stamped in GPS time and have metricaccuracies ranging from 2m up to 10m near buildings andtree shades In this work GNSS positions were interpolated atthe step frequency in order to be fusedwith the PDRestimatesof headings and step lengths

53 Reference Trajectories Besides collecting data withULISS device in hand all volunteers were equipped with anindependent GNSS receiver carried in their backpacks anda small antenna attached to their caps GNSS measurementswere then postprocessed in order to calculate reference trajec-tories by differential GNSS This was performed using RTK-LIB 242p12 software [22]Measurements from the embarkedGNSS receiver and from a nearby base station were used toperform relative GNSS positioning Obtained positions had

0 100 200 300 400East (m)

0

100

200

300N

orth

(m)

PDR trajectoryEstimated trajectory

Differential GNSS trajectory

Figure 7 Estimated trajectory (in red) The blue trajectory givesthe PDR position estimates and the green one gives the referencetrajectory (only texting)

decimetric accuracies up to several meters near buildings andother elevated features Afterwards thresholding was appliedto three parameters in order to reject some outlier positionestimates These parameters are first the number of visiblesatellites second the ratio factor of ambiguity resolutionand finally the horizontal dilution of precisionThe resultingpositions had precisions below 1m and were adequate foraccuracy assessment in this work

6 Results

61 Trajectory Analysis Figures 7 8 and 9 show the esti-mated trajectories (red) with the Alowast-generated routing graphand the particle filter described in Section 43 The blue

Wireless Communications and Mobile Computing 7

0 100 200 300 400East (m)

0

100

200

300

400

Nor

th (m

)

Figure 8 Trajectories for the second dataset (only texting)

0 100 200 300 400 500East (m)

0

100

200

300

400

500

Nor

th (m

)

Figure 9 Trajectories for the third dataset (swinging texting)

pattern corresponds to the PDR trajectory while the greenone is the reference trajectory obtained by differential GNSS

Figures 7 and 8 correspond to acquisitions performed inthe texting modeThe starting point for both acquisitions lieson the top right extremity of the trajectories Both subjectsmade a closed loop around the building with an intermediateoutdoor travel (bottom side of the figures) before reaching thestarting point back

For both acquisitions the travel distance seems to beoverestimated Yet heading determination is more accurate

for the first acquisition as the shape of the trajectory is morefaithful to the building structure The drift is higher for thesecond dataset and can be visually observed at the end of thePDR trajectory

Where the drift is most important the PDR trajectorypresents major inconsistencies with the map According toFigures 7 8 and 9 the drift has been corrected as theestimated trajectories are more compliant with the buildingstructure and with footpaths in outdoor space This has beenachieved thanks to the proposed particle filter and to anincreased conformitywith pedestrianmotion demonstratingthat the positioning accuracy for the texting scenario has beensignificantly enhanced using our approach

Figure 9 corresponds to data collected for both theswinging and texting modes The acquisition started at thebottom of the figure where the reference and the filteredtrajectories overlap Starting from this position the subjectentered the building and then went outside through theNorth-East building entrance This travel was made in theswinging mode The top right extremity of the trajectoryunderlines a U-turn before the subject entered the buildingback to reach the starting point This part of the travel wasperformed in the texting mode

Unlike the two first acquisitions there is a gap in the PDRtrajectory because the subject did not go around the buildingas the two first subjects didThis gap is retrieved in the filteredtrajectory

Obviously this scenario implies a less accurate headingdetermination and even an alteration in the travel distanceestimation In fact thewalking distance is underestimated forthe swinging mode (first part of the travel until the subjectreached the outdoor) and overestimated for the texting mode(second part of the travel until the ending point)These errorscan be noticed in the PDR trajectory

The filtered trajectory shows that the drift has beencorrected resulting in a shape that is more compliant withthe map and with the reference trajectory outdoors Yet thepositioning accuracy seems to be decreased as comparedwith the two first datasets Following section discusses theaccuracy of estimated trajectories

62 Error Computation In order to assess the accuracy ofour positioning method filtered positions were comparedto the reference positions interpolated at the step frequencyThe average plane error ranges from 4 to 5 meters for thethree datasets (Figure 10) Accuracy is dependent on thequality of the PDR trajectory Therefore computed errorsare more important for the second dataset considering thetexting scenario For the third acquisition that includes bothswinging and texting the accuracy is significantly decreasedand more outliers (precisions above 15m) which are givenby the red plus signs in Figure 10 are detected due tomismatching errors (ie choosing wrong edges of the graph)These errors are mainly due to uncalibrated PDR parametersand are discussed in the following sections

63 Heading Misalignment Estimation Figures 11 12 and 13show the estimated heading misalignment values for each

8 Wireless Communications and Mobile Computing

2 31Datasets

0

5

10

15

20

Erro

r (m

)

Figure 10 Plane errors for each dataset

2 4 6 80 10 12Time (min)

minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 11 Full texting scenario (1)

dataset For the first trial (Figure 11) angular misalignmentis comprised between minus10∘ and +10∘ and varies around anapproximate mean value of 0∘ According to this distributionthe angular difference between walking directions and thepointing direction of the device is minimal Hence appliedcorrections compensate only for the gyro driftwhich is ratherlogical regarding the texting mode scenario

For the second dataset (Figure 12) estimated headingmisalignment values are between minus15∘ and +12∘ They arenot equally distributed around 0∘ (eg between the 1st and2nd minutes the mean value is over 5∘) From this analysisnonnegligible hand motion can be assumed even if thesubject intended to perform the experiment in the textingmode

Figure 13 gives the estimated heading misalignment forthe third trial The first part (until 95min) of the travel wasperformed in the swingingmodeHence the estimated valuesvary significantly (minus15∘ to +20∘) For the second part of theplot (gt10min) angular misalignment variations occur withlower magnitudes comprising plusmn5∘ over a mean value of 0∘which reflects the texting mode of the acquisition

64 Step Length Model Calibration In this paper a scalefactor was introduced on the step length in order to calibratethe walking distance though due to degraded GNSS signaland to short walking periods in outdoor space the scale

10 12 14Time (m)

2 4 6 80minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 12 Full texting scenario (2)

minus20minus15minus10minus5

05

101520

Hea

ding

misa

lignm

ent(

∘ )

6 8 141210 16 182 40

Time (min)

Figure 13 Swinging-texting scenario

factor was not calibrated In fact regular and accurate GNSSpositions are needed through straight line travels as cited inour introduction in order to calibrate the walking distanceThese conditions were not verified during our experiments

As a result only corrected headings were relied on in theselection process Therefore distance calibration occurredonly at junctions of the graph when a change in heading wasdetected This explains the fact that the accuracy of filteredtrajectories is still enhanced and compliance with the mapimproved

Though the uncalibrated walking distances caused somemismatching errors because direction change was detectedtoo late or too early in the process this can be noticed in theNorthern part of Figure 9 Indeed while the pedestrian wasintending to exit the building the filtered trajectory indicatesthat he was walking towards a corridor This happened fortwo reasons First the real trajectory is quite unusual interms of pedestrian behavior In fact there is a change inheading (observed in the PDR trajectory) that is independentof space configuration invalidating the assumptions thatallowed constructing our navigation network Second theuncalibrated walking distance prevented the particles fromreaching outdoor space at the right time Another mismatch-ing error occurred at the middle of the building (between300m and 400m North) because direction change wasdetected too late due to uncalibrated walking distance Lateron the particle filter corrected for this error and convergedover the right corridor thanks to the particle dispersion overthe graph and to the multihypothesis approach

Wireless Communications and Mobile Computing 9

7 Conclusions

Amap-aided PDR approach where a routing graph is used asmotion model has been proposed Main contribution of thispaper is Alowast algorithm adaptation to elaborate a pedestriannetwork that is capable of cancelling the gyro drift and themisalignment between the device orientation and thewalkingdirection even in large spaces These are GNSS-deprived andobstacle-free areas where the limitations of map-aided PDRalgorithms are most important In fact widespread map-aided PDR approaches do not compensate for these errorswhen pedestrian motion is unconstrained mainly duringthe transition between outdoor and indoor spaces and whenobstacles are absent The Alowast-based routing graph mitigatesthe lack of obstacles thanks to a set of waypoints implementedaccording to human spatial cognition and to a weightednavigation mesh This allows building a realistic motionmodel thatmeets the requirements ofmap-aided localizationIndeed the proposed routing graph is well exploited becauseit gives prior knowledge about the pedestrianrsquos destinationand provides reliable measurements of walking directionsResults show that it is adequate for a seamless transitionbetween outdoor and indoor environments and for enhanc-ing the positioning accuracy even in large spaces Achievedaccuracies range from 3 to 5 meters and the drift is almostcancelled with the help of the routing graph though somemismatching errors due to uncalibrated walking distanceespecially while carrying the device in the swinging modemight induce important positioning errors Indeed properconditions of sky visibility and sufficient period of outdoorwalking are prerequisite for the step length calibration beforethe pedestrian reaches indoor space

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This researchworkwas done as part of theHappyHand (2015ndash2018) project which is funded by the 19th ldquoFonds UniqueInterministerielrdquo (French national RampD funding)

References

[1] J B Bancroft and G Lachapelle ldquoData fusion algorithms formultiple inertial measurement unitsrdquo Sensors vol 11 no 7 pp6771ndash6798 2011

[2] Y Hu L Sheng and S J Zhang ldquoDesign of Continuous IndoorNavigation System Based on INS and Wifirdquo Applied Mechanicsand Materials vol 303-306 pp 2046ndash2049 2013

[3] T Lee B Shin J H Lee et al ldquoAn Indoor Positioning SystemUsing Vision aided Advanced PDR technology without imageDB andwithmotion recognitionrdquo inProceedings of the ION2013Pacific PNT Meeting pp 490ndash497

[4] P Hafner T Moder M Wieser and T Bernoulli ldquoEvaluationof smartphone-based indoor positioning using different Bayesfiltersrdquo in Proceedings of the 2013 International Conference on

Indoor Positioning and Indoor Navigation IPIN 2013 October2013

[5] T Willemsen F Keller and H Sternberg ldquoA topologicalapproach with MEMS in smartphones based on routing-graphrdquo in Proceedings of the International Conference on IndoorPositioning and Indoor Navigation IPIN 2015 October 2015

[6] F T Alaoui D Betaille and V Renaudin ldquoA multi-hypothesisparticle filtering approach for pedestrian dead reckoningrdquo inProceedings of the 2016 International Conference on IndoorPositioning and Indoor Navigation IPIN 2016 October 2016

[7] T Fetzer F Ebner F Deinzer L Koping and M GrzegorzekldquoOn Monte Carlo smoothing in multi sensor indoor localisa-tionrdquo in Proceedings of the 2016 International Conference onIndoor Positioning and Indoor Navigation IPIN 2016 October2016

[8] M R F Mendonca H S Bernardino and R F Neto ldquoStealthypath planning using navigation meshesrdquo in Proceedings of the4th Brazilian Conference on Intelligent Systems BRACIS 2015pp 31ndash36 bra November 2015

[9] ldquoRecast Detail API documentation for the members declaredin Recasth sdot recastnavigationrecastnavigation6f5c9f9rdquoGitHub httpsgithubcomrecastnavigationcommit6f5c9f9-b82418efc44b85974e604095b95354ada

[10] I Afyouni C Ray and C Claramunt ldquoSpatial models forcontext-aware indoor navigation systems a surveyrdquo Journal ofSpatial Information Science vol 4 no 1 pp 85ndash123 2012

[11] W Van Toll A F Cook and R Geraerts ldquoNavigation meshesfor realistic multi-layered environmentsrdquo in Proceedings of the2011 IEEERSJ International Conference on Intelligent Robots andSystems Celebrating 50Years of Robotics IROSrsquo11 pp 3526ndash3532September 2011

[12] DETOUR ldquoGitHubrdquo httpsgithubcomrecastnavigationreca-stnavigationtreemasterDetourSource 2017

[13] F Mortari S Zlatanova L Liu and E Clementini ldquoImprovedgeometric network model (IGNM) a novel approach for deriv-ing connectivity graphs for indoor navigationrdquo ISPRS Annalsof Photogrammetry Remote Sensing and Spatial InformationSciences vol II-4 pp 45ndash51 2014

[14] N O Eraghi F Lopez-Colino A De Castro and J GarridoldquoPath length comparison in grid maps of planning algorithmsHCTNav A and Dijkstrardquo Design of Circuits and IntegratedSystems pp 1ndash6 2014

[15] C Gaisbauer and A U Frank ldquoWayfinding Model For Pedes-trian Navigationrdquo in Proceedings of the 11th AGILE InternationalConference on Geographic Information Science Spain 2008

[16] N Victor evaluation des deplacements pietons quotidiens Uni-versite de Lyon 2016

[17] L Yang and M Worboys ldquoGeneration of navigation graphs forindoor spacerdquo International Journal of Geographical InformationScience vol 29 no 10 pp 1737ndash1756 2015

[18] M Susi V Renaudin and G Lachapelle ldquoMotion moderecognition and step detection algorithms for mobile phoneusersrdquo Sensors vol 13 no 2 pp 1539ndash1562 2013

[19] V RenaudinM Susi andG Lachapelle ldquoStep length estimationusing handheld inertial sensorsrdquo Sensors (Switzerland) vol 12no 7 pp 8507ndash8525 2012

[20] V Renaudin and C Combettes ldquoMagnetic acceleration fieldsand gyroscope quaternion (MAGYQ)-based attitude estimationwith smartphone sensors for indoor pedestrian navigationrdquoSensors (Switzerland) vol 14 no 12 pp 22864ndash22890 2014

10 Wireless Communications and Mobile Computing

[21] ULISS httpwwwifsttar-geolocfrindexphpenequipment44-uliss

[22] ldquoGitHub - tomojitakasuRTKLIB_binrdquo httpsgithubcomto-mojitakasuRTKLIB_bin 2017

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Page 2: Pedestrian Dead Reckoning Navigation with the Help of -Based Routing …downloads.hindawi.com/journals/wcmc/2017/7951346.pdf · ResearchArticle Pedestrian Dead Reckoning Navigation

2 Wireless Communications and Mobile Computing

First there is no ubiquitous process that generates combinedindoor- outdoor pedestrian graph networks Their construc-tion can be time-consuming and sometimes inadequatewith pedestrian motion While the graph is expected tocounterbalance PDR limitations it may introduce additionalerrors due to inconsistencies with real displacement Secondthey do not handle ldquopseudorandomrdquo trajectories withinobstacle-free areas and during the transition between out-door and indoor spaces as the freedomofmotion increases Inthese cases routing graphs become inefficient as no motionassumptions can be relied on to design a realistic networkGrid-based models are well-suited for exploring large spaces[7] Yet no calibration is possible with this approach Thiseven leads to overestimating the travel distance becauselinearity of movement is lost due to the grid structure As aresult the PDR process is no longer assisted and the basicissues of gyro drift and uncalibrated travel distance are notsolved

A solution to the vulnerability of routing graph-assistedPDR navigation in obstacle-free spaces is investigated inthis study The proposal is to make use of Alowast algorithmwhich is commonly used for path planning within virtualworlds or for vehicle route guidance in order to design arealistic routing graph in obstacle-free spaces The graph isconstructed on the basis of a set of waypoints that are crucialfor pedestrian navigation In fact the latter are expected tobalance the absence of geometrical constraints by providingstrategic locations between which path computation with Alowastalgorithmwould be relevantTherefore the first contributionof this study is to mitigate the drift in the PDR approach evenin large spaces Moreover the issue of accumulating errorsduring the transition between indoor and outdoor spaces isaddressed by improving the routing graph relevance This isaccomplished through the computation of the likely pathsa pedestrian may take from outdoor strategic locations toreach buildings entrance doors In addition the choice of Alowastalgorithm is well-suited for handling different mobility pro-files (eg personal disease that impacts the path of the targetuser) so that the approach can be customized for the specificneeds of each pedestrian Indeed this is ensured thanks to auser-relative weighting of the map which is directly involvedin the graph construction This is because Alowast algorithmcomputes optimal paths according to walkability rates givenby a map of navigation (eg walkability rate according toslopes pavement) For instance one would take the shortestpath to cross an open area whereas a disabled person mayfollow another path such as walking near walls or dedicatedtracks

2 Alowast Pathfinding within Navigation Meshes

21 Overview of Navigation Mesh Generation A navigationmesh (NavMesh) is a set of several 2D or 3D polygonsreachable by some given user [8] This structure can beobtained automatically using GIS software Others make useof RECAST [9] an open source librarywithin the communityof virtual world designers Several studies were conductedon how to construct navigation meshes from raw 3D models[10 11] NavMesh construction allows assigning weights that

Figure 1Walkable zones have beenmeshed (green polygons) Blankspaces are also walkable but the NavMesh is restricted to our zoneof interest The gray blocks are the stores considered as obstacles

translate the polygon walkability rates according to somegiven characteristics (slope type = stairs flat ground )and to the target userrsquos mobility profile (ability to walk toclimb to take the stairs etc)Theweighting of theNavMesh ispretty important for considering the userrsquos mobility profile inorder to compute suitable routing paths DETOUR [12] is anexample of open source software that allows Alowast pathfindingcalculations on the basis of a NavMesh

22 A Shopping Mall NavMesh Generation The proposedmethod was applied to a shopping mall (Figure 1) The latteris mapped in Google Maps with the names of stores themall principal gates and the set of walkable zones Weused QGIS 2121 software to extract the map in the formof a raster with an 11 cm resolution Walkable zones wereextracted by digitalizing the map and eliminating obstaclesDelaunay triangulation was used to create the NavMesh Allpolygons were assigned the same weight as all experimentsare conducted by healthy persons This implies that Alowastcomputed paths will minimize distance and may result in ashape that can be obtained with different approaches [10 13]

23 Alowast Pathfinding Algorithm Alowast is an improved version ofthe Dijkstra algorithm [14] The latter aims at calculating theoptimal path between two points according to one given costfunction Optimization applies to the travel distance the timeof travel the expended energy and so forthThis adaptabilityallows dealing with the issue of reduced mobility due to adisease or handicapTheAlowast optimization process is a discretesearch scheme where space is modelled by a grid composedof cells Each cell is explored according to an adjacency graphthat models connections within the grid Once visited a cellis stored together with its assigned cost until the target cellis reached The optimal path is given by the sequence of cellsfor which the sum of costs is minimal Alowast is an extensionto the principle of Dijkstra that calculates the cost of visitedcells according to their distance from the starting point(which is the same as the Dijkstra cost) but also to theirassumed distance to the target cell in a heuristic approachThe a priori cost-to-go is assumed to be lower than the actual(unknown) cost from the current cell to the target cell Alowast isfaster than Dijkstra algorithm due to the sorted explorationof cells according to their costs-to-go The Alowast pathfinding

Wireless Communications and Mobile Computing 3

scheme can be applied to any rasterized NavMesh wherepolygon weights are inherited by the raster cells

3 Alowast-Based Routing Graph Generation

31 Routing Graph Generation on the Basis of Waypoints Awaypoint is a punctual element that intervenes significantlyin the pathfinding process The significance of a waypoint isrelated to the role it plays in the pedestrianrsquos decision makingduring herhis travel Two situations are worth consideringThe first is when the pedestrian intends to reach a knowndestination within a familiar environment In this case it isobvious that (s)he follows the itinerary best suited to herhismobility profile Generally it is the shortest one but it couldbe any path depending on the cost function that definesthe userrsquos mobility profile This situation is handled by Alowastalgorithm using a weighted NavMesh The target destinationis known if it is visible from the pedestrianrsquos current positionfor example if the pedestrian is located at a graph node that isvisible from herhis target nodeThe second situation is whenthe pedestrianrsquos destination is invisible due to obstacles Inthis case waypoints are built in order to discretize the possibletarget destinations which may be either final or intermediatedestinations

