multi-people tracking across multiple cameras

8/12/2019 MULTI-PEOPLE TRACKING ACROSS MULTIPLE CAMERAS

1/11

23

Multi-

people

tracking

across

multiple

cameras

Faouzi Alaya Cheikh1, Sajib Kumar Saha2, Victoria Rudakova3, Peng Wang4

1,2Gjvik University College, Norway3cole Nationale Des Ponts et Chausses