To better understand the implementation of waypointsspace has to be considered in relation to human spatialcognition Indeed walkable areas within a building floorcould be corridors rooms or halls Outdoors walkable areasare the space comprising sidewalks footpaths squares carparks and so forth (roads belong to the drivable areas) Thisintuitive classification of space components allows definingscenes and extends what is called decision scenes to outdoorspace and corridors Decision scenes have been definedpreviously in [15] as the places that ldquocan be entered and leftand [are] physically bounded by buildings and other solidobstacles that prevent movementrdquo The extension of decisionscenes to corridors and footpaths in this paper is motivatedby the fact that these elements have borders (physical bordersfor corridors and geometrical borders for footpaths) andthat they can be accessed and left through specific pointsFor example two corridors are two separate scenes Theycan be accessed and left through their intersection whichis a building corner a corridor and hall are also distinctscenes between which traffic flow is possible through theirintersection Two elements are important for pedestriantravel inside and between scenes The first is related to theisovist which is visible space from the pedestrianrsquos currentposition The isovist is mainly influenced by the presenceof obstacles that have an impact on both visibility andthe pedestrianrsquos trajectory [16] Second element is the setof portals that allow flows of pedestrians from scene toscene [15] such as corridor extremities building gates storeentries or footpath extremities Indeed key elements in thepedestrian wayfinding behavior depend mainly on portalsand obstacle borders (which determine the isovist) Theymaterialize the fact that visibility and purpose (destination)are the parameters that give a shape to onersquos trajectory Inthis study the set of waypoints is composed of portals (cor-ridor extremities footpath extremities and outdoorindoor

Figure 2 Main area where Alowast algorithm has served for the routinggraph construction

doors) and obstacle corners The graph construction basedon waypoints is obviously more realistic than kernel-basedmethods where scenes are modelled by their centers [17]implying erroneously that the pedestrianwalks systematicallythrough the center of the scene Besides the structure of thegraph is not entirely determined by the geometry of spacebut handles behavioral-based paths generated according towaypoints and reflecting pedestrian travel strategies

Once waypoints have been constructed they can beexported in a database that will assist the pathfinding processEach pair of waypoints are related by a set of itinerariesthat are generated with Alowast algorithm and are part of thefinal routing graph Multihypotheses motion is handled asin classical routing graphs by exploring several paths andkeeping the ones that are best adapted to IMUmeasurementsThe graph structure is stored in the form of a GIS databasethat contains the graph segments identifiers their extremitynodes coordinates their length and their connections withinthe graph in a node-connectivity approach Each edge of thegraph is oriented and has an entry node A and an exit nodeB The node-connectivity design is involved in the motionmodel used for filtering

32 Alowast Path Planning for Seamless Transition betweenOutdoorand Indoor Spaces and within Obstacle-Free Areas Previouswork demonstrated that indoor layout-based graphs areefficient for enhancing the PDR localization within geomet-rically constrained areas [5 6] Hence the proposed methodis applied only to the area that precedes the mall principalentrance as well as in the obstacle-free hall (Figure 2) Thelatter represent the critical GNSS-deprived places where mapinformation fails to provide a calibration to the PDR posi-tioning process emphasizing the main issue being addressedin this paper Outdoors the routing graph is constructedaccording to the geometry of sidewalks crosswalks andfootpaths

Figure 3 shows the generated routing graph within thetest area Three waypoints materialize the building entrancegate (drawn in black in Figure 2) The focus is made onthe waypoint where red paths intersect It represents the leftside of the mall entrance Red paths are the Alowast-calculated

4 Wireless Communications and Mobile Computing

Figure 3 The highlighted segments of the graph in red colorshow the Alowast itineraries relating a portal waypoint (left side of thebuilding entrance door) to different destinations The latter areeither extremities of corridors or footpaths obstacle corners or storeentries

itineraries relating the left side of the gate to other places ofinterest such as stores entries or the buildingrsquosmain corridorsThe latter are modelled by a series of straight lines accordingto corridors main directions Three paths relate the outdoorto the indoor space through the left side of the mall entrance

This representation shows that straight line travels areprivileged for practical and fast displacement They are alsomore realistic in regard to pedestrian motion In fact somehypotheses are eliminated such as walking towards walls ormaking a series of turns to attend a place that can be reachedstraight ahead Moreover this provides a measurement ofwalking directions given by the graph segments in obstacle-free zones which would be impossible if a grid-basedapproach had been applied requiring additional informationsuch as radio beacons signal to compensate for the PDRerrors

4 IMU Fusion with GNSS and the RoutingGraph with a Particle Filter

41 Step Detection and Heading Calculation The step detec-tion is realized after motion classification according to [18] bydetecting peaks on the acceleration signal using an adaptivethreshold algorithm [19] According to the same reference ageneric model is used to estimate the step length This modelrelies on a set of three parameters trained on 10 subjects andis given by

119904 = 119896 sdot (ℎ (119886119891 + 119887) + 119888) (1)

where 119904 is the step length ℎ is the userrsquos height 119891 is the stepfrequency 119886 119887 119888 is the generic model parameters and 119896 isa scale factor that is expected to calibrate the model on thepedestrian

Headings have been calculated with MAGYQ attitudeestimation filter [20] that fuses signals from a triaxisaccelerometer a triaxis gyroscope and a triaxis magnetome-ter Heading calculation considers different carrying modessuch as the swinging mode or the texting mode The deviceorientation in 3D-space is then obtained and the yaw angle

deduced giving the orientation of the device relative to thetrue North

42 Calibration of the PDR Parameters Using GNSS Positionsand the Routing Graph ThePDR parameters to be calibratedare the step length and headingsThe step lengthmodel needsto be adjusted to each user as the model parameters aredependent on the pedestrianrsquos physiological features and gaitcycle whereas headings are potentially misaligned with theactual walking direction because of gyro drift and the devicecarrying mode In fact the device may be oriented towards adirection which is different from the walking directionTheseerrors are compensated by fusing the IMU with GNSS andthe routing graph The fusion is realized thanks to a particlefilter that models the state (vector of unknown variables) bya set of particles (a set of sampled state vectors) The statevector contains necessary variables for determining the userrsquosposition and is presented in detail in Section 43

The step length model is adjusted to the pedestrian usingboth the graph and GNSS positions when available In factthe graph allows keeping the positions on plausible path(s)This directly impacts the travel distance Besides GNSSdecreases the particles dispersion by bringing them next tothe GNSS position On the other hand walking directionsare given by path headings They are particularly reliableindoors as the paths are calculated by the Alowast algorithm or aredirectly given by the corridors main directions The routinggraph-derived walking direction is then compared to theIMU pointing direction and the most likely path is selectedThe difference between both headings gives the IMU angularmisalignment with the pedestrianrsquos walking direction

The step length model is calibrated using a scale factor 119896The latter is assumed constant as it depends mainly onthe pedestrianrsquos physiological parameters It is modelled bya Gaussian variable with 1m mean and low variations asthe step length is restricted to realistic values comprising060mndash1m 119896 is expected to converge within straight linetravels where an optimal calibration is guaranteed

Heading misalignment 120573 is the angular differencebetween the actual walking direction and the pointingdirection of the IMU It is time-dependent because it varieswith handmotion Uniformly distributed samples of headingmisalignment values are drawn with variations up to 30∘

43 Model Implementation with a Particle Filter The statevector is given by

IdDp 120575 119896 120573 119908119901119905 (2)

where Id is the edge identifier in the graph database Dp is thecurvilinear abscissa 120575 represents the travel direction whichcan be 1 if the edge is walked forward (from A to B) and minus1otherwise (fromB toA) 119896 is the scale factor that calibrates thestep lengthmodel120573 is themisalignment between thewalkingdirection and the pointing direction of the handheld device119908 is the particle weight and 119901 and 119905 refer respectively to theparticle and time

Each particle is moved along the current path accordingto the travel distance given by the step length The particle

Wireless Communications and Mobile Computing 5

BA connectivity BB connectivity

$5 gt 0

ltL)>minus1gtL)>minus1

)>t = )>tminus1 )>t = )>tminus1 )>t = )>tminus1

$Jt = $5

t = tminus1 t = tminus1

$Jt = $5 minus L)>minus1$Jt = L)>

minus

t = minustminus1

($5 minus L)>minus1)

Figure 4 Dp computation according to DU

filter introduces white noise on each parameter to spread theparticles and optimize the graph exploration The particleposition is determined by the identifier of the graph edge andits curvilinear abscissa If the latter is above the edge lengthone of its connected edges will be explored The transitionmodel is

DU = (119904) sdot 120575 sdot 120572 + 119899119866 + Dp119905minus1 (3)

where DU is compared to the current edge length 119871 Id tocompute the curvilinear abscissa Dp (Figure 4) 119904 is thestep length and 119899119866 is the modelling error represented by aGaussian noise 120572 isin 0 1 2 and determines the numberof steps to be made by one particle It allows detectingovermisdetected steps

Figure 4 gives the transition test process that allowscomputing the curvilinear abscissa Dp when an exit nodeB is exceeded (ie DU gt 0) 119905 is the time index BABBconnectivity translates the search in the graph database inorder to determine if the edge being explored is connectedto the exceeded one via an A or a B node The same logic isadopted if an entry node A is exceeded (ie DU lt 0)

When GNSS positions are accurate enough they areused together with the graph headings in order to calibrateboth the step length model and PDR headings The graphheadings provide ameasurement of walking directions whichare compared to PDR headings taking into account potentialangular misalignment This is performed according to

119908119905= 119908119905minus1sdot exp(minus12 sdot (

1003816100381610038161003816120579PDR minus (120579Id + 120573)10038161003816100381610038162(120590120579Id2 + 120590120579PDR 2) + Δ1015840Σminus1Δ))

Δ = [119864119875 minus 119864GPS119873119875 minus 119873GPS]

Σminus1 = [[[[[

1(120590119864GPS2 + 120590map

119864

2) 00 1

(120590119873GPS2 + 120590map119873

2)]]]]]

(4)

Figure 5 Ubiquitous Localization Unit with Inertial Sensors andSatellites

where 119908119905 is the particle weight and 119864119875 and 119864GPS arerespectively the predicted East coordinate of a particle and itscorresponding GPS observation 119873119875 and 119873GPS are the sameparameters along the North direction Σminus1 is the weightingmatrix considering the graph and GPS position accuracies

Indoors some GNSS positions are still available but thelatter are generally unreliable and are rejected The rejectiontest is made by thresholding the signal to noise ratio (SNR)and the horizontal dilution of precision (HDOP) Only thegraph headings are exploited in the weighting process Theweighting equation indoors is

119908119905 = 119908119905minus1 sdot exp(minus12 times1003816100381610038161003816120579PDR minus (120579Id + 120573)10038161003816100381610038162(120590120579Id2 + 120590120579PDR 2) ) (5)

where 120579PDR is the walking direction estimated by the PDRalgorithms 120579Id is the predicted heading of the current path120573 is the heading misalignment between the path directionand the PDR estimate 120590120579Id is the standard deviation of thepath heading and 120590120579PDR is the standard deviation of the PDRwalking direction estimate

Once the particles have been weighted some of themare assigned low weights and become useless in the processResampling is performed to duplicate the particles with highweights and delete the others The particles amount is keptconstant and their weights equal after each update in order toexplore enough hypotheses of motion

6 Wireless Communications and Mobile Computing

Figure 6 Data collection by a pedestrian with ULISS unit in hand

5 Experiments

51 Data Collection with a HSGNSS and IMU in Hand Threehealthy volunteers (2 men 1 woman) collected data withULISS (Ubiquitous Localization Unit with Inertial Sensorsand Satellites) device (Figure 5) held in hand (Figure 6) Datawere collected in both outdoor and indoor environments fortwo different device carrying modes Two acquisitions weremade in a textingmode and one in both swinging and textingmodes The average duration of acquisitions was about 14minutes and the walking distance for each trial almost 15 km

Different scenarios were chosen to perform the experi-ments These will be explained in detail in Section 61 wherethe text is enhanced by figures

52 Input Data ULISS device [21] comprises 9 degreesof freedom inertial mobile unit a high sensitivity GNSS(HSGNSS) receiver and antenna a memory card and abattery It delivers measurements that are time-stampedin GPS time Inertial sensors and magnetometers providemeasurements at a 200Hz frequency

TheHSGNSS receiver operates in a standalone mode anddelivers positions in real time at a 5Hz frequency Deliveredpositions are time-stamped in GPS time and have metricaccuracies ranging from 2m up to 10m near buildings andtree shades In this work GNSS positions were interpolated atthe step frequency in order to be fusedwith the PDRestimatesof headings and step lengths

53 Reference Trajectories Besides collecting data withULISS device in hand all volunteers were equipped with anindependent GNSS receiver carried in their backpacks anda small antenna attached to their caps GNSS measurementswere then postprocessed in order to calculate reference trajec-tories by differential GNSS This was performed using RTK-LIB 242p12 software [22]Measurements from the embarkedGNSS receiver and from a nearby base station were used toperform relative GNSS positioning Obtained positions had

0 100 200 300 400East (m)

0

100

200

300N

orth

(m)

PDR trajectoryEstimated trajectory

Differential GNSS trajectory

Figure 7 Estimated trajectory (in red) The blue trajectory givesthe PDR position estimates and the green one gives the referencetrajectory (only texting)

decimetric accuracies up to several meters near buildings andother elevated features Afterwards thresholding was appliedto three parameters in order to reject some outlier positionestimates These parameters are first the number of visiblesatellites second the ratio factor of ambiguity resolutionand finally the horizontal dilution of precisionThe resultingpositions had precisions below 1m and were adequate foraccuracy assessment in this work

6 Results

61 Trajectory Analysis Figures 7 8 and 9 show the esti-mated trajectories (red) with the Alowast-generated routing graphand the particle filter described in Section 43 The blue

Wireless Communications and Mobile Computing 7

0 100 200 300 400East (m)

0

100

200

300

400

Nor

th (m

)

Figure 8 Trajectories for the second dataset (only texting)

0 100 200 300 400 500East (m)

0

100

200

300

400

500

Nor

th (m

)

Figure 9 Trajectories for the third dataset (swinging texting)

pattern corresponds to the PDR trajectory while the greenone is the reference trajectory obtained by differential GNSS

Figures 7 and 8 correspond to acquisitions performed inthe texting modeThe starting point for both acquisitions lieson the top right extremity of the trajectories Both subjectsmade a closed loop around the building with an intermediateoutdoor travel (bottom side of the figures) before reaching thestarting point back

For both acquisitions the travel distance seems to beoverestimated Yet heading determination is more accurate

for the first acquisition as the shape of the trajectory is morefaithful to the building structure The drift is higher for thesecond dataset and can be visually observed at the end of thePDR trajectory

Where the drift is most important the PDR trajectorypresents major inconsistencies with the map According toFigures 7 8 and 9 the drift has been corrected as theestimated trajectories are more compliant with the buildingstructure and with footpaths in outdoor space This has beenachieved thanks to the proposed particle filter and to anincreased conformitywith pedestrianmotion demonstratingthat the positioning accuracy for the texting scenario has beensignificantly enhanced using our approach

Figure 9 corresponds to data collected for both theswinging and texting modes The acquisition started at thebottom of the figure where the reference and the filteredtrajectories overlap Starting from this position the subjectentered the building and then went outside through theNorth-East building entrance This travel was made in theswinging mode The top right extremity of the trajectoryunderlines a U-turn before the subject entered the buildingback to reach the starting point This part of the travel wasperformed in the texting mode

Unlike the two first acquisitions there is a gap in the PDRtrajectory because the subject did not go around the buildingas the two first subjects didThis gap is retrieved in the filteredtrajectory

Obviously this scenario implies a less accurate headingdetermination and even an alteration in the travel distanceestimation In fact thewalking distance is underestimated forthe swinging mode (first part of the travel until the subjectreached the outdoor) and overestimated for the texting mode(second part of the travel until the ending point)These errorscan be noticed in the PDR trajectory

The filtered trajectory shows that the drift has beencorrected resulting in a shape that is more compliant withthe map and with the reference trajectory outdoors Yet thepositioning accuracy seems to be decreased as comparedwith the two first datasets Following section discusses theaccuracy of estimated trajectories

62 Error Computation In order to assess the accuracy ofour positioning method filtered positions were comparedto the reference positions interpolated at the step frequencyThe average plane error ranges from 4 to 5 meters for thethree datasets (Figure 10) Accuracy is dependent on thequality of the PDR trajectory Therefore computed errorsare more important for the second dataset considering thetexting scenario For the third acquisition that includes bothswinging and texting the accuracy is significantly decreasedand more outliers (precisions above 15m) which are givenby the red plus signs in Figure 10 are detected due tomismatching errors (ie choosing wrong edges of the graph)These errors are mainly due to uncalibrated PDR parametersand are discussed in the following sections

63 Heading Misalignment Estimation Figures 11 12 and 13show the estimated heading misalignment values for each

8 Wireless Communications and Mobile Computing

2 31Datasets

0

5

10

15

20

Erro

r (m

)

Figure 10 Plane errors for each dataset

2 4 6 80 10 12Time (min)

minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 11 Full texting scenario (1)

dataset For the first trial (Figure 11) angular misalignmentis comprised between minus10∘ and +10∘ and varies around anapproximate mean value of 0∘ According to this distributionthe angular difference between walking directions and thepointing direction of the device is minimal Hence appliedcorrections compensate only for the gyro driftwhich is ratherlogical regarding the texting mode scenario

For the second dataset (Figure 12) estimated headingmisalignment values are between minus15∘ and +12∘ They arenot equally distributed around 0∘ (eg between the 1st and2nd minutes the mean value is over 5∘) From this analysisnonnegligible hand motion can be assumed even if thesubject intended to perform the experiment in the textingmode

Figure 13 gives the estimated heading misalignment forthe third trial The first part (until 95min) of the travel wasperformed in the swingingmodeHence the estimated valuesvary significantly (minus15∘ to +20∘) For the second part of theplot (gt10min) angular misalignment variations occur withlower magnitudes comprising plusmn5∘ over a mean value of 0∘which reflects the texting mode of the acquisition

64 Step Length Model Calibration In this paper a scalefactor was introduced on the step length in order to calibratethe walking distance though due to degraded GNSS signaland to short walking periods in outdoor space the scale

10 12 14Time (m)

2 4 6 80minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 12 Full texting scenario (2)

minus20minus15minus10minus5

05

101520

Hea

ding

misa

lignm

ent(

∘ )

6 8 141210 16 182 40

Time (min)

Figure 13 Swinging-texting scenario

factor was not calibrated In fact regular and accurate GNSSpositions are needed through straight line travels as cited inour introduction in order to calibrate the walking distanceThese conditions were not verified during our experiments

As a result only corrected headings were relied on in theselection process Therefore distance calibration occurredonly at junctions of the graph when a change in heading wasdetected This explains the fact that the accuracy of filteredtrajectories is still enhanced and compliance with the mapimproved

Though the uncalibrated walking distances caused somemismatching errors because direction change was detectedtoo late or too early in the process this can be noticed in theNorthern part of Figure 9 Indeed while the pedestrian wasintending to exit the building the filtered trajectory indicatesthat he was walking towards a corridor This happened fortwo reasons First the real trajectory is quite unusual interms of pedestrian behavior In fact there is a change inheading (observed in the PDR trajectory) that is independentof space configuration invalidating the assumptions thatallowed constructing our navigation network Second theuncalibrated walking distance prevented the particles fromreaching outdoor space at the right time Another mismatch-ing error occurred at the middle of the building (between300m and 400m North) because direction change wasdetected too late due to uncalibrated walking distance Lateron the particle filter corrected for this error and convergedover the right corridor thanks to the particle dispersion overthe graph and to the multihypothesis approach

Wireless Communications and Mobile Computing 9

7 Conclusions

Amap-aided PDR approach where a routing graph is used asmotion model has been proposed Main contribution of thispaper is Alowast algorithm adaptation to elaborate a pedestriannetwork that is capable of cancelling the gyro drift and themisalignment between the device orientation and thewalkingdirection even in large spaces These are GNSS-deprived andobstacle-free areas where the limitations of map-aided PDRalgorithms are most important In fact widespread map-aided PDR approaches do not compensate for these errorswhen pedestrian motion is unconstrained mainly duringthe transition between outdoor and indoor spaces and whenobstacles are absent The Alowast-based routing graph mitigatesthe lack of obstacles thanks to a set of waypoints implementedaccording to human spatial cognition and to a weightednavigation mesh This allows building a realistic motionmodel thatmeets the requirements ofmap-aided localizationIndeed the proposed routing graph is well exploited becauseit gives prior knowledge about the pedestrianrsquos destinationand provides reliable measurements of walking directionsResults show that it is adequate for a seamless transitionbetween outdoor and indoor environments and for enhanc-ing the positioning accuracy even in large spaces Achievedaccuracies range from 3 to 5 meters and the drift is almostcancelled with the help of the routing graph though somemismatching errors due to uncalibrated walking distanceespecially while carrying the device in the swinging modemight induce important positioning errors Indeed properconditions of sky visibility and sufficient period of outdoorwalking are prerequisite for the step length calibration beforethe pedestrian reaches indoor space

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This researchworkwas done as part of theHappyHand (2015ndash2018) project which is funded by the 19th ldquoFonds UniqueInterministerielrdquo (French national RampD funding)

References

[1] J B Bancroft and G Lachapelle ldquoData fusion algorithms formultiple inertial measurement unitsrdquo Sensors vol 11 no 7 pp6771ndash6798 2011

[2] Y Hu L Sheng and S J Zhang ldquoDesign of Continuous IndoorNavigation System Based on INS and Wifirdquo Applied Mechanicsand Materials vol 303-306 pp 2046ndash2049 2013

[3] T Lee B Shin J H Lee et al ldquoAn Indoor Positioning SystemUsing Vision aided Advanced PDR technology without imageDB andwithmotion recognitionrdquo inProceedings of the ION2013Pacific PNT Meeting pp 490ndash497

[4] P Hafner T Moder M Wieser and T Bernoulli ldquoEvaluationof smartphone-based indoor positioning using different Bayesfiltersrdquo in Proceedings of the 2013 International Conference on

Indoor Positioning and Indoor Navigation IPIN 2013 October2013

[5] T Willemsen F Keller and H Sternberg ldquoA topologicalapproach with MEMS in smartphones based on routing-graphrdquo in Proceedings of the International Conference on IndoorPositioning and Indoor Navigation IPIN 2015 October 2015

[6] F T Alaoui D Betaille and V Renaudin ldquoA multi-hypothesisparticle filtering approach for pedestrian dead reckoningrdquo inProceedings of the 2016 International Conference on IndoorPositioning and Indoor Navigation IPIN 2016 October 2016

[7] T Fetzer F Ebner F Deinzer L Koping and M GrzegorzekldquoOn Monte Carlo smoothing in multi sensor indoor localisa-tionrdquo in Proceedings of the 2016 International Conference onIndoor Positioning and Indoor Navigation IPIN 2016 October2016

[8] M R F Mendonca H S Bernardino and R F Neto ldquoStealthypath planning using navigation meshesrdquo in Proceedings of the4th Brazilian Conference on Intelligent Systems BRACIS 2015pp 31ndash36 bra November 2015

[9] ldquoRecast Detail API documentation for the members declaredin Recasth sdot recastnavigationrecastnavigation6f5c9f9rdquoGitHub httpsgithubcomrecastnavigationcommit6f5c9f9-b82418efc44b85974e604095b95354ada

[10] I Afyouni C Ray and C Claramunt ldquoSpatial models forcontext-aware indoor navigation systems a surveyrdquo Journal ofSpatial Information Science vol 4 no 1 pp 85ndash123 2012

[11] W Van Toll A F Cook and R Geraerts ldquoNavigation meshesfor realistic multi-layered environmentsrdquo in Proceedings of the2011 IEEERSJ International Conference on Intelligent Robots andSystems Celebrating 50Years of Robotics IROSrsquo11 pp 3526ndash3532September 2011

[12] DETOUR ldquoGitHubrdquo httpsgithubcomrecastnavigationreca-stnavigationtreemasterDetourSource 2017

[13] F Mortari S Zlatanova L Liu and E Clementini ldquoImprovedgeometric network model (IGNM) a novel approach for deriv-ing connectivity graphs for indoor navigationrdquo ISPRS Annalsof Photogrammetry Remote Sensing and Spatial InformationSciences vol II-4 pp 45ndash51 2014

[14] N O Eraghi F Lopez-Colino A De Castro and J GarridoldquoPath length comparison in grid maps of planning algorithmsHCTNav A and Dijkstrardquo Design of Circuits and IntegratedSystems pp 1ndash6 2014

[15] C Gaisbauer and A U Frank ldquoWayfinding Model For Pedes-trian Navigationrdquo in Proceedings of the 11th AGILE InternationalConference on Geographic Information Science Spain 2008

[16] N Victor evaluation des deplacements pietons quotidiens Uni-versite de Lyon 2016

[17] L Yang and M Worboys ldquoGeneration of navigation graphs forindoor spacerdquo International Journal of Geographical InformationScience vol 29 no 10 pp 1737ndash1756 2015

[18] M Susi V Renaudin and G Lachapelle ldquoMotion moderecognition and step detection algorithms for mobile phoneusersrdquo Sensors vol 13 no 2 pp 1539ndash1562 2013

[19] V RenaudinM Susi andG Lachapelle ldquoStep length estimationusing handheld inertial sensorsrdquo Sensors (Switzerland) vol 12no 7 pp 8507ndash8525 2012

[20] V Renaudin and C Combettes ldquoMagnetic acceleration fieldsand gyroscope quaternion (MAGYQ)-based attitude estimationwith smartphone sensors for indoor pedestrian navigationrdquoSensors (Switzerland) vol 14 no 12 pp 22864ndash22890 2014

10 Wireless Communications and Mobile Computing

[21] ULISS httpwwwifsttar-geolocfrindexphpenequipment44-uliss

[22] ldquoGitHub - tomojitakasuRTKLIB_binrdquo httpsgithubcomto-mojitakasuRTKLIB_bin 2017

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Page 3: Pedestrian Dead Reckoning Navigation with the Help of -Based Routing …downloads.hindawi.com/journals/wcmc/2017/7951346.pdf · ResearchArticle Pedestrian Dead Reckoning Navigation

Wireless Communications and Mobile Computing 3

scheme can be applied to any rasterized NavMesh wherepolygon weights are inherited by the raster cells

3 Alowast-Based Routing Graph Generation

31 Routing Graph Generation on the Basis of Waypoints Awaypoint is a punctual element that intervenes significantlyin the pathfinding process The significance of a waypoint isrelated to the role it plays in the pedestrianrsquos decision makingduring herhis travel Two situations are worth consideringThe first is when the pedestrian intends to reach a knowndestination within a familiar environment In this case it isobvious that (s)he follows the itinerary best suited to herhismobility profile Generally it is the shortest one but it couldbe any path depending on the cost function that definesthe userrsquos mobility profile This situation is handled by Alowastalgorithm using a weighted NavMesh The target destinationis known if it is visible from the pedestrianrsquos current positionfor example if the pedestrian is located at a graph node that isvisible from herhis target nodeThe second situation is whenthe pedestrianrsquos destination is invisible due to obstacles Inthis case waypoints are built in order to discretize the possibletarget destinations which may be either final or intermediatedestinations

To better understand the implementation of waypointsspace has to be considered in relation to human spatialcognition Indeed walkable areas within a building floorcould be corridors rooms or halls Outdoors walkable areasare the space comprising sidewalks footpaths squares carparks and so forth (roads belong to the drivable areas) Thisintuitive classification of space components allows definingscenes and extends what is called decision scenes to outdoorspace and corridors Decision scenes have been definedpreviously in [15] as the places that ldquocan be entered and leftand [are] physically bounded by buildings and other solidobstacles that prevent movementrdquo The extension of decisionscenes to corridors and footpaths in this paper is motivatedby the fact that these elements have borders (physical bordersfor corridors and geometrical borders for footpaths) andthat they can be accessed and left through specific pointsFor example two corridors are two separate scenes Theycan be accessed and left through their intersection whichis a building corner a corridor and hall are also distinctscenes between which traffic flow is possible through theirintersection Two elements are important for pedestriantravel inside and between scenes The first is related to theisovist which is visible space from the pedestrianrsquos currentposition The isovist is mainly influenced by the presenceof obstacles that have an impact on both visibility andthe pedestrianrsquos trajectory [16] Second element is the setof portals that allow flows of pedestrians from scene toscene [15] such as corridor extremities building gates storeentries or footpath extremities Indeed key elements in thepedestrian wayfinding behavior depend mainly on portalsand obstacle borders (which determine the isovist) Theymaterialize the fact that visibility and purpose (destination)are the parameters that give a shape to onersquos trajectory Inthis study the set of waypoints is composed of portals (cor-ridor extremities footpath extremities and outdoorindoor

Figure 2 Main area where Alowast algorithm has served for the routinggraph construction

doors) and obstacle corners The graph construction basedon waypoints is obviously more realistic than kernel-basedmethods where scenes are modelled by their centers [17]implying erroneously that the pedestrianwalks systematicallythrough the center of the scene Besides the structure of thegraph is not entirely determined by the geometry of spacebut handles behavioral-based paths generated according towaypoints and reflecting pedestrian travel strategies

Once waypoints have been constructed they can beexported in a database that will assist the pathfinding processEach pair of waypoints are related by a set of itinerariesthat are generated with Alowast algorithm and are part of thefinal routing graph Multihypotheses motion is handled asin classical routing graphs by exploring several paths andkeeping the ones that are best adapted to IMUmeasurementsThe graph structure is stored in the form of a GIS databasethat contains the graph segments identifiers their extremitynodes coordinates their length and their connections withinthe graph in a node-connectivity approach Each edge of thegraph is oriented and has an entry node A and an exit nodeB The node-connectivity design is involved in the motionmodel used for filtering

32 Alowast Path Planning for Seamless Transition betweenOutdoorand Indoor Spaces and within Obstacle-Free Areas Previouswork demonstrated that indoor layout-based graphs areefficient for enhancing the PDR localization within geomet-rically constrained areas [5 6] Hence the proposed methodis applied only to the area that precedes the mall principalentrance as well as in the obstacle-free hall (Figure 2) Thelatter represent the critical GNSS-deprived places where mapinformation fails to provide a calibration to the PDR posi-tioning process emphasizing the main issue being addressedin this paper Outdoors the routing graph is constructedaccording to the geometry of sidewalks crosswalks andfootpaths

Figure 3 shows the generated routing graph within thetest area Three waypoints materialize the building entrancegate (drawn in black in Figure 2) The focus is made onthe waypoint where red paths intersect It represents the leftside of the mall entrance Red paths are the Alowast-calculated

4 Wireless Communications and Mobile Computing

Figure 3 The highlighted segments of the graph in red colorshow the Alowast itineraries relating a portal waypoint (left side of thebuilding entrance door) to different destinations The latter areeither extremities of corridors or footpaths obstacle corners or storeentries

itineraries relating the left side of the gate to other places ofinterest such as stores entries or the buildingrsquosmain corridorsThe latter are modelled by a series of straight lines accordingto corridors main directions Three paths relate the outdoorto the indoor space through the left side of the mall entrance

This representation shows that straight line travels areprivileged for practical and fast displacement They are alsomore realistic in regard to pedestrian motion In fact somehypotheses are eliminated such as walking towards walls ormaking a series of turns to attend a place that can be reachedstraight ahead Moreover this provides a measurement ofwalking directions given by the graph segments in obstacle-free zones which would be impossible if a grid-basedapproach had been applied requiring additional informationsuch as radio beacons signal to compensate for the PDRerrors

4 IMU Fusion with GNSS and the RoutingGraph with a Particle Filter

41 Step Detection and Heading Calculation The step detec-tion is realized after motion classification according to [18] bydetecting peaks on the acceleration signal using an adaptivethreshold algorithm [19] According to the same reference ageneric model is used to estimate the step length This modelrelies on a set of three parameters trained on 10 subjects andis given by

119904 = 119896 sdot (ℎ (119886119891 + 119887) + 119888) (1)

where 119904 is the step length ℎ is the userrsquos height 119891 is the stepfrequency 119886 119887 119888 is the generic model parameters and 119896 isa scale factor that is expected to calibrate the model on thepedestrian

Headings have been calculated with MAGYQ attitudeestimation filter [20] that fuses signals from a triaxisaccelerometer a triaxis gyroscope and a triaxis magnetome-ter Heading calculation considers different carrying modessuch as the swinging mode or the texting mode The deviceorientation in 3D-space is then obtained and the yaw angle

deduced giving the orientation of the device relative to thetrue North

42 Calibration of the PDR Parameters Using GNSS Positionsand the Routing Graph ThePDR parameters to be calibratedare the step length and headingsThe step lengthmodel needsto be adjusted to each user as the model parameters aredependent on the pedestrianrsquos physiological features and gaitcycle whereas headings are potentially misaligned with theactual walking direction because of gyro drift and the devicecarrying mode In fact the device may be oriented towards adirection which is different from the walking directionTheseerrors are compensated by fusing the IMU with GNSS andthe routing graph The fusion is realized thanks to a particlefilter that models the state (vector of unknown variables) bya set of particles (a set of sampled state vectors) The statevector contains necessary variables for determining the userrsquosposition and is presented in detail in Section 43

The step length model is adjusted to the pedestrian usingboth the graph and GNSS positions when available In factthe graph allows keeping the positions on plausible path(s)This directly impacts the travel distance Besides GNSSdecreases the particles dispersion by bringing them next tothe GNSS position On the other hand walking directionsare given by path headings They are particularly reliableindoors as the paths are calculated by the Alowast algorithm or aredirectly given by the corridors main directions The routinggraph-derived walking direction is then compared to theIMU pointing direction and the most likely path is selectedThe difference between both headings gives the IMU angularmisalignment with the pedestrianrsquos walking direction

The step length model is calibrated using a scale factor 119896The latter is assumed constant as it depends mainly onthe pedestrianrsquos physiological parameters It is modelled bya Gaussian variable with 1m mean and low variations asthe step length is restricted to realistic values comprising060mndash1m 119896 is expected to converge within straight linetravels where an optimal calibration is guaranteed

Heading misalignment 120573 is the angular differencebetween the actual walking direction and the pointingdirection of the IMU It is time-dependent because it varieswith handmotion Uniformly distributed samples of headingmisalignment values are drawn with variations up to 30∘

43 Model Implementation with a Particle Filter The statevector is given by

IdDp 120575 119896 120573 119908119901119905 (2)

where Id is the edge identifier in the graph database Dp is thecurvilinear abscissa 120575 represents the travel direction whichcan be 1 if the edge is walked forward (from A to B) and minus1otherwise (fromB toA) 119896 is the scale factor that calibrates thestep lengthmodel120573 is themisalignment between thewalkingdirection and the pointing direction of the handheld device119908 is the particle weight and 119901 and 119905 refer respectively to theparticle and time

Each particle is moved along the current path accordingto the travel distance given by the step length The particle

Wireless Communications and Mobile Computing 5

BA connectivity BB connectivity

$5 gt 0

ltL)>minus1gtL)>minus1

)>t = )>tminus1 )>t = )>tminus1 )>t = )>tminus1

$Jt = $5

t = tminus1 t = tminus1

$Jt = $5 minus L)>minus1$Jt = L)>

minus

t = minustminus1

($5 minus L)>minus1)

Figure 4 Dp computation according to DU

filter introduces white noise on each parameter to spread theparticles and optimize the graph exploration The particleposition is determined by the identifier of the graph edge andits curvilinear abscissa If the latter is above the edge lengthone of its connected edges will be explored The transitionmodel is

DU = (119904) sdot 120575 sdot 120572 + 119899119866 + Dp119905minus1 (3)

where DU is compared to the current edge length 119871 Id tocompute the curvilinear abscissa Dp (Figure 4) 119904 is thestep length and 119899119866 is the modelling error represented by aGaussian noise 120572 isin 0 1 2 and determines the numberof steps to be made by one particle It allows detectingovermisdetected steps

Figure 4 gives the transition test process that allowscomputing the curvilinear abscissa Dp when an exit nodeB is exceeded (ie DU gt 0) 119905 is the time index BABBconnectivity translates the search in the graph database inorder to determine if the edge being explored is connectedto the exceeded one via an A or a B node The same logic isadopted if an entry node A is exceeded (ie DU lt 0)

When GNSS positions are accurate enough they areused together with the graph headings in order to calibrateboth the step length model and PDR headings The graphheadings provide ameasurement of walking directions whichare compared to PDR headings taking into account potentialangular misalignment This is performed according to

119908119905= 119908119905minus1sdot exp(minus12 sdot (

1003816100381610038161003816120579PDR minus (120579Id + 120573)10038161003816100381610038162(120590120579Id2 + 120590120579PDR 2) + Δ1015840Σminus1Δ))

Δ = [119864119875 minus 119864GPS119873119875 minus 119873GPS]

Σminus1 = [[[[[

1(120590119864GPS2 + 120590map

119864

2) 00 1

(120590119873GPS2 + 120590map119873

2)]]]]]

(4)

Figure 5 Ubiquitous Localization Unit with Inertial Sensors andSatellites

where 119908119905 is the particle weight and 119864119875 and 119864GPS arerespectively the predicted East coordinate of a particle and itscorresponding GPS observation 119873119875 and 119873GPS are the sameparameters along the North direction Σminus1 is the weightingmatrix considering the graph and GPS position accuracies

Indoors some GNSS positions are still available but thelatter are generally unreliable and are rejected The rejectiontest is made by thresholding the signal to noise ratio (SNR)and the horizontal dilution of precision (HDOP) Only thegraph headings are exploited in the weighting process Theweighting equation indoors is

119908119905 = 119908119905minus1 sdot exp(minus12 times1003816100381610038161003816120579PDR minus (120579Id + 120573)10038161003816100381610038162(120590120579Id2 + 120590120579PDR 2) ) (5)

where 120579PDR is the walking direction estimated by the PDRalgorithms 120579Id is the predicted heading of the current path120573 is the heading misalignment between the path directionand the PDR estimate 120590120579Id is the standard deviation of thepath heading and 120590120579PDR is the standard deviation of the PDRwalking direction estimate

Once the particles have been weighted some of themare assigned low weights and become useless in the processResampling is performed to duplicate the particles with highweights and delete the others The particles amount is keptconstant and their weights equal after each update in order toexplore enough hypotheses of motion

6 Wireless Communications and Mobile Computing

Figure 6 Data collection by a pedestrian with ULISS unit in hand

5 Experiments

51 Data Collection with a HSGNSS and IMU in Hand Threehealthy volunteers (2 men 1 woman) collected data withULISS (Ubiquitous Localization Unit with Inertial Sensorsand Satellites) device (Figure 5) held in hand (Figure 6) Datawere collected in both outdoor and indoor environments fortwo different device carrying modes Two acquisitions weremade in a textingmode and one in both swinging and textingmodes The average duration of acquisitions was about 14minutes and the walking distance for each trial almost 15 km

Different scenarios were chosen to perform the experi-ments These will be explained in detail in Section 61 wherethe text is enhanced by figures

52 Input Data ULISS device [21] comprises 9 degreesof freedom inertial mobile unit a high sensitivity GNSS(HSGNSS) receiver and antenna a memory card and abattery It delivers measurements that are time-stampedin GPS time Inertial sensors and magnetometers providemeasurements at a 200Hz frequency

TheHSGNSS receiver operates in a standalone mode anddelivers positions in real time at a 5Hz frequency Deliveredpositions are time-stamped in GPS time and have metricaccuracies ranging from 2m up to 10m near buildings andtree shades In this work GNSS positions were interpolated atthe step frequency in order to be fusedwith the PDRestimatesof headings and step lengths

53 Reference Trajectories Besides collecting data withULISS device in hand all volunteers were equipped with anindependent GNSS receiver carried in their backpacks anda small antenna attached to their caps GNSS measurementswere then postprocessed in order to calculate reference trajec-tories by differential GNSS This was performed using RTK-LIB 242p12 software [22]Measurements from the embarkedGNSS receiver and from a nearby base station were used toperform relative GNSS positioning Obtained positions had

0 100 200 300 400East (m)

0

100

200

300N

orth

(m)

PDR trajectoryEstimated trajectory

Differential GNSS trajectory

Figure 7 Estimated trajectory (in red) The blue trajectory givesthe PDR position estimates and the green one gives the referencetrajectory (only texting)

decimetric accuracies up to several meters near buildings andother elevated features Afterwards thresholding was appliedto three parameters in order to reject some outlier positionestimates These parameters are first the number of visiblesatellites second the ratio factor of ambiguity resolutionand finally the horizontal dilution of precisionThe resultingpositions had precisions below 1m and were adequate foraccuracy assessment in this work

6 Results

61 Trajectory Analysis Figures 7 8 and 9 show the esti-mated trajectories (red) with the Alowast-generated routing graphand the particle filter described in Section 43 The blue

Wireless Communications and Mobile Computing 7

0 100 200 300 400East (m)

0

100

200

300

400

Nor

th (m

)

Figure 8 Trajectories for the second dataset (only texting)

0 100 200 300 400 500East (m)

0

100

200

300

400

500

Nor

th (m

)

Figure 9 Trajectories for the third dataset (swinging texting)

pattern corresponds to the PDR trajectory while the greenone is the reference trajectory obtained by differential GNSS

Figures 7 and 8 correspond to acquisitions performed inthe texting modeThe starting point for both acquisitions lieson the top right extremity of the trajectories Both subjectsmade a closed loop around the building with an intermediateoutdoor travel (bottom side of the figures) before reaching thestarting point back

For both acquisitions the travel distance seems to beoverestimated Yet heading determination is more accurate

for the first acquisition as the shape of the trajectory is morefaithful to the building structure The drift is higher for thesecond dataset and can be visually observed at the end of thePDR trajectory

Where the drift is most important the PDR trajectorypresents major inconsistencies with the map According toFigures 7 8 and 9 the drift has been corrected as theestimated trajectories are more compliant with the buildingstructure and with footpaths in outdoor space This has beenachieved thanks to the proposed particle filter and to anincreased conformitywith pedestrianmotion demonstratingthat the positioning accuracy for the texting scenario has beensignificantly enhanced using our approach

Figure 9 corresponds to data collected for both theswinging and texting modes The acquisition started at thebottom of the figure where the reference and the filteredtrajectories overlap Starting from this position the subjectentered the building and then went outside through theNorth-East building entrance This travel was made in theswinging mode The top right extremity of the trajectoryunderlines a U-turn before the subject entered the buildingback to reach the starting point This part of the travel wasperformed in the texting mode

Unlike the two first acquisitions there is a gap in the PDRtrajectory because the subject did not go around the buildingas the two first subjects didThis gap is retrieved in the filteredtrajectory

Obviously this scenario implies a less accurate headingdetermination and even an alteration in the travel distanceestimation In fact thewalking distance is underestimated forthe swinging mode (first part of the travel until the subjectreached the outdoor) and overestimated for the texting mode(second part of the travel until the ending point)These errorscan be noticed in the PDR trajectory

The filtered trajectory shows that the drift has beencorrected resulting in a shape that is more compliant withthe map and with the reference trajectory outdoors Yet thepositioning accuracy seems to be decreased as comparedwith the two first datasets Following section discusses theaccuracy of estimated trajectories

62 Error Computation In order to assess the accuracy ofour positioning method filtered positions were comparedto the reference positions interpolated at the step frequencyThe average plane error ranges from 4 to 5 meters for thethree datasets (Figure 10) Accuracy is dependent on thequality of the PDR trajectory Therefore computed errorsare more important for the second dataset considering thetexting scenario For the third acquisition that includes bothswinging and texting the accuracy is significantly decreasedand more outliers (precisions above 15m) which are givenby the red plus signs in Figure 10 are detected due tomismatching errors (ie choosing wrong edges of the graph)These errors are mainly due to uncalibrated PDR parametersand are discussed in the following sections

63 Heading Misalignment Estimation Figures 11 12 and 13show the estimated heading misalignment values for each

8 Wireless Communications and Mobile Computing

2 31Datasets

0

5

10

15

20

Erro

r (m

)

Figure 10 Plane errors for each dataset

2 4 6 80 10 12Time (min)

minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 11 Full texting scenario (1)

dataset For the first trial (Figure 11) angular misalignmentis comprised between minus10∘ and +10∘ and varies around anapproximate mean value of 0∘ According to this distributionthe angular difference between walking directions and thepointing direction of the device is minimal Hence appliedcorrections compensate only for the gyro driftwhich is ratherlogical regarding the texting mode scenario

For the second dataset (Figure 12) estimated headingmisalignment values are between minus15∘ and +12∘ They arenot equally distributed around 0∘ (eg between the 1st and2nd minutes the mean value is over 5∘) From this analysisnonnegligible hand motion can be assumed even if thesubject intended to perform the experiment in the textingmode

Figure 13 gives the estimated heading misalignment forthe third trial The first part (until 95min) of the travel wasperformed in the swingingmodeHence the estimated valuesvary significantly (minus15∘ to +20∘) For the second part of theplot (gt10min) angular misalignment variations occur withlower magnitudes comprising plusmn5∘ over a mean value of 0∘which reflects the texting mode of the acquisition

64 Step Length Model Calibration In this paper a scalefactor was introduced on the step length in order to calibratethe walking distance though due to degraded GNSS signaland to short walking periods in outdoor space the scale

10 12 14Time (m)

2 4 6 80minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 12 Full texting scenario (2)

minus20minus15minus10minus5

05

101520

Hea

ding

misa

lignm

ent(

∘ )

6 8 141210 16 182 40

Time (min)

Figure 13 Swinging-texting scenario

factor was not calibrated In fact regular and accurate GNSSpositions are needed through straight line travels as cited inour introduction in order to calibrate the walking distanceThese conditions were not verified during our experiments

As a result only corrected headings were relied on in theselection process Therefore distance calibration occurredonly at junctions of the graph when a change in heading wasdetected This explains the fact that the accuracy of filteredtrajectories is still enhanced and compliance with the mapimproved

Though the uncalibrated walking distances caused somemismatching errors because direction change was detectedtoo late or too early in the process this can be noticed in theNorthern part of Figure 9 Indeed while the pedestrian wasintending to exit the building the filtered trajectory indicatesthat he was walking towards a corridor This happened fortwo reasons First the real trajectory is quite unusual interms of pedestrian behavior In fact there is a change inheading (observed in the PDR trajectory) that is independentof space configuration invalidating the assumptions thatallowed constructing our navigation network Second theuncalibrated walking distance prevented the particles fromreaching outdoor space at the right time Another mismatch-ing error occurred at the middle of the building (between300m and 400m North) because direction change wasdetected too late due to uncalibrated walking distance Lateron the particle filter corrected for this error and convergedover the right corridor thanks to the particle dispersion overthe graph and to the multihypothesis approach

Wireless Communications and Mobile Computing 9

7 Conclusions

Amap-aided PDR approach where a routing graph is used asmotion model has been proposed Main contribution of thispaper is Alowast algorithm adaptation to elaborate a pedestriannetwork that is capable of cancelling the gyro drift and themisalignment between the device orientation and thewalkingdirection even in large spaces These are GNSS-deprived andobstacle-free areas where the limitations of map-aided PDRalgorithms are most important In fact widespread map-aided PDR approaches do not compensate for these errorswhen pedestrian motion is unconstrained mainly duringthe transition between outdoor and indoor spaces and whenobstacles are absent The Alowast-based routing graph mitigatesthe lack of obstacles thanks to a set of waypoints implementedaccording to human spatial cognition and to a weightednavigation mesh This allows building a realistic motionmodel thatmeets the requirements ofmap-aided localizationIndeed the proposed routing graph is well exploited becauseit gives prior knowledge about the pedestrianrsquos destinationand provides reliable measurements of walking directionsResults show that it is adequate for a seamless transitionbetween outdoor and indoor environments and for enhanc-ing the positioning accuracy even in large spaces Achievedaccuracies range from 3 to 5 meters and the drift is almostcancelled with the help of the routing graph though somemismatching errors due to uncalibrated walking distanceespecially while carrying the device in the swinging modemight induce important positioning errors Indeed properconditions of sky visibility and sufficient period of outdoorwalking are prerequisite for the step length calibration beforethe pedestrian reaches indoor space

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This researchworkwas done as part of theHappyHand (2015ndash2018) project which is funded by the 19th ldquoFonds UniqueInterministerielrdquo (French national RampD funding)

References

[1] J B Bancroft and G Lachapelle ldquoData fusion algorithms formultiple inertial measurement unitsrdquo Sensors vol 11 no 7 pp6771ndash6798 2011

[2] Y Hu L Sheng and S J Zhang ldquoDesign of Continuous IndoorNavigation System Based on INS and Wifirdquo Applied Mechanicsand Materials vol 303-306 pp 2046ndash2049 2013

[3] T Lee B Shin J H Lee et al ldquoAn Indoor Positioning SystemUsing Vision aided Advanced PDR technology without imageDB andwithmotion recognitionrdquo inProceedings of the ION2013Pacific PNT Meeting pp 490ndash497

[4] P Hafner T Moder M Wieser and T Bernoulli ldquoEvaluationof smartphone-based indoor positioning using different Bayesfiltersrdquo in Proceedings of the 2013 International Conference on

Indoor Positioning and Indoor Navigation IPIN 2013 October2013

[5] T Willemsen F Keller and H Sternberg ldquoA topologicalapproach with MEMS in smartphones based on routing-graphrdquo in Proceedings of the International Conference on IndoorPositioning and Indoor Navigation IPIN 2015 October 2015

[6] F T Alaoui D Betaille and V Renaudin ldquoA multi-hypothesisparticle filtering approach for pedestrian dead reckoningrdquo inProceedings of the 2016 International Conference on IndoorPositioning and Indoor Navigation IPIN 2016 October 2016

[7] T Fetzer F Ebner F Deinzer L Koping and M GrzegorzekldquoOn Monte Carlo smoothing in multi sensor indoor localisa-tionrdquo in Proceedings of the 2016 International Conference onIndoor Positioning and Indoor Navigation IPIN 2016 October2016

[8] M R F Mendonca H S Bernardino and R F Neto ldquoStealthypath planning using navigation meshesrdquo in Proceedings of the4th Brazilian Conference on Intelligent Systems BRACIS 2015pp 31ndash36 bra November 2015

[9] ldquoRecast Detail API documentation for the members declaredin Recasth sdot recastnavigationrecastnavigation6f5c9f9rdquoGitHub httpsgithubcomrecastnavigationcommit6f5c9f9-b82418efc44b85974e604095b95354ada

[10] I Afyouni C Ray and C Claramunt ldquoSpatial models forcontext-aware indoor navigation systems a surveyrdquo Journal ofSpatial Information Science vol 4 no 1 pp 85ndash123 2012

[11] W Van Toll A F Cook and R Geraerts ldquoNavigation meshesfor realistic multi-layered environmentsrdquo in Proceedings of the2011 IEEERSJ International Conference on Intelligent Robots andSystems Celebrating 50Years of Robotics IROSrsquo11 pp 3526ndash3532September 2011

[12] DETOUR ldquoGitHubrdquo httpsgithubcomrecastnavigationreca-stnavigationtreemasterDetourSource 2017

[13] F Mortari S Zlatanova L Liu and E Clementini ldquoImprovedgeometric network model (IGNM) a novel approach for deriv-ing connectivity graphs for indoor navigationrdquo ISPRS Annalsof Photogrammetry Remote Sensing and Spatial InformationSciences vol II-4 pp 45ndash51 2014

[14] N O Eraghi F Lopez-Colino A De Castro and J GarridoldquoPath length comparison in grid maps of planning algorithmsHCTNav A and Dijkstrardquo Design of Circuits and IntegratedSystems pp 1ndash6 2014

[15] C Gaisbauer and A U Frank ldquoWayfinding Model For Pedes-trian Navigationrdquo in Proceedings of the 11th AGILE InternationalConference on Geographic Information Science Spain 2008

[16] N Victor evaluation des deplacements pietons quotidiens Uni-versite de Lyon 2016

[17] L Yang and M Worboys ldquoGeneration of navigation graphs forindoor spacerdquo International Journal of Geographical InformationScience vol 29 no 10 pp 1737ndash1756 2015

[18] M Susi V Renaudin and G Lachapelle ldquoMotion moderecognition and step detection algorithms for mobile phoneusersrdquo Sensors vol 13 no 2 pp 1539ndash1562 2013

[19] V RenaudinM Susi andG Lachapelle ldquoStep length estimationusing handheld inertial sensorsrdquo Sensors (Switzerland) vol 12no 7 pp 8507ndash8525 2012

[20] V Renaudin and C Combettes ldquoMagnetic acceleration fieldsand gyroscope quaternion (MAGYQ)-based attitude estimationwith smartphone sensors for indoor pedestrian navigationrdquoSensors (Switzerland) vol 14 no 12 pp 22864ndash22890 2014

10 Wireless Communications and Mobile Computing

[21] ULISS httpwwwifsttar-geolocfrindexphpenequipment44-uliss

[22] ldquoGitHub - tomojitakasuRTKLIB_binrdquo httpsgithubcomto-mojitakasuRTKLIB_bin 2017

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Page 4: Pedestrian Dead Reckoning Navigation with the Help of -Based Routing …downloads.hindawi.com/journals/wcmc/2017/7951346.pdf · ResearchArticle Pedestrian Dead Reckoning Navigation

4 Wireless Communications and Mobile Computing

Figure 3 The highlighted segments of the graph in red colorshow the Alowast itineraries relating a portal waypoint (left side of thebuilding entrance door) to different destinations The latter areeither extremities of corridors or footpaths obstacle corners or storeentries

itineraries relating the left side of the gate to other places ofinterest such as stores entries or the buildingrsquosmain corridorsThe latter are modelled by a series of straight lines accordingto corridors main directions Three paths relate the outdoorto the indoor space through the left side of the mall entrance

This representation shows that straight line travels areprivileged for practical and fast displacement They are alsomore realistic in regard to pedestrian motion In fact somehypotheses are eliminated such as walking towards walls ormaking a series of turns to attend a place that can be reachedstraight ahead Moreover this provides a measurement ofwalking directions given by the graph segments in obstacle-free zones which would be impossible if a grid-basedapproach had been applied requiring additional informationsuch as radio beacons signal to compensate for the PDRerrors

4 IMU Fusion with GNSS and the RoutingGraph with a Particle Filter

41 Step Detection and Heading Calculation The step detec-tion is realized after motion classification according to [18] bydetecting peaks on the acceleration signal using an adaptivethreshold algorithm [19] According to the same reference ageneric model is used to estimate the step length This modelrelies on a set of three parameters trained on 10 subjects andis given by

119904 = 119896 sdot (ℎ (119886119891 + 119887) + 119888) (1)

where 119904 is the step length ℎ is the userrsquos height 119891 is the stepfrequency 119886 119887 119888 is the generic model parameters and 119896 isa scale factor that is expected to calibrate the model on thepedestrian

Headings have been calculated with MAGYQ attitudeestimation filter [20] that fuses signals from a triaxisaccelerometer a triaxis gyroscope and a triaxis magnetome-ter Heading calculation considers different carrying modessuch as the swinging mode or the texting mode The deviceorientation in 3D-space is then obtained and the yaw angle

deduced giving the orientation of the device relative to thetrue North

42 Calibration of the PDR Parameters Using GNSS Positionsand the Routing Graph ThePDR parameters to be calibratedare the step length and headingsThe step lengthmodel needsto be adjusted to each user as the model parameters aredependent on the pedestrianrsquos physiological features and gaitcycle whereas headings are potentially misaligned with theactual walking direction because of gyro drift and the devicecarrying mode In fact the device may be oriented towards adirection which is different from the walking directionTheseerrors are compensated by fusing the IMU with GNSS andthe routing graph The fusion is realized thanks to a particlefilter that models the state (vector of unknown variables) bya set of particles (a set of sampled state vectors) The statevector contains necessary variables for determining the userrsquosposition and is presented in detail in Section 43

The step length model is adjusted to the pedestrian usingboth the graph and GNSS positions when available In factthe graph allows keeping the positions on plausible path(s)This directly impacts the travel distance Besides GNSSdecreases the particles dispersion by bringing them next tothe GNSS position On the other hand walking directionsare given by path headings They are particularly reliableindoors as the paths are calculated by the Alowast algorithm or aredirectly given by the corridors main directions The routinggraph-derived walking direction is then compared to theIMU pointing direction and the most likely path is selectedThe difference between both headings gives the IMU angularmisalignment with the pedestrianrsquos walking direction

The step length model is calibrated using a scale factor 119896The latter is assumed constant as it depends mainly onthe pedestrianrsquos physiological parameters It is modelled bya Gaussian variable with 1m mean and low variations asthe step length is restricted to realistic values comprising060mndash1m 119896 is expected to converge within straight linetravels where an optimal calibration is guaranteed

Heading misalignment 120573 is the angular differencebetween the actual walking direction and the pointingdirection of the IMU It is time-dependent because it varieswith handmotion Uniformly distributed samples of headingmisalignment values are drawn with variations up to 30∘

43 Model Implementation with a Particle Filter The statevector is given by

IdDp 120575 119896 120573 119908119901119905 (2)

where Id is the edge identifier in the graph database Dp is thecurvilinear abscissa 120575 represents the travel direction whichcan be 1 if the edge is walked forward (from A to B) and minus1otherwise (fromB toA) 119896 is the scale factor that calibrates thestep lengthmodel120573 is themisalignment between thewalkingdirection and the pointing direction of the handheld device119908 is the particle weight and 119901 and 119905 refer respectively to theparticle and time

Each particle is moved along the current path accordingto the travel distance given by the step length The particle

Wireless Communications and Mobile Computing 5

BA connectivity BB connectivity

$5 gt 0

ltL)>minus1gtL)>minus1

)>t = )>tminus1 )>t = )>tminus1 )>t = )>tminus1

$Jt = $5

t = tminus1 t = tminus1

$Jt = $5 minus L)>minus1$Jt = L)>

minus

t = minustminus1

($5 minus L)>minus1)

Figure 4 Dp computation according to DU

filter introduces white noise on each parameter to spread theparticles and optimize the graph exploration The particleposition is determined by the identifier of the graph edge andits curvilinear abscissa If the latter is above the edge lengthone of its connected edges will be explored The transitionmodel is

DU = (119904) sdot 120575 sdot 120572 + 119899119866 + Dp119905minus1 (3)

where DU is compared to the current edge length 119871 Id tocompute the curvilinear abscissa Dp (Figure 4) 119904 is thestep length and 119899119866 is the modelling error represented by aGaussian noise 120572 isin 0 1 2 and determines the numberof steps to be made by one particle It allows detectingovermisdetected steps

Figure 4 gives the transition test process that allowscomputing the curvilinear abscissa Dp when an exit nodeB is exceeded (ie DU gt 0) 119905 is the time index BABBconnectivity translates the search in the graph database inorder to determine if the edge being explored is connectedto the exceeded one via an A or a B node The same logic isadopted if an entry node A is exceeded (ie DU lt 0)

When GNSS positions are accurate enough they areused together with the graph headings in order to calibrateboth the step length model and PDR headings The graphheadings provide ameasurement of walking directions whichare compared to PDR headings taking into account potentialangular misalignment This is performed according to

119908119905= 119908119905minus1sdot exp(minus12 sdot (

1003816100381610038161003816120579PDR minus (120579Id + 120573)10038161003816100381610038162(120590120579Id2 + 120590120579PDR 2) + Δ1015840Σminus1Δ))

Δ = [119864119875 minus 119864GPS119873119875 minus 119873GPS]

Σminus1 = [[[[[

1(120590119864GPS2 + 120590map

119864

2) 00 1

(120590119873GPS2 + 120590map119873

2)]]]]]

(4)

Figure 5 Ubiquitous Localization Unit with Inertial Sensors andSatellites

where 119908119905 is the particle weight and 119864119875 and 119864GPS arerespectively the predicted East coordinate of a particle and itscorresponding GPS observation 119873119875 and 119873GPS are the sameparameters along the North direction Σminus1 is the weightingmatrix considering the graph and GPS position accuracies

Indoors some GNSS positions are still available but thelatter are generally unreliable and are rejected The rejectiontest is made by thresholding the signal to noise ratio (SNR)and the horizontal dilution of precision (HDOP) Only thegraph headings are exploited in the weighting process Theweighting equation indoors is

119908119905 = 119908119905minus1 sdot exp(minus12 times1003816100381610038161003816120579PDR minus (120579Id + 120573)10038161003816100381610038162(120590120579Id2 + 120590120579PDR 2) ) (5)

where 120579PDR is the walking direction estimated by the PDRalgorithms 120579Id is the predicted heading of the current path120573 is the heading misalignment between the path directionand the PDR estimate 120590120579Id is the standard deviation of thepath heading and 120590120579PDR is the standard deviation of the PDRwalking direction estimate

Once the particles have been weighted some of themare assigned low weights and become useless in the processResampling is performed to duplicate the particles with highweights and delete the others The particles amount is keptconstant and their weights equal after each update in order toexplore enough hypotheses of motion

6 Wireless Communications and Mobile Computing

Figure 6 Data collection by a pedestrian with ULISS unit in hand

5 Experiments

51 Data Collection with a HSGNSS and IMU in Hand Threehealthy volunteers (2 men 1 woman) collected data withULISS (Ubiquitous Localization Unit with Inertial Sensorsand Satellites) device (Figure 5) held in hand (Figure 6) Datawere collected in both outdoor and indoor environments fortwo different device carrying modes Two acquisitions weremade in a textingmode and one in both swinging and textingmodes The average duration of acquisitions was about 14minutes and the walking distance for each trial almost 15 km

Different scenarios were chosen to perform the experi-ments These will be explained in detail in Section 61 wherethe text is enhanced by figures

52 Input Data ULISS device [21] comprises 9 degreesof freedom inertial mobile unit a high sensitivity GNSS(HSGNSS) receiver and antenna a memory card and abattery It delivers measurements that are time-stampedin GPS time Inertial sensors and magnetometers providemeasurements at a 200Hz frequency

TheHSGNSS receiver operates in a standalone mode anddelivers positions in real time at a 5Hz frequency Deliveredpositions are time-stamped in GPS time and have metricaccuracies ranging from 2m up to 10m near buildings andtree shades In this work GNSS positions were interpolated atthe step frequency in order to be fusedwith the PDRestimatesof headings and step lengths

53 Reference Trajectories Besides collecting data withULISS device in hand all volunteers were equipped with anindependent GNSS receiver carried in their backpacks anda small antenna attached to their caps GNSS measurementswere then postprocessed in order to calculate reference trajec-tories by differential GNSS This was performed using RTK-LIB 242p12 software [22]Measurements from the embarkedGNSS receiver and from a nearby base station were used toperform relative GNSS positioning Obtained positions had

0 100 200 300 400East (m)

0

100

200

300N

orth

(m)

PDR trajectoryEstimated trajectory

Differential GNSS trajectory

Figure 7 Estimated trajectory (in red) The blue trajectory givesthe PDR position estimates and the green one gives the referencetrajectory (only texting)

decimetric accuracies up to several meters near buildings andother elevated features Afterwards thresholding was appliedto three parameters in order to reject some outlier positionestimates These parameters are first the number of visiblesatellites second the ratio factor of ambiguity resolutionand finally the horizontal dilution of precisionThe resultingpositions had precisions below 1m and were adequate foraccuracy assessment in this work

6 Results

61 Trajectory Analysis Figures 7 8 and 9 show the esti-mated trajectories (red) with the Alowast-generated routing graphand the particle filter described in Section 43 The blue

Wireless Communications and Mobile Computing 7

0 100 200 300 400East (m)

0

100

200

300

400

Nor

th (m

)

Figure 8 Trajectories for the second dataset (only texting)

0 100 200 300 400 500East (m)

0

100

200

300

400

500

Nor

th (m

)

Figure 9 Trajectories for the third dataset (swinging texting)

pattern corresponds to the PDR trajectory while the greenone is the reference trajectory obtained by differential GNSS

Figures 7 and 8 correspond to acquisitions performed inthe texting modeThe starting point for both acquisitions lieson the top right extremity of the trajectories Both subjectsmade a closed loop around the building with an intermediateoutdoor travel (bottom side of the figures) before reaching thestarting point back

For both acquisitions the travel distance seems to beoverestimated Yet heading determination is more accurate

for the first acquisition as the shape of the trajectory is morefaithful to the building structure The drift is higher for thesecond dataset and can be visually observed at the end of thePDR trajectory

Where the drift is most important the PDR trajectorypresents major inconsistencies with the map According toFigures 7 8 and 9 the drift has been corrected as theestimated trajectories are more compliant with the buildingstructure and with footpaths in outdoor space This has beenachieved thanks to the proposed particle filter and to anincreased conformitywith pedestrianmotion demonstratingthat the positioning accuracy for the texting scenario has beensignificantly enhanced using our approach

Figure 9 corresponds to data collected for both theswinging and texting modes The acquisition started at thebottom of the figure where the reference and the filteredtrajectories overlap Starting from this position the subjectentered the building and then went outside through theNorth-East building entrance This travel was made in theswinging mode The top right extremity of the trajectoryunderlines a U-turn before the subject entered the buildingback to reach the starting point This part of the travel wasperformed in the texting mode

Unlike the two first acquisitions there is a gap in the PDRtrajectory because the subject did not go around the buildingas the two first subjects didThis gap is retrieved in the filteredtrajectory

Obviously this scenario implies a less accurate headingdetermination and even an alteration in the travel distanceestimation In fact thewalking distance is underestimated forthe swinging mode (first part of the travel until the subjectreached the outdoor) and overestimated for the texting mode(second part of the travel until the ending point)These errorscan be noticed in the PDR trajectory

The filtered trajectory shows that the drift has beencorrected resulting in a shape that is more compliant withthe map and with the reference trajectory outdoors Yet thepositioning accuracy seems to be decreased as comparedwith the two first datasets Following section discusses theaccuracy of estimated trajectories

62 Error Computation In order to assess the accuracy ofour positioning method filtered positions were comparedto the reference positions interpolated at the step frequencyThe average plane error ranges from 4 to 5 meters for thethree datasets (Figure 10) Accuracy is dependent on thequality of the PDR trajectory Therefore computed errorsare more important for the second dataset considering thetexting scenario For the third acquisition that includes bothswinging and texting the accuracy is significantly decreasedand more outliers (precisions above 15m) which are givenby the red plus signs in Figure 10 are detected due tomismatching errors (ie choosing wrong edges of the graph)These errors are mainly due to uncalibrated PDR parametersand are discussed in the following sections

63 Heading Misalignment Estimation Figures 11 12 and 13show the estimated heading misalignment values for each

8 Wireless Communications and Mobile Computing

2 31Datasets

0

5

10

15

20

Erro

r (m

)

Figure 10 Plane errors for each dataset

2 4 6 80 10 12Time (min)

minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 11 Full texting scenario (1)

dataset For the first trial (Figure 11) angular misalignmentis comprised between minus10∘ and +10∘ and varies around anapproximate mean value of 0∘ According to this distributionthe angular difference between walking directions and thepointing direction of the device is minimal Hence appliedcorrections compensate only for the gyro driftwhich is ratherlogical regarding the texting mode scenario

For the second dataset (Figure 12) estimated headingmisalignment values are between minus15∘ and +12∘ They arenot equally distributed around 0∘ (eg between the 1st and2nd minutes the mean value is over 5∘) From this analysisnonnegligible hand motion can be assumed even if thesubject intended to perform the experiment in the textingmode

Figure 13 gives the estimated heading misalignment forthe third trial The first part (until 95min) of the travel wasperformed in the swingingmodeHence the estimated valuesvary significantly (minus15∘ to +20∘) For the second part of theplot (gt10min) angular misalignment variations occur withlower magnitudes comprising plusmn5∘ over a mean value of 0∘which reflects the texting mode of the acquisition

64 Step Length Model Calibration In this paper a scalefactor was introduced on the step length in order to calibratethe walking distance though due to degraded GNSS signaland to short walking periods in outdoor space the scale

10 12 14Time (m)

2 4 6 80minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 12 Full texting scenario (2)

minus20minus15minus10minus5

05

101520

Hea

ding

misa

lignm

ent(

∘ )

6 8 141210 16 182 40

Time (min)

Figure 13 Swinging-texting scenario

factor was not calibrated In fact regular and accurate GNSSpositions are needed through straight line travels as cited inour introduction in order to calibrate the walking distanceThese conditions were not verified during our experiments

As a result only corrected headings were relied on in theselection process Therefore distance calibration occurredonly at junctions of the graph when a change in heading wasdetected This explains the fact that the accuracy of filteredtrajectories is still enhanced and compliance with the mapimproved

Though the uncalibrated walking distances caused somemismatching errors because direction change was detectedtoo late or too early in the process this can be noticed in theNorthern part of Figure 9 Indeed while the pedestrian wasintending to exit the building the filtered trajectory indicatesthat he was walking towards a corridor This happened fortwo reasons First the real trajectory is quite unusual interms of pedestrian behavior In fact there is a change inheading (observed in the PDR trajectory) that is independentof space configuration invalidating the assumptions thatallowed constructing our navigation network Second theuncalibrated walking distance prevented the particles fromreaching outdoor space at the right time Another mismatch-ing error occurred at the middle of the building (between300m and 400m North) because direction change wasdetected too late due to uncalibrated walking distance Lateron the particle filter corrected for this error and convergedover the right corridor thanks to the particle dispersion overthe graph and to the multihypothesis approach

Wireless Communications and Mobile Computing 9

7 Conclusions

Amap-aided PDR approach where a routing graph is used asmotion model has been proposed Main contribution of thispaper is Alowast algorithm adaptation to elaborate a pedestriannetwork that is capable of cancelling the gyro drift and themisalignment between the device orientation and thewalkingdirection even in large spaces These are GNSS-deprived andobstacle-free areas where the limitations of map-aided PDRalgorithms are most important In fact widespread map-aided PDR approaches do not compensate for these errorswhen pedestrian motion is unconstrained mainly duringthe transition between outdoor and indoor spaces and whenobstacles are absent The Alowast-based routing graph mitigatesthe lack of obstacles thanks to a set of waypoints implementedaccording to human spatial cognition and to a weightednavigation mesh This allows building a realistic motionmodel thatmeets the requirements ofmap-aided localizationIndeed the proposed routing graph is well exploited becauseit gives prior knowledge about the pedestrianrsquos destinationand provides reliable measurements of walking directionsResults show that it is adequate for a seamless transitionbetween outdoor and indoor environments and for enhanc-ing the positioning accuracy even in large spaces Achievedaccuracies range from 3 to 5 meters and the drift is almostcancelled with the help of the routing graph though somemismatching errors due to uncalibrated walking distanceespecially while carrying the device in the swinging modemight induce important positioning errors Indeed properconditions of sky visibility and sufficient period of outdoorwalking are prerequisite for the step length calibration beforethe pedestrian reaches indoor space

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This researchworkwas done as part of theHappyHand (2015ndash2018) project which is funded by the 19th ldquoFonds UniqueInterministerielrdquo (French national RampD funding)

References

[1] J B Bancroft and G Lachapelle ldquoData fusion algorithms formultiple inertial measurement unitsrdquo Sensors vol 11 no 7 pp6771ndash6798 2011

[2] Y Hu L Sheng and S J Zhang ldquoDesign of Continuous IndoorNavigation System Based on INS and Wifirdquo Applied Mechanicsand Materials vol 303-306 pp 2046ndash2049 2013

[3] T Lee B Shin J H Lee et al ldquoAn Indoor Positioning SystemUsing Vision aided Advanced PDR technology without imageDB andwithmotion recognitionrdquo inProceedings of the ION2013Pacific PNT Meeting pp 490ndash497

[4] P Hafner T Moder M Wieser and T Bernoulli ldquoEvaluationof smartphone-based indoor positioning using different Bayesfiltersrdquo in Proceedings of the 2013 International Conference on

Indoor Positioning and Indoor Navigation IPIN 2013 October2013

[5] T Willemsen F Keller and H Sternberg ldquoA topologicalapproach with MEMS in smartphones based on routing-graphrdquo in Proceedings of the International Conference on IndoorPositioning and Indoor Navigation IPIN 2015 October 2015

[6] F T Alaoui D Betaille and V Renaudin ldquoA multi-hypothesisparticle filtering approach for pedestrian dead reckoningrdquo inProceedings of the 2016 International Conference on IndoorPositioning and Indoor Navigation IPIN 2016 October 2016

[7] T Fetzer F Ebner F Deinzer L Koping and M GrzegorzekldquoOn Monte Carlo smoothing in multi sensor indoor localisa-tionrdquo in Proceedings of the 2016 International Conference onIndoor Positioning and Indoor Navigation IPIN 2016 October2016

[8] M R F Mendonca H S Bernardino and R F Neto ldquoStealthypath planning using navigation meshesrdquo in Proceedings of the4th Brazilian Conference on Intelligent Systems BRACIS 2015pp 31ndash36 bra November 2015

[9] ldquoRecast Detail API documentation for the members declaredin Recasth sdot recastnavigationrecastnavigation6f5c9f9rdquoGitHub httpsgithubcomrecastnavigationcommit6f5c9f9-b82418efc44b85974e604095b95354ada

[10] I Afyouni C Ray and C Claramunt ldquoSpatial models forcontext-aware indoor navigation systems a surveyrdquo Journal ofSpatial Information Science vol 4 no 1 pp 85ndash123 2012

[11] W Van Toll A F Cook and R Geraerts ldquoNavigation meshesfor realistic multi-layered environmentsrdquo in Proceedings of the2011 IEEERSJ International Conference on Intelligent Robots andSystems Celebrating 50Years of Robotics IROSrsquo11 pp 3526ndash3532September 2011

[12] DETOUR ldquoGitHubrdquo httpsgithubcomrecastnavigationreca-stnavigationtreemasterDetourSource 2017

[13] F Mortari S Zlatanova L Liu and E Clementini ldquoImprovedgeometric network model (IGNM) a novel approach for deriv-ing connectivity graphs for indoor navigationrdquo ISPRS Annalsof Photogrammetry Remote Sensing and Spatial InformationSciences vol II-4 pp 45ndash51 2014

[14] N O Eraghi F Lopez-Colino A De Castro and J GarridoldquoPath length comparison in grid maps of planning algorithmsHCTNav A and Dijkstrardquo Design of Circuits and IntegratedSystems pp 1ndash6 2014

[15] C Gaisbauer and A U Frank ldquoWayfinding Model For Pedes-trian Navigationrdquo in Proceedings of the 11th AGILE InternationalConference on Geographic Information Science Spain 2008

[16] N Victor evaluation des deplacements pietons quotidiens Uni-versite de Lyon 2016

[17] L Yang and M Worboys ldquoGeneration of navigation graphs forindoor spacerdquo International Journal of Geographical InformationScience vol 29 no 10 pp 1737ndash1756 2015

[18] M Susi V Renaudin and G Lachapelle ldquoMotion moderecognition and step detection algorithms for mobile phoneusersrdquo Sensors vol 13 no 2 pp 1539ndash1562 2013

[19] V RenaudinM Susi andG Lachapelle ldquoStep length estimationusing handheld inertial sensorsrdquo Sensors (Switzerland) vol 12no 7 pp 8507ndash8525 2012

[20] V Renaudin and C Combettes ldquoMagnetic acceleration fieldsand gyroscope quaternion (MAGYQ)-based attitude estimationwith smartphone sensors for indoor pedestrian navigationrdquoSensors (Switzerland) vol 14 no 12 pp 22864ndash22890 2014

10 Wireless Communications and Mobile Computing

[21] ULISS httpwwwifsttar-geolocfrindexphpenequipment44-uliss

[22] ldquoGitHub - tomojitakasuRTKLIB_binrdquo httpsgithubcomto-mojitakasuRTKLIB_bin 2017

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Page 5: Pedestrian Dead Reckoning Navigation with the Help of -Based Routing …downloads.hindawi.com/journals/wcmc/2017/7951346.pdf · ResearchArticle Pedestrian Dead Reckoning Navigation

Wireless Communications and Mobile Computing 5

BA connectivity BB connectivity

$5 gt 0

ltL)>minus1gtL)>minus1

)>t = )>tminus1 )>t = )>tminus1 )>t = )>tminus1

$Jt = $5

t = tminus1 t = tminus1

$Jt = $5 minus L)>minus1$Jt = L)>

minus

t = minustminus1

($5 minus L)>minus1)

Figure 4 Dp computation according to DU

filter introduces white noise on each parameter to spread theparticles and optimize the graph exploration The particleposition is determined by the identifier of the graph edge andits curvilinear abscissa If the latter is above the edge lengthone of its connected edges will be explored The transitionmodel is

DU = (119904) sdot 120575 sdot 120572 + 119899119866 + Dp119905minus1 (3)

where DU is compared to the current edge length 119871 Id tocompute the curvilinear abscissa Dp (Figure 4) 119904 is thestep length and 119899119866 is the modelling error represented by aGaussian noise 120572 isin 0 1 2 and determines the numberof steps to be made by one particle It allows detectingovermisdetected steps

Figure 4 gives the transition test process that allowscomputing the curvilinear abscissa Dp when an exit nodeB is exceeded (ie DU gt 0) 119905 is the time index BABBconnectivity translates the search in the graph database inorder to determine if the edge being explored is connectedto the exceeded one via an A or a B node The same logic isadopted if an entry node A is exceeded (ie DU lt 0)

When GNSS positions are accurate enough they areused together with the graph headings in order to calibrateboth the step length model and PDR headings The graphheadings provide ameasurement of walking directions whichare compared to PDR headings taking into account potentialangular misalignment This is performed according to

119908119905= 119908119905minus1sdot exp(minus12 sdot (

1003816100381610038161003816120579PDR minus (120579Id + 120573)10038161003816100381610038162(120590120579Id2 + 120590120579PDR 2) + Δ1015840Σminus1Δ))

Δ = [119864119875 minus 119864GPS119873119875 minus 119873GPS]

Σminus1 = [[[[[

1(120590119864GPS2 + 120590map

119864

2) 00 1

(120590119873GPS2 + 120590map119873

2)]]]]]

(4)

Figure 5 Ubiquitous Localization Unit with Inertial Sensors andSatellites

where 119908119905 is the particle weight and 119864119875 and 119864GPS arerespectively the predicted East coordinate of a particle and itscorresponding GPS observation 119873119875 and 119873GPS are the sameparameters along the North direction Σminus1 is the weightingmatrix considering the graph and GPS position accuracies

Indoors some GNSS positions are still available but thelatter are generally unreliable and are rejected The rejectiontest is made by thresholding the signal to noise ratio (SNR)and the horizontal dilution of precision (HDOP) Only thegraph headings are exploited in the weighting process Theweighting equation indoors is

119908119905 = 119908119905minus1 sdot exp(minus12 times1003816100381610038161003816120579PDR minus (120579Id + 120573)10038161003816100381610038162(120590120579Id2 + 120590120579PDR 2) ) (5)

where 120579PDR is the walking direction estimated by the PDRalgorithms 120579Id is the predicted heading of the current path120573 is the heading misalignment between the path directionand the PDR estimate 120590120579Id is the standard deviation of thepath heading and 120590120579PDR is the standard deviation of the PDRwalking direction estimate

Once the particles have been weighted some of themare assigned low weights and become useless in the processResampling is performed to duplicate the particles with highweights and delete the others The particles amount is keptconstant and their weights equal after each update in order toexplore enough hypotheses of motion

6 Wireless Communications and Mobile Computing

Figure 6 Data collection by a pedestrian with ULISS unit in hand

5 Experiments

51 Data Collection with a HSGNSS and IMU in Hand Threehealthy volunteers (2 men 1 woman) collected data withULISS (Ubiquitous Localization Unit with Inertial Sensorsand Satellites) device (Figure 5) held in hand (Figure 6) Datawere collected in both outdoor and indoor environments fortwo different device carrying modes Two acquisitions weremade in a textingmode and one in both swinging and textingmodes The average duration of acquisitions was about 14minutes and the walking distance for each trial almost 15 km

Different scenarios were chosen to perform the experi-ments These will be explained in detail in Section 61 wherethe text is enhanced by figures

52 Input Data ULISS device [21] comprises 9 degreesof freedom inertial mobile unit a high sensitivity GNSS(HSGNSS) receiver and antenna a memory card and abattery It delivers measurements that are time-stampedin GPS time Inertial sensors and magnetometers providemeasurements at a 200Hz frequency

TheHSGNSS receiver operates in a standalone mode anddelivers positions in real time at a 5Hz frequency Deliveredpositions are time-stamped in GPS time and have metricaccuracies ranging from 2m up to 10m near buildings andtree shades In this work GNSS positions were interpolated atthe step frequency in order to be fusedwith the PDRestimatesof headings and step lengths

53 Reference Trajectories Besides collecting data withULISS device in hand all volunteers were equipped with anindependent GNSS receiver carried in their backpacks anda small antenna attached to their caps GNSS measurementswere then postprocessed in order to calculate reference trajec-tories by differential GNSS This was performed using RTK-LIB 242p12 software [22]Measurements from the embarkedGNSS receiver and from a nearby base station were used toperform relative GNSS positioning Obtained positions had

0 100 200 300 400East (m)

0

100

200

300N

orth

(m)

PDR trajectoryEstimated trajectory

Differential GNSS trajectory

Figure 7 Estimated trajectory (in red) The blue trajectory givesthe PDR position estimates and the green one gives the referencetrajectory (only texting)

decimetric accuracies up to several meters near buildings andother elevated features Afterwards thresholding was appliedto three parameters in order to reject some outlier positionestimates These parameters are first the number of visiblesatellites second the ratio factor of ambiguity resolutionand finally the horizontal dilution of precisionThe resultingpositions had precisions below 1m and were adequate foraccuracy assessment in this work

6 Results

61 Trajectory Analysis Figures 7 8 and 9 show the esti-mated trajectories (red) with the Alowast-generated routing graphand the particle filter described in Section 43 The blue

Wireless Communications and Mobile Computing 7

0 100 200 300 400East (m)

0

100

200

300

400

Nor

th (m

)

Figure 8 Trajectories for the second dataset (only texting)

0 100 200 300 400 500East (m)

0

100

200

300

400

500

Nor

th (m

)

Figure 9 Trajectories for the third dataset (swinging texting)

pattern corresponds to the PDR trajectory while the greenone is the reference trajectory obtained by differential GNSS

Figures 7 and 8 correspond to acquisitions performed inthe texting modeThe starting point for both acquisitions lieson the top right extremity of the trajectories Both subjectsmade a closed loop around the building with an intermediateoutdoor travel (bottom side of the figures) before reaching thestarting point back

For both acquisitions the travel distance seems to beoverestimated Yet heading determination is more accurate

for the first acquisition as the shape of the trajectory is morefaithful to the building structure The drift is higher for thesecond dataset and can be visually observed at the end of thePDR trajectory

Where the drift is most important the PDR trajectorypresents major inconsistencies with the map According toFigures 7 8 and 9 the drift has been corrected as theestimated trajectories are more compliant with the buildingstructure and with footpaths in outdoor space This has beenachieved thanks to the proposed particle filter and to anincreased conformitywith pedestrianmotion demonstratingthat the positioning accuracy for the texting scenario has beensignificantly enhanced using our approach

Figure 9 corresponds to data collected for both theswinging and texting modes The acquisition started at thebottom of the figure where the reference and the filteredtrajectories overlap Starting from this position the subjectentered the building and then went outside through theNorth-East building entrance This travel was made in theswinging mode The top right extremity of the trajectoryunderlines a U-turn before the subject entered the buildingback to reach the starting point This part of the travel wasperformed in the texting mode

Unlike the two first acquisitions there is a gap in the PDRtrajectory because the subject did not go around the buildingas the two first subjects didThis gap is retrieved in the filteredtrajectory

Obviously this scenario implies a less accurate headingdetermination and even an alteration in the travel distanceestimation In fact thewalking distance is underestimated forthe swinging mode (first part of the travel until the subjectreached the outdoor) and overestimated for the texting mode(second part of the travel until the ending point)These errorscan be noticed in the PDR trajectory

The filtered trajectory shows that the drift has beencorrected resulting in a shape that is more compliant withthe map and with the reference trajectory outdoors Yet thepositioning accuracy seems to be decreased as comparedwith the two first datasets Following section discusses theaccuracy of estimated trajectories

62 Error Computation In order to assess the accuracy ofour positioning method filtered positions were comparedto the reference positions interpolated at the step frequencyThe average plane error ranges from 4 to 5 meters for thethree datasets (Figure 10) Accuracy is dependent on thequality of the PDR trajectory Therefore computed errorsare more important for the second dataset considering thetexting scenario For the third acquisition that includes bothswinging and texting the accuracy is significantly decreasedand more outliers (precisions above 15m) which are givenby the red plus signs in Figure 10 are detected due tomismatching errors (ie choosing wrong edges of the graph)These errors are mainly due to uncalibrated PDR parametersand are discussed in the following sections

63 Heading Misalignment Estimation Figures 11 12 and 13show the estimated heading misalignment values for each

8 Wireless Communications and Mobile Computing

2 31Datasets

0

5

10

15

20

Erro

r (m

)

Figure 10 Plane errors for each dataset

2 4 6 80 10 12Time (min)

minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 11 Full texting scenario (1)

dataset For the first trial (Figure 11) angular misalignmentis comprised between minus10∘ and +10∘ and varies around anapproximate mean value of 0∘ According to this distributionthe angular difference between walking directions and thepointing direction of the device is minimal Hence appliedcorrections compensate only for the gyro driftwhich is ratherlogical regarding the texting mode scenario

For the second dataset (Figure 12) estimated headingmisalignment values are between minus15∘ and +12∘ They arenot equally distributed around 0∘ (eg between the 1st and2nd minutes the mean value is over 5∘) From this analysisnonnegligible hand motion can be assumed even if thesubject intended to perform the experiment in the textingmode

Figure 13 gives the estimated heading misalignment forthe third trial The first part (until 95min) of the travel wasperformed in the swingingmodeHence the estimated valuesvary significantly (minus15∘ to +20∘) For the second part of theplot (gt10min) angular misalignment variations occur withlower magnitudes comprising plusmn5∘ over a mean value of 0∘which reflects the texting mode of the acquisition

64 Step Length Model Calibration In this paper a scalefactor was introduced on the step length in order to calibratethe walking distance though due to degraded GNSS signaland to short walking periods in outdoor space the scale

10 12 14Time (m)

2 4 6 80minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 12 Full texting scenario (2)

minus20minus15minus10minus5

05

101520

Hea

ding

misa

lignm

ent(

∘ )

6 8 141210 16 182 40

Time (min)

Figure 13 Swinging-texting scenario

factor was not calibrated In fact regular and accurate GNSSpositions are needed through straight line travels as cited inour introduction in order to calibrate the walking distanceThese conditions were not verified during our experiments

As a result only corrected headings were relied on in theselection process Therefore distance calibration occurredonly at junctions of the graph when a change in heading wasdetected This explains the fact that the accuracy of filteredtrajectories is still enhanced and compliance with the mapimproved

Though the uncalibrated walking distances caused somemismatching errors because direction change was detectedtoo late or too early in the process this can be noticed in theNorthern part of Figure 9 Indeed while the pedestrian wasintending to exit the building the filtered trajectory indicatesthat he was walking towards a corridor This happened fortwo reasons First the real trajectory is quite unusual interms of pedestrian behavior In fact there is a change inheading (observed in the PDR trajectory) that is independentof space configuration invalidating the assumptions thatallowed constructing our navigation network Second theuncalibrated walking distance prevented the particles fromreaching outdoor space at the right time Another mismatch-ing error occurred at the middle of the building (between300m and 400m North) because direction change wasdetected too late due to uncalibrated walking distance Lateron the particle filter corrected for this error and convergedover the right corridor thanks to the particle dispersion overthe graph and to the multihypothesis approach

Wireless Communications and Mobile Computing 9

7 Conclusions

Amap-aided PDR approach where a routing graph is used asmotion model has been proposed Main contribution of thispaper is Alowast algorithm adaptation to elaborate a pedestriannetwork that is capable of cancelling the gyro drift and themisalignment between the device orientation and thewalkingdirection even in large spaces These are GNSS-deprived andobstacle-free areas where the limitations of map-aided PDRalgorithms are most important In fact widespread map-aided PDR approaches do not compensate for these errorswhen pedestrian motion is unconstrained mainly duringthe transition between outdoor and indoor spaces and whenobstacles are absent The Alowast-based routing graph mitigatesthe lack of obstacles thanks to a set of waypoints implementedaccording to human spatial cognition and to a weightednavigation mesh This allows building a realistic motionmodel thatmeets the requirements ofmap-aided localizationIndeed the proposed routing graph is well exploited becauseit gives prior knowledge about the pedestrianrsquos destinationand provides reliable measurements of walking directionsResults show that it is adequate for a seamless transitionbetween outdoor and indoor environments and for enhanc-ing the positioning accuracy even in large spaces Achievedaccuracies range from 3 to 5 meters and the drift is almostcancelled with the help of the routing graph though somemismatching errors due to uncalibrated walking distanceespecially while carrying the device in the swinging modemight induce important positioning errors Indeed properconditions of sky visibility and sufficient period of outdoorwalking are prerequisite for the step length calibration beforethe pedestrian reaches indoor space

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This researchworkwas done as part of theHappyHand (2015ndash2018) project which is funded by the 19th ldquoFonds UniqueInterministerielrdquo (French national RampD funding)

References

[1] J B Bancroft and G Lachapelle ldquoData fusion algorithms formultiple inertial measurement unitsrdquo Sensors vol 11 no 7 pp6771ndash6798 2011

[2] Y Hu L Sheng and S J Zhang ldquoDesign of Continuous IndoorNavigation System Based on INS and Wifirdquo Applied Mechanicsand Materials vol 303-306 pp 2046ndash2049 2013

[3] T Lee B Shin J H Lee et al ldquoAn Indoor Positioning SystemUsing Vision aided Advanced PDR technology without imageDB andwithmotion recognitionrdquo inProceedings of the ION2013Pacific PNT Meeting pp 490ndash497

[4] P Hafner T Moder M Wieser and T Bernoulli ldquoEvaluationof smartphone-based indoor positioning using different Bayesfiltersrdquo in Proceedings of the 2013 International Conference on

Indoor Positioning and Indoor Navigation IPIN 2013 October2013

[5] T Willemsen F Keller and H Sternberg ldquoA topologicalapproach with MEMS in smartphones based on routing-graphrdquo in Proceedings of the International Conference on IndoorPositioning and Indoor Navigation IPIN 2015 October 2015

[6] F T Alaoui D Betaille and V Renaudin ldquoA multi-hypothesisparticle filtering approach for pedestrian dead reckoningrdquo inProceedings of the 2016 International Conference on IndoorPositioning and Indoor Navigation IPIN 2016 October 2016

[7] T Fetzer F Ebner F Deinzer L Koping and M GrzegorzekldquoOn Monte Carlo smoothing in multi sensor indoor localisa-tionrdquo in Proceedings of the 2016 International Conference onIndoor Positioning and Indoor Navigation IPIN 2016 October2016

[8] M R F Mendonca H S Bernardino and R F Neto ldquoStealthypath planning using navigation meshesrdquo in Proceedings of the4th Brazilian Conference on Intelligent Systems BRACIS 2015pp 31ndash36 bra November 2015

[9] ldquoRecast Detail API documentation for the members declaredin Recasth sdot recastnavigationrecastnavigation6f5c9f9rdquoGitHub httpsgithubcomrecastnavigationcommit6f5c9f9-b82418efc44b85974e604095b95354ada

[10] I Afyouni C Ray and C Claramunt ldquoSpatial models forcontext-aware indoor navigation systems a surveyrdquo Journal ofSpatial Information Science vol 4 no 1 pp 85ndash123 2012

[11] W Van Toll A F Cook and R Geraerts ldquoNavigation meshesfor realistic multi-layered environmentsrdquo in Proceedings of the2011 IEEERSJ International Conference on Intelligent Robots andSystems Celebrating 50Years of Robotics IROSrsquo11 pp 3526ndash3532September 2011

[12] DETOUR ldquoGitHubrdquo httpsgithubcomrecastnavigationreca-stnavigationtreemasterDetourSource 2017

[13] F Mortari S Zlatanova L Liu and E Clementini ldquoImprovedgeometric network model (IGNM) a novel approach for deriv-ing connectivity graphs for indoor navigationrdquo ISPRS Annalsof Photogrammetry Remote Sensing and Spatial InformationSciences vol II-4 pp 45ndash51 2014

[14] N O Eraghi F Lopez-Colino A De Castro and J GarridoldquoPath length comparison in grid maps of planning algorithmsHCTNav A and Dijkstrardquo Design of Circuits and IntegratedSystems pp 1ndash6 2014

[15] C Gaisbauer and A U Frank ldquoWayfinding Model For Pedes-trian Navigationrdquo in Proceedings of the 11th AGILE InternationalConference on Geographic Information Science Spain 2008

[16] N Victor evaluation des deplacements pietons quotidiens Uni-versite de Lyon 2016

[17] L Yang and M Worboys ldquoGeneration of navigation graphs forindoor spacerdquo International Journal of Geographical InformationScience vol 29 no 10 pp 1737ndash1756 2015

[18] M Susi V Renaudin and G Lachapelle ldquoMotion moderecognition and step detection algorithms for mobile phoneusersrdquo Sensors vol 13 no 2 pp 1539ndash1562 2013

[19] V RenaudinM Susi andG Lachapelle ldquoStep length estimationusing handheld inertial sensorsrdquo Sensors (Switzerland) vol 12no 7 pp 8507ndash8525 2012

[20] V Renaudin and C Combettes ldquoMagnetic acceleration fieldsand gyroscope quaternion (MAGYQ)-based attitude estimationwith smartphone sensors for indoor pedestrian navigationrdquoSensors (Switzerland) vol 14 no 12 pp 22864ndash22890 2014

10 Wireless Communications and Mobile Computing

[21] ULISS httpwwwifsttar-geolocfrindexphpenequipment44-uliss

[22] ldquoGitHub - tomojitakasuRTKLIB_binrdquo httpsgithubcomto-mojitakasuRTKLIB_bin 2017

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Page 6: Pedestrian Dead Reckoning Navigation with the Help of -Based Routing …downloads.hindawi.com/journals/wcmc/2017/7951346.pdf · ResearchArticle Pedestrian Dead Reckoning Navigation

6 Wireless Communications and Mobile Computing

Figure 6 Data collection by a pedestrian with ULISS unit in hand

5 Experiments

51 Data Collection with a HSGNSS and IMU in Hand Threehealthy volunteers (2 men 1 woman) collected data withULISS (Ubiquitous Localization Unit with Inertial Sensorsand Satellites) device (Figure 5) held in hand (Figure 6) Datawere collected in both outdoor and indoor environments fortwo different device carrying modes Two acquisitions weremade in a textingmode and one in both swinging and textingmodes The average duration of acquisitions was about 14minutes and the walking distance for each trial almost 15 km

Different scenarios were chosen to perform the experi-ments These will be explained in detail in Section 61 wherethe text is enhanced by figures

52 Input Data ULISS device [21] comprises 9 degreesof freedom inertial mobile unit a high sensitivity GNSS(HSGNSS) receiver and antenna a memory card and abattery It delivers measurements that are time-stampedin GPS time Inertial sensors and magnetometers providemeasurements at a 200Hz frequency

TheHSGNSS receiver operates in a standalone mode anddelivers positions in real time at a 5Hz frequency Deliveredpositions are time-stamped in GPS time and have metricaccuracies ranging from 2m up to 10m near buildings andtree shades In this work GNSS positions were interpolated atthe step frequency in order to be fusedwith the PDRestimatesof headings and step lengths

53 Reference Trajectories Besides collecting data withULISS device in hand all volunteers were equipped with anindependent GNSS receiver carried in their backpacks anda small antenna attached to their caps GNSS measurementswere then postprocessed in order to calculate reference trajec-tories by differential GNSS This was performed using RTK-LIB 242p12 software [22]Measurements from the embarkedGNSS receiver and from a nearby base station were used toperform relative GNSS positioning Obtained positions had

0 100 200 300 400East (m)

0

100

200

300N

orth

(m)

PDR trajectoryEstimated trajectory

Differential GNSS trajectory

Figure 7 Estimated trajectory (in red) The blue trajectory givesthe PDR position estimates and the green one gives the referencetrajectory (only texting)

decimetric accuracies up to several meters near buildings andother elevated features Afterwards thresholding was appliedto three parameters in order to reject some outlier positionestimates These parameters are first the number of visiblesatellites second the ratio factor of ambiguity resolutionand finally the horizontal dilution of precisionThe resultingpositions had precisions below 1m and were adequate foraccuracy assessment in this work

6 Results

61 Trajectory Analysis Figures 7 8 and 9 show the esti-mated trajectories (red) with the Alowast-generated routing graphand the particle filter described in Section 43 The blue

Wireless Communications and Mobile Computing 7

0 100 200 300 400East (m)

0

100

200

300

400

Nor

th (m

)

Figure 8 Trajectories for the second dataset (only texting)

0 100 200 300 400 500East (m)

0

100

200

300

400

500

Nor

th (m

)

Figure 9 Trajectories for the third dataset (swinging texting)

pattern corresponds to the PDR trajectory while the greenone is the reference trajectory obtained by differential GNSS

Figures 7 and 8 correspond to acquisitions performed inthe texting modeThe starting point for both acquisitions lieson the top right extremity of the trajectories Both subjectsmade a closed loop around the building with an intermediateoutdoor travel (bottom side of the figures) before reaching thestarting point back

For both acquisitions the travel distance seems to beoverestimated Yet heading determination is more accurate

for the first acquisition as the shape of the trajectory is morefaithful to the building structure The drift is higher for thesecond dataset and can be visually observed at the end of thePDR trajectory

Where the drift is most important the PDR trajectorypresents major inconsistencies with the map According toFigures 7 8 and 9 the drift has been corrected as theestimated trajectories are more compliant with the buildingstructure and with footpaths in outdoor space This has beenachieved thanks to the proposed particle filter and to anincreased conformitywith pedestrianmotion demonstratingthat the positioning accuracy for the texting scenario has beensignificantly enhanced using our approach

Figure 9 corresponds to data collected for both theswinging and texting modes The acquisition started at thebottom of the figure where the reference and the filteredtrajectories overlap Starting from this position the subjectentered the building and then went outside through theNorth-East building entrance This travel was made in theswinging mode The top right extremity of the trajectoryunderlines a U-turn before the subject entered the buildingback to reach the starting point This part of the travel wasperformed in the texting mode

Unlike the two first acquisitions there is a gap in the PDRtrajectory because the subject did not go around the buildingas the two first subjects didThis gap is retrieved in the filteredtrajectory

Obviously this scenario implies a less accurate headingdetermination and even an alteration in the travel distanceestimation In fact thewalking distance is underestimated forthe swinging mode (first part of the travel until the subjectreached the outdoor) and overestimated for the texting mode(second part of the travel until the ending point)These errorscan be noticed in the PDR trajectory

The filtered trajectory shows that the drift has beencorrected resulting in a shape that is more compliant withthe map and with the reference trajectory outdoors Yet thepositioning accuracy seems to be decreased as comparedwith the two first datasets Following section discusses theaccuracy of estimated trajectories

62 Error Computation In order to assess the accuracy ofour positioning method filtered positions were comparedto the reference positions interpolated at the step frequencyThe average plane error ranges from 4 to 5 meters for thethree datasets (Figure 10) Accuracy is dependent on thequality of the PDR trajectory Therefore computed errorsare more important for the second dataset considering thetexting scenario For the third acquisition that includes bothswinging and texting the accuracy is significantly decreasedand more outliers (precisions above 15m) which are givenby the red plus signs in Figure 10 are detected due tomismatching errors (ie choosing wrong edges of the graph)These errors are mainly due to uncalibrated PDR parametersand are discussed in the following sections

63 Heading Misalignment Estimation Figures 11 12 and 13show the estimated heading misalignment values for each

8 Wireless Communications and Mobile Computing

2 31Datasets

0

5

10

15

20

Erro

r (m

)

Figure 10 Plane errors for each dataset

2 4 6 80 10 12Time (min)

minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 11 Full texting scenario (1)

dataset For the first trial (Figure 11) angular misalignmentis comprised between minus10∘ and +10∘ and varies around anapproximate mean value of 0∘ According to this distributionthe angular difference between walking directions and thepointing direction of the device is minimal Hence appliedcorrections compensate only for the gyro driftwhich is ratherlogical regarding the texting mode scenario

For the second dataset (Figure 12) estimated headingmisalignment values are between minus15∘ and +12∘ They arenot equally distributed around 0∘ (eg between the 1st and2nd minutes the mean value is over 5∘) From this analysisnonnegligible hand motion can be assumed even if thesubject intended to perform the experiment in the textingmode

Figure 13 gives the estimated heading misalignment forthe third trial The first part (until 95min) of the travel wasperformed in the swingingmodeHence the estimated valuesvary significantly (minus15∘ to +20∘) For the second part of theplot (gt10min) angular misalignment variations occur withlower magnitudes comprising plusmn5∘ over a mean value of 0∘which reflects the texting mode of the acquisition

64 Step Length Model Calibration In this paper a scalefactor was introduced on the step length in order to calibratethe walking distance though due to degraded GNSS signaland to short walking periods in outdoor space the scale

10 12 14Time (m)

2 4 6 80minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 12 Full texting scenario (2)

minus20minus15minus10minus5

05

101520

Hea

ding

misa

lignm

ent(

∘ )

6 8 141210 16 182 40

Time (min)

Figure 13 Swinging-texting scenario

factor was not calibrated In fact regular and accurate GNSSpositions are needed through straight line travels as cited inour introduction in order to calibrate the walking distanceThese conditions were not verified during our experiments

As a result only corrected headings were relied on in theselection process Therefore distance calibration occurredonly at junctions of the graph when a change in heading wasdetected This explains the fact that the accuracy of filteredtrajectories is still enhanced and compliance with the mapimproved

Though the uncalibrated walking distances caused somemismatching errors because direction change was detectedtoo late or too early in the process this can be noticed in theNorthern part of Figure 9 Indeed while the pedestrian wasintending to exit the building the filtered trajectory indicatesthat he was walking towards a corridor This happened fortwo reasons First the real trajectory is quite unusual interms of pedestrian behavior In fact there is a change inheading (observed in the PDR trajectory) that is independentof space configuration invalidating the assumptions thatallowed constructing our navigation network Second theuncalibrated walking distance prevented the particles fromreaching outdoor space at the right time Another mismatch-ing error occurred at the middle of the building (between300m and 400m North) because direction change wasdetected too late due to uncalibrated walking distance Lateron the particle filter corrected for this error and convergedover the right corridor thanks to the particle dispersion overthe graph and to the multihypothesis approach

Wireless Communications and Mobile Computing 9

7 Conclusions

Amap-aided PDR approach where a routing graph is used asmotion model has been proposed Main contribution of thispaper is Alowast algorithm adaptation to elaborate a pedestriannetwork that is capable of cancelling the gyro drift and themisalignment between the device orientation and thewalkingdirection even in large spaces These are GNSS-deprived andobstacle-free areas where the limitations of map-aided PDRalgorithms are most important In fact widespread map-aided PDR approaches do not compensate for these errorswhen pedestrian motion is unconstrained mainly duringthe transition between outdoor and indoor spaces and whenobstacles are absent The Alowast-based routing graph mitigatesthe lack of obstacles thanks to a set of waypoints implementedaccording to human spatial cognition and to a weightednavigation mesh This allows building a realistic motionmodel thatmeets the requirements ofmap-aided localizationIndeed the proposed routing graph is well exploited becauseit gives prior knowledge about the pedestrianrsquos destinationand provides reliable measurements of walking directionsResults show that it is adequate for a seamless transitionbetween outdoor and indoor environments and for enhanc-ing the positioning accuracy even in large spaces Achievedaccuracies range from 3 to 5 meters and the drift is almostcancelled with the help of the routing graph though somemismatching errors due to uncalibrated walking distanceespecially while carrying the device in the swinging modemight induce important positioning errors Indeed properconditions of sky visibility and sufficient period of outdoorwalking are prerequisite for the step length calibration beforethe pedestrian reaches indoor space

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This researchworkwas done as part of theHappyHand (2015ndash2018) project which is funded by the 19th ldquoFonds UniqueInterministerielrdquo (French national RampD funding)

References

[1] J B Bancroft and G Lachapelle ldquoData fusion algorithms formultiple inertial measurement unitsrdquo Sensors vol 11 no 7 pp6771ndash6798 2011

[2] Y Hu L Sheng and S J Zhang ldquoDesign of Continuous IndoorNavigation System Based on INS and Wifirdquo Applied Mechanicsand Materials vol 303-306 pp 2046ndash2049 2013

[3] T Lee B Shin J H Lee et al ldquoAn Indoor Positioning SystemUsing Vision aided Advanced PDR technology without imageDB andwithmotion recognitionrdquo inProceedings of the ION2013Pacific PNT Meeting pp 490ndash497

[4] P Hafner T Moder M Wieser and T Bernoulli ldquoEvaluationof smartphone-based indoor positioning using different Bayesfiltersrdquo in Proceedings of the 2013 International Conference on

Indoor Positioning and Indoor Navigation IPIN 2013 October2013

[5] T Willemsen F Keller and H Sternberg ldquoA topologicalapproach with MEMS in smartphones based on routing-graphrdquo in Proceedings of the International Conference on IndoorPositioning and Indoor Navigation IPIN 2015 October 2015

[6] F T Alaoui D Betaille and V Renaudin ldquoA multi-hypothesisparticle filtering approach for pedestrian dead reckoningrdquo inProceedings of the 2016 International Conference on IndoorPositioning and Indoor Navigation IPIN 2016 October 2016

[7] T Fetzer F Ebner F Deinzer L Koping and M GrzegorzekldquoOn Monte Carlo smoothing in multi sensor indoor localisa-tionrdquo in Proceedings of the 2016 International Conference onIndoor Positioning and Indoor Navigation IPIN 2016 October2016

[8] M R F Mendonca H S Bernardino and R F Neto ldquoStealthypath planning using navigation meshesrdquo in Proceedings of the4th Brazilian Conference on Intelligent Systems BRACIS 2015pp 31ndash36 bra November 2015

[9] ldquoRecast Detail API documentation for the members declaredin Recasth sdot recastnavigationrecastnavigation6f5c9f9rdquoGitHub httpsgithubcomrecastnavigationcommit6f5c9f9-b82418efc44b85974e604095b95354ada

[10] I Afyouni C Ray and C Claramunt ldquoSpatial models forcontext-aware indoor navigation systems a surveyrdquo Journal ofSpatial Information Science vol 4 no 1 pp 85ndash123 2012

[11] W Van Toll A F Cook and R Geraerts ldquoNavigation meshesfor realistic multi-layered environmentsrdquo in Proceedings of the2011 IEEERSJ International Conference on Intelligent Robots andSystems Celebrating 50Years of Robotics IROSrsquo11 pp 3526ndash3532September 2011

[12] DETOUR ldquoGitHubrdquo httpsgithubcomrecastnavigationreca-stnavigationtreemasterDetourSource 2017

[13] F Mortari S Zlatanova L Liu and E Clementini ldquoImprovedgeometric network model (IGNM) a novel approach for deriv-ing connectivity graphs for indoor navigationrdquo ISPRS Annalsof Photogrammetry Remote Sensing and Spatial InformationSciences vol II-4 pp 45ndash51 2014

[14] N O Eraghi F Lopez-Colino A De Castro and J GarridoldquoPath length comparison in grid maps of planning algorithmsHCTNav A and Dijkstrardquo Design of Circuits and IntegratedSystems pp 1ndash6 2014

[15] C Gaisbauer and A U Frank ldquoWayfinding Model For Pedes-trian Navigationrdquo in Proceedings of the 11th AGILE InternationalConference on Geographic Information Science Spain 2008

[16] N Victor evaluation des deplacements pietons quotidiens Uni-versite de Lyon 2016

[17] L Yang and M Worboys ldquoGeneration of navigation graphs forindoor spacerdquo International Journal of Geographical InformationScience vol 29 no 10 pp 1737ndash1756 2015

[18] M Susi V Renaudin and G Lachapelle ldquoMotion moderecognition and step detection algorithms for mobile phoneusersrdquo Sensors vol 13 no 2 pp 1539ndash1562 2013

[19] V RenaudinM Susi andG Lachapelle ldquoStep length estimationusing handheld inertial sensorsrdquo Sensors (Switzerland) vol 12no 7 pp 8507ndash8525 2012

[20] V Renaudin and C Combettes ldquoMagnetic acceleration fieldsand gyroscope quaternion (MAGYQ)-based attitude estimationwith smartphone sensors for indoor pedestrian navigationrdquoSensors (Switzerland) vol 14 no 12 pp 22864ndash22890 2014

10 Wireless Communications and Mobile Computing

[21] ULISS httpwwwifsttar-geolocfrindexphpenequipment44-uliss

[22] ldquoGitHub - tomojitakasuRTKLIB_binrdquo httpsgithubcomto-mojitakasuRTKLIB_bin 2017

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Page 7: Pedestrian Dead Reckoning Navigation with the Help of -Based Routing …downloads.hindawi.com/journals/wcmc/2017/7951346.pdf · ResearchArticle Pedestrian Dead Reckoning Navigation

Wireless Communications and Mobile Computing 7

0 100 200 300 400East (m)

0

100

200

300

400

Nor

th (m

)

Figure 8 Trajectories for the second dataset (only texting)

0 100 200 300 400 500East (m)

0

100

200

300

400

500

Nor

th (m

)

Figure 9 Trajectories for the third dataset (swinging texting)

pattern corresponds to the PDR trajectory while the greenone is the reference trajectory obtained by differential GNSS

Figures 7 and 8 correspond to acquisitions performed inthe texting modeThe starting point for both acquisitions lieson the top right extremity of the trajectories Both subjectsmade a closed loop around the building with an intermediateoutdoor travel (bottom side of the figures) before reaching thestarting point back

For both acquisitions the travel distance seems to beoverestimated Yet heading determination is more accurate

for the first acquisition as the shape of the trajectory is morefaithful to the building structure The drift is higher for thesecond dataset and can be visually observed at the end of thePDR trajectory

Where the drift is most important the PDR trajectorypresents major inconsistencies with the map According toFigures 7 8 and 9 the drift has been corrected as theestimated trajectories are more compliant with the buildingstructure and with footpaths in outdoor space This has beenachieved thanks to the proposed particle filter and to anincreased conformitywith pedestrianmotion demonstratingthat the positioning accuracy for the texting scenario has beensignificantly enhanced using our approach

Figure 9 corresponds to data collected for both theswinging and texting modes The acquisition started at thebottom of the figure where the reference and the filteredtrajectories overlap Starting from this position the subjectentered the building and then went outside through theNorth-East building entrance This travel was made in theswinging mode The top right extremity of the trajectoryunderlines a U-turn before the subject entered the buildingback to reach the starting point This part of the travel wasperformed in the texting mode

Unlike the two first acquisitions there is a gap in the PDRtrajectory because the subject did not go around the buildingas the two first subjects didThis gap is retrieved in the filteredtrajectory

Obviously this scenario implies a less accurate headingdetermination and even an alteration in the travel distanceestimation In fact thewalking distance is underestimated forthe swinging mode (first part of the travel until the subjectreached the outdoor) and overestimated for the texting mode(second part of the travel until the ending point)These errorscan be noticed in the PDR trajectory

The filtered trajectory shows that the drift has beencorrected resulting in a shape that is more compliant withthe map and with the reference trajectory outdoors Yet thepositioning accuracy seems to be decreased as comparedwith the two first datasets Following section discusses theaccuracy of estimated trajectories

62 Error Computation In order to assess the accuracy ofour positioning method filtered positions were comparedto the reference positions interpolated at the step frequencyThe average plane error ranges from 4 to 5 meters for thethree datasets (Figure 10) Accuracy is dependent on thequality of the PDR trajectory Therefore computed errorsare more important for the second dataset considering thetexting scenario For the third acquisition that includes bothswinging and texting the accuracy is significantly decreasedand more outliers (precisions above 15m) which are givenby the red plus signs in Figure 10 are detected due tomismatching errors (ie choosing wrong edges of the graph)These errors are mainly due to uncalibrated PDR parametersand are discussed in the following sections

63 Heading Misalignment Estimation Figures 11 12 and 13show the estimated heading misalignment values for each

8 Wireless Communications and Mobile Computing

2 31Datasets

0

5

10

15

20

Erro

r (m

)

Figure 10 Plane errors for each dataset

2 4 6 80 10 12Time (min)

minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 11 Full texting scenario (1)

dataset For the first trial (Figure 11) angular misalignmentis comprised between minus10∘ and +10∘ and varies around anapproximate mean value of 0∘ According to this distributionthe angular difference between walking directions and thepointing direction of the device is minimal Hence appliedcorrections compensate only for the gyro driftwhich is ratherlogical regarding the texting mode scenario

For the second dataset (Figure 12) estimated headingmisalignment values are between minus15∘ and +12∘ They arenot equally distributed around 0∘ (eg between the 1st and2nd minutes the mean value is over 5∘) From this analysisnonnegligible hand motion can be assumed even if thesubject intended to perform the experiment in the textingmode

Figure 13 gives the estimated heading misalignment forthe third trial The first part (until 95min) of the travel wasperformed in the swingingmodeHence the estimated valuesvary significantly (minus15∘ to +20∘) For the second part of theplot (gt10min) angular misalignment variations occur withlower magnitudes comprising plusmn5∘ over a mean value of 0∘which reflects the texting mode of the acquisition

64 Step Length Model Calibration In this paper a scalefactor was introduced on the step length in order to calibratethe walking distance though due to degraded GNSS signaland to short walking periods in outdoor space the scale

10 12 14Time (m)

2 4 6 80minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 12 Full texting scenario (2)

minus20minus15minus10minus5

05

101520

Hea

ding

misa

lignm

ent(

∘ )

6 8 141210 16 182 40

Time (min)

Figure 13 Swinging-texting scenario

factor was not calibrated In fact regular and accurate GNSSpositions are needed through straight line travels as cited inour introduction in order to calibrate the walking distanceThese conditions were not verified during our experiments

As a result only corrected headings were relied on in theselection process Therefore distance calibration occurredonly at junctions of the graph when a change in heading wasdetected This explains the fact that the accuracy of filteredtrajectories is still enhanced and compliance with the mapimproved

Though the uncalibrated walking distances caused somemismatching errors because direction change was detectedtoo late or too early in the process this can be noticed in theNorthern part of Figure 9 Indeed while the pedestrian wasintending to exit the building the filtered trajectory indicatesthat he was walking towards a corridor This happened fortwo reasons First the real trajectory is quite unusual interms of pedestrian behavior In fact there is a change inheading (observed in the PDR trajectory) that is independentof space configuration invalidating the assumptions thatallowed constructing our navigation network Second theuncalibrated walking distance prevented the particles fromreaching outdoor space at the right time Another mismatch-ing error occurred at the middle of the building (between300m and 400m North) because direction change wasdetected too late due to uncalibrated walking distance Lateron the particle filter corrected for this error and convergedover the right corridor thanks to the particle dispersion overthe graph and to the multihypothesis approach

Wireless Communications and Mobile Computing 9

7 Conclusions

Amap-aided PDR approach where a routing graph is used asmotion model has been proposed Main contribution of thispaper is Alowast algorithm adaptation to elaborate a pedestriannetwork that is capable of cancelling the gyro drift and themisalignment between the device orientation and thewalkingdirection even in large spaces These are GNSS-deprived andobstacle-free areas where the limitations of map-aided PDRalgorithms are most important In fact widespread map-aided PDR approaches do not compensate for these errorswhen pedestrian motion is unconstrained mainly duringthe transition between outdoor and indoor spaces and whenobstacles are absent The Alowast-based routing graph mitigatesthe lack of obstacles thanks to a set of waypoints implementedaccording to human spatial cognition and to a weightednavigation mesh This allows building a realistic motionmodel thatmeets the requirements ofmap-aided localizationIndeed the proposed routing graph is well exploited becauseit gives prior knowledge about the pedestrianrsquos destinationand provides reliable measurements of walking directionsResults show that it is adequate for a seamless transitionbetween outdoor and indoor environments and for enhanc-ing the positioning accuracy even in large spaces Achievedaccuracies range from 3 to 5 meters and the drift is almostcancelled with the help of the routing graph though somemismatching errors due to uncalibrated walking distanceespecially while carrying the device in the swinging modemight induce important positioning errors Indeed properconditions of sky visibility and sufficient period of outdoorwalking are prerequisite for the step length calibration beforethe pedestrian reaches indoor space

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This researchworkwas done as part of theHappyHand (2015ndash2018) project which is funded by the 19th ldquoFonds UniqueInterministerielrdquo (French national RampD funding)

References

[1] J B Bancroft and G Lachapelle ldquoData fusion algorithms formultiple inertial measurement unitsrdquo Sensors vol 11 no 7 pp6771ndash6798 2011

[2] Y Hu L Sheng and S J Zhang ldquoDesign of Continuous IndoorNavigation System Based on INS and Wifirdquo Applied Mechanicsand Materials vol 303-306 pp 2046ndash2049 2013

[3] T Lee B Shin J H Lee et al ldquoAn Indoor Positioning SystemUsing Vision aided Advanced PDR technology without imageDB andwithmotion recognitionrdquo inProceedings of the ION2013Pacific PNT Meeting pp 490ndash497

[4] P Hafner T Moder M Wieser and T Bernoulli ldquoEvaluationof smartphone-based indoor positioning using different Bayesfiltersrdquo in Proceedings of the 2013 International Conference on

Indoor Positioning and Indoor Navigation IPIN 2013 October2013

[5] T Willemsen F Keller and H Sternberg ldquoA topologicalapproach with MEMS in smartphones based on routing-graphrdquo in Proceedings of the International Conference on IndoorPositioning and Indoor Navigation IPIN 2015 October 2015

[6] F T Alaoui D Betaille and V Renaudin ldquoA multi-hypothesisparticle filtering approach for pedestrian dead reckoningrdquo inProceedings of the 2016 International Conference on IndoorPositioning and Indoor Navigation IPIN 2016 October 2016

[7] T Fetzer F Ebner F Deinzer L Koping and M GrzegorzekldquoOn Monte Carlo smoothing in multi sensor indoor localisa-tionrdquo in Proceedings of the 2016 International Conference onIndoor Positioning and Indoor Navigation IPIN 2016 October2016

[8] M R F Mendonca H S Bernardino and R F Neto ldquoStealthypath planning using navigation meshesrdquo in Proceedings of the4th Brazilian Conference on Intelligent Systems BRACIS 2015pp 31ndash36 bra November 2015

[9] ldquoRecast Detail API documentation for the members declaredin Recasth sdot recastnavigationrecastnavigation6f5c9f9rdquoGitHub httpsgithubcomrecastnavigationcommit6f5c9f9-b82418efc44b85974e604095b95354ada

[10] I Afyouni C Ray and C Claramunt ldquoSpatial models forcontext-aware indoor navigation systems a surveyrdquo Journal ofSpatial Information Science vol 4 no 1 pp 85ndash123 2012

[11] W Van Toll A F Cook and R Geraerts ldquoNavigation meshesfor realistic multi-layered environmentsrdquo in Proceedings of the2011 IEEERSJ International Conference on Intelligent Robots andSystems Celebrating 50Years of Robotics IROSrsquo11 pp 3526ndash3532September 2011

[12] DETOUR ldquoGitHubrdquo httpsgithubcomrecastnavigationreca-stnavigationtreemasterDetourSource 2017

[13] F Mortari S Zlatanova L Liu and E Clementini ldquoImprovedgeometric network model (IGNM) a novel approach for deriv-ing connectivity graphs for indoor navigationrdquo ISPRS Annalsof Photogrammetry Remote Sensing and Spatial InformationSciences vol II-4 pp 45ndash51 2014

[14] N O Eraghi F Lopez-Colino A De Castro and J GarridoldquoPath length comparison in grid maps of planning algorithmsHCTNav A and Dijkstrardquo Design of Circuits and IntegratedSystems pp 1ndash6 2014

[15] C Gaisbauer and A U Frank ldquoWayfinding Model For Pedes-trian Navigationrdquo in Proceedings of the 11th AGILE InternationalConference on Geographic Information Science Spain 2008

[16] N Victor evaluation des deplacements pietons quotidiens Uni-versite de Lyon 2016

[17] L Yang and M Worboys ldquoGeneration of navigation graphs forindoor spacerdquo International Journal of Geographical InformationScience vol 29 no 10 pp 1737ndash1756 2015

[18] M Susi V Renaudin and G Lachapelle ldquoMotion moderecognition and step detection algorithms for mobile phoneusersrdquo Sensors vol 13 no 2 pp 1539ndash1562 2013

[19] V RenaudinM Susi andG Lachapelle ldquoStep length estimationusing handheld inertial sensorsrdquo Sensors (Switzerland) vol 12no 7 pp 8507ndash8525 2012

[20] V Renaudin and C Combettes ldquoMagnetic acceleration fieldsand gyroscope quaternion (MAGYQ)-based attitude estimationwith smartphone sensors for indoor pedestrian navigationrdquoSensors (Switzerland) vol 14 no 12 pp 22864ndash22890 2014

10 Wireless Communications and Mobile Computing

[21] ULISS httpwwwifsttar-geolocfrindexphpenequipment44-uliss

[22] ldquoGitHub - tomojitakasuRTKLIB_binrdquo httpsgithubcomto-mojitakasuRTKLIB_bin 2017

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Page 8: Pedestrian Dead Reckoning Navigation with the Help of -Based Routing …downloads.hindawi.com/journals/wcmc/2017/7951346.pdf · ResearchArticle Pedestrian Dead Reckoning Navigation

8 Wireless Communications and Mobile Computing

2 31Datasets

0

5

10

15

20

Erro

r (m

)

Figure 10 Plane errors for each dataset

2 4 6 80 10 12Time (min)

minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 11 Full texting scenario (1)

dataset For the first trial (Figure 11) angular misalignmentis comprised between minus10∘ and +10∘ and varies around anapproximate mean value of 0∘ According to this distributionthe angular difference between walking directions and thepointing direction of the device is minimal Hence appliedcorrections compensate only for the gyro driftwhich is ratherlogical regarding the texting mode scenario

For the second dataset (Figure 12) estimated headingmisalignment values are between minus15∘ and +12∘ They arenot equally distributed around 0∘ (eg between the 1st and2nd minutes the mean value is over 5∘) From this analysisnonnegligible hand motion can be assumed even if thesubject intended to perform the experiment in the textingmode

Figure 13 gives the estimated heading misalignment forthe third trial The first part (until 95min) of the travel wasperformed in the swingingmodeHence the estimated valuesvary significantly (minus15∘ to +20∘) For the second part of theplot (gt10min) angular misalignment variations occur withlower magnitudes comprising plusmn5∘ over a mean value of 0∘which reflects the texting mode of the acquisition

64 Step Length Model Calibration In this paper a scalefactor was introduced on the step length in order to calibratethe walking distance though due to degraded GNSS signaland to short walking periods in outdoor space the scale

10 12 14Time (m)

2 4 6 80minus15

minus10

minus5

0

5

10

15

Hea

ding

misa

lignm

ent(

∘ )

Figure 12 Full texting scenario (2)

minus20minus15minus10minus5

05

101520

Hea

ding

misa

lignm

ent(

∘ )

6 8 141210 16 182 40

Time (min)

Figure 13 Swinging-texting scenario

factor was not calibrated In fact regular and accurate GNSSpositions are needed through straight line travels as cited inour introduction in order to calibrate the walking distanceThese conditions were not verified during our experiments

As a result only corrected headings were relied on in theselection process Therefore distance calibration occurredonly at junctions of the graph when a change in heading wasdetected This explains the fact that the accuracy of filteredtrajectories is still enhanced and compliance with the mapimproved

Though the uncalibrated walking distances caused somemismatching errors because direction change was detectedtoo late or too early in the process this can be noticed in theNorthern part of Figure 9 Indeed while the pedestrian wasintending to exit the building the filtered trajectory indicatesthat he was walking towards a corridor This happened fortwo reasons First the real trajectory is quite unusual interms of pedestrian behavior In fact there is a change inheading (observed in the PDR trajectory) that is independentof space configuration invalidating the assumptions thatallowed constructing our navigation network Second theuncalibrated walking distance prevented the particles fromreaching outdoor space at the right time Another mismatch-ing error occurred at the middle of the building (between300m and 400m North) because direction change wasdetected too late due to uncalibrated walking distance Lateron the particle filter corrected for this error and convergedover the right corridor thanks to the particle dispersion overthe graph and to the multihypothesis approach

Wireless Communications and Mobile Computing 9

7 Conclusions

Amap-aided PDR approach where a routing graph is used asmotion model has been proposed Main contribution of thispaper is Alowast algorithm adaptation to elaborate a pedestriannetwork that is capable of cancelling the gyro drift and themisalignment between the device orientation and thewalkingdirection even in large spaces These are GNSS-deprived andobstacle-free areas where the limitations of map-aided PDRalgorithms are most important In fact widespread map-aided PDR approaches do not compensate for these errorswhen pedestrian motion is unconstrained mainly duringthe transition between outdoor and indoor spaces and whenobstacles are absent The Alowast-based routing graph mitigatesthe lack of obstacles thanks to a set of waypoints implementedaccording to human spatial cognition and to a weightednavigation mesh This allows building a realistic motionmodel thatmeets the requirements ofmap-aided localizationIndeed the proposed routing graph is well exploited becauseit gives prior knowledge about the pedestrianrsquos destinationand provides reliable measurements of walking directionsResults show that it is adequate for a seamless transitionbetween outdoor and indoor environments and for enhanc-ing the positioning accuracy even in large spaces Achievedaccuracies range from 3 to 5 meters and the drift is almostcancelled with the help of the routing graph though somemismatching errors due to uncalibrated walking distanceespecially while carrying the device in the swinging modemight induce important positioning errors Indeed properconditions of sky visibility and sufficient period of outdoorwalking are prerequisite for the step length calibration beforethe pedestrian reaches indoor space

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This researchworkwas done as part of theHappyHand (2015ndash2018) project which is funded by the 19th ldquoFonds UniqueInterministerielrdquo (French national RampD funding)

References

[1] J B Bancroft and G Lachapelle ldquoData fusion algorithms formultiple inertial measurement unitsrdquo Sensors vol 11 no 7 pp6771ndash6798 2011

[2] Y Hu L Sheng and S J Zhang ldquoDesign of Continuous IndoorNavigation System Based on INS and Wifirdquo Applied Mechanicsand Materials vol 303-306 pp 2046ndash2049 2013

[3] T Lee B Shin J H Lee et al ldquoAn Indoor Positioning SystemUsing Vision aided Advanced PDR technology without imageDB andwithmotion recognitionrdquo inProceedings of the ION2013Pacific PNT Meeting pp 490ndash497

[4] P Hafner T Moder M Wieser and T Bernoulli ldquoEvaluationof smartphone-based indoor positioning using different Bayesfiltersrdquo in Proceedings of the 2013 International Conference on

Indoor Positioning and Indoor Navigation IPIN 2013 October2013

[5] T Willemsen F Keller and H Sternberg ldquoA topologicalapproach with MEMS in smartphones based on routing-graphrdquo in Proceedings of the International Conference on IndoorPositioning and Indoor Navigation IPIN 2015 October 2015

[6] F T Alaoui D Betaille and V Renaudin ldquoA multi-hypothesisparticle filtering approach for pedestrian dead reckoningrdquo inProceedings of the 2016 International Conference on IndoorPositioning and Indoor Navigation IPIN 2016 October 2016

[7] T Fetzer F Ebner F Deinzer L Koping and M GrzegorzekldquoOn Monte Carlo smoothing in multi sensor indoor localisa-tionrdquo in Proceedings of the 2016 International Conference onIndoor Positioning and Indoor Navigation IPIN 2016 October2016

[8] M R F Mendonca H S Bernardino and R F Neto ldquoStealthypath planning using navigation meshesrdquo in Proceedings of the4th Brazilian Conference on Intelligent Systems BRACIS 2015pp 31ndash36 bra November 2015

[9] ldquoRecast Detail API documentation for the members declaredin Recasth sdot recastnavigationrecastnavigation6f5c9f9rdquoGitHub httpsgithubcomrecastnavigationcommit6f5c9f9-b82418efc44b85974e604095b95354ada

[10] I Afyouni C Ray and C Claramunt ldquoSpatial models forcontext-aware indoor navigation systems a surveyrdquo Journal ofSpatial Information Science vol 4 no 1 pp 85ndash123 2012

[11] W Van Toll A F Cook and R Geraerts ldquoNavigation meshesfor realistic multi-layered environmentsrdquo in Proceedings of the2011 IEEERSJ International Conference on Intelligent Robots andSystems Celebrating 50Years of Robotics IROSrsquo11 pp 3526ndash3532September 2011

[12] DETOUR ldquoGitHubrdquo httpsgithubcomrecastnavigationreca-stnavigationtreemasterDetourSource 2017

[13] F Mortari S Zlatanova L Liu and E Clementini ldquoImprovedgeometric network model (IGNM) a novel approach for deriv-ing connectivity graphs for indoor navigationrdquo ISPRS Annalsof Photogrammetry Remote Sensing and Spatial InformationSciences vol II-4 pp 45ndash51 2014

[14] N O Eraghi F Lopez-Colino A De Castro and J GarridoldquoPath length comparison in grid maps of planning algorithmsHCTNav A and Dijkstrardquo Design of Circuits and IntegratedSystems pp 1ndash6 2014

[15] C Gaisbauer and A U Frank ldquoWayfinding Model For Pedes-trian Navigationrdquo in Proceedings of the 11th AGILE InternationalConference on Geographic Information Science Spain 2008

[16] N Victor evaluation des deplacements pietons quotidiens Uni-versite de Lyon 2016

[17] L Yang and M Worboys ldquoGeneration of navigation graphs forindoor spacerdquo International Journal of Geographical InformationScience vol 29 no 10 pp 1737ndash1756 2015

[18] M Susi V Renaudin and G Lachapelle ldquoMotion moderecognition and step detection algorithms for mobile phoneusersrdquo Sensors vol 13 no 2 pp 1539ndash1562 2013

[19] V RenaudinM Susi andG Lachapelle ldquoStep length estimationusing handheld inertial sensorsrdquo Sensors (Switzerland) vol 12no 7 pp 8507ndash8525 2012

[20] V Renaudin and C Combettes ldquoMagnetic acceleration fieldsand gyroscope quaternion (MAGYQ)-based attitude estimationwith smartphone sensors for indoor pedestrian navigationrdquoSensors (Switzerland) vol 14 no 12 pp 22864ndash22890 2014

10 Wireless Communications and Mobile Computing

[21] ULISS httpwwwifsttar-geolocfrindexphpenequipment44-uliss

[22] ldquoGitHub - tomojitakasuRTKLIB_binrdquo httpsgithubcomto-mojitakasuRTKLIB_bin 2017

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Page 9: Pedestrian Dead Reckoning Navigation with the Help of -Based Routing …downloads.hindawi.com/journals/wcmc/2017/7951346.pdf · ResearchArticle Pedestrian Dead Reckoning Navigation

Wireless Communications and Mobile Computing 9

7 Conclusions

Amap-aided PDR approach where a routing graph is used asmotion model has been proposed Main contribution of thispaper is Alowast algorithm adaptation to elaborate a pedestriannetwork that is capable of cancelling the gyro drift and themisalignment between the device orientation and thewalkingdirection even in large spaces These are GNSS-deprived andobstacle-free areas where the limitations of map-aided PDRalgorithms are most important In fact widespread map-aided PDR approaches do not compensate for these errorswhen pedestrian motion is unconstrained mainly duringthe transition between outdoor and indoor spaces and whenobstacles are absent The Alowast-based routing graph mitigatesthe lack of obstacles thanks to a set of waypoints implementedaccording to human spatial cognition and to a weightednavigation mesh This allows building a realistic motionmodel thatmeets the requirements ofmap-aided localizationIndeed the proposed routing graph is well exploited becauseit gives prior knowledge about the pedestrianrsquos destinationand provides reliable measurements of walking directionsResults show that it is adequate for a seamless transitionbetween outdoor and indoor environments and for enhanc-ing the positioning accuracy even in large spaces Achievedaccuracies range from 3 to 5 meters and the drift is almostcancelled with the help of the routing graph though somemismatching errors due to uncalibrated walking distanceespecially while carrying the device in the swinging modemight induce important positioning errors Indeed properconditions of sky visibility and sufficient period of outdoorwalking are prerequisite for the step length calibration beforethe pedestrian reaches indoor space

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This researchworkwas done as part of theHappyHand (2015ndash2018) project which is funded by the 19th ldquoFonds UniqueInterministerielrdquo (French national RampD funding)

References

[1] J B Bancroft and G Lachapelle ldquoData fusion algorithms formultiple inertial measurement unitsrdquo Sensors vol 11 no 7 pp6771ndash6798 2011

[2] Y Hu L Sheng and S J Zhang ldquoDesign of Continuous IndoorNavigation System Based on INS and Wifirdquo Applied Mechanicsand Materials vol 303-306 pp 2046ndash2049 2013

[3] T Lee B Shin J H Lee et al ldquoAn Indoor Positioning SystemUsing Vision aided Advanced PDR technology without imageDB andwithmotion recognitionrdquo inProceedings of the ION2013Pacific PNT Meeting pp 490ndash497

[4] P Hafner T Moder M Wieser and T Bernoulli ldquoEvaluationof smartphone-based indoor positioning using different Bayesfiltersrdquo in Proceedings of the 2013 International Conference on

Indoor Positioning and Indoor Navigation IPIN 2013 October2013

[5] T Willemsen F Keller and H Sternberg ldquoA topologicalapproach with MEMS in smartphones based on routing-graphrdquo in Proceedings of the International Conference on IndoorPositioning and Indoor Navigation IPIN 2015 October 2015

[6] F T Alaoui D Betaille and V Renaudin ldquoA multi-hypothesisparticle filtering approach for pedestrian dead reckoningrdquo inProceedings of the 2016 International Conference on IndoorPositioning and Indoor Navigation IPIN 2016 October 2016

[7] T Fetzer F Ebner F Deinzer L Koping and M GrzegorzekldquoOn Monte Carlo smoothing in multi sensor indoor localisa-tionrdquo in Proceedings of the 2016 International Conference onIndoor Positioning and Indoor Navigation IPIN 2016 October2016

[8] M R F Mendonca H S Bernardino and R F Neto ldquoStealthypath planning using navigation meshesrdquo in Proceedings of the4th Brazilian Conference on Intelligent Systems BRACIS 2015pp 31ndash36 bra November 2015

[9] ldquoRecast Detail API documentation for the members declaredin Recasth sdot recastnavigationrecastnavigation6f5c9f9rdquoGitHub httpsgithubcomrecastnavigationcommit6f5c9f9-b82418efc44b85974e604095b95354ada

[10] I Afyouni C Ray and C Claramunt ldquoSpatial models forcontext-aware indoor navigation systems a surveyrdquo Journal ofSpatial Information Science vol 4 no 1 pp 85ndash123 2012

[11] W Van Toll A F Cook and R Geraerts ldquoNavigation meshesfor realistic multi-layered environmentsrdquo in Proceedings of the2011 IEEERSJ International Conference on Intelligent Robots andSystems Celebrating 50Years of Robotics IROSrsquo11 pp 3526ndash3532September 2011

[12] DETOUR ldquoGitHubrdquo httpsgithubcomrecastnavigationreca-stnavigationtreemasterDetourSource 2017

[13] F Mortari S Zlatanova L Liu and E Clementini ldquoImprovedgeometric network model (IGNM) a novel approach for deriv-ing connectivity graphs for indoor navigationrdquo ISPRS Annalsof Photogrammetry Remote Sensing and Spatial InformationSciences vol II-4 pp 45ndash51 2014

[14] N O Eraghi F Lopez-Colino A De Castro and J GarridoldquoPath length comparison in grid maps of planning algorithmsHCTNav A and Dijkstrardquo Design of Circuits and IntegratedSystems pp 1ndash6 2014

[15] C Gaisbauer and A U Frank ldquoWayfinding Model For Pedes-trian Navigationrdquo in Proceedings of the 11th AGILE InternationalConference on Geographic Information Science Spain 2008

[16] N Victor evaluation des deplacements pietons quotidiens Uni-versite de Lyon 2016

[17] L Yang and M Worboys ldquoGeneration of navigation graphs forindoor spacerdquo International Journal of Geographical InformationScience vol 29 no 10 pp 1737ndash1756 2015

[18] M Susi V Renaudin and G Lachapelle ldquoMotion moderecognition and step detection algorithms for mobile phoneusersrdquo Sensors vol 13 no 2 pp 1539ndash1562 2013

[19] V RenaudinM Susi andG Lachapelle ldquoStep length estimationusing handheld inertial sensorsrdquo Sensors (Switzerland) vol 12no 7 pp 8507ndash8525 2012

[20] V Renaudin and C Combettes ldquoMagnetic acceleration fieldsand gyroscope quaternion (MAGYQ)-based attitude estimationwith smartphone sensors for indoor pedestrian navigationrdquoSensors (Switzerland) vol 14 no 12 pp 22864ndash22890 2014

10 Wireless Communications and Mobile Computing

[21] ULISS httpwwwifsttar-geolocfrindexphpenequipment44-uliss

[22] ldquoGitHub - tomojitakasuRTKLIB_binrdquo httpsgithubcomto-mojitakasuRTKLIB_bin 2017

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Page 10: Pedestrian Dead Reckoning Navigation with the Help of -Based Routing …downloads.hindawi.com/journals/wcmc/2017/7951346.pdf · ResearchArticle Pedestrian Dead Reckoning Navigation

10 Wireless Communications and Mobile Computing

[21] ULISS httpwwwifsttar-geolocfrindexphpenequipment44-uliss

[22] ldquoGitHub - tomojitakasuRTKLIB_binrdquo httpsgithubcomto-mojitakasuRTKLIB_bin 2017

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Page 11: Pedestrian Dead Reckoning Navigation with the Help of -Based Routing …downloads.hindawi.com/journals/wcmc/2017/7951346.pdf · ResearchArticle Pedestrian Dead Reckoning Navigation

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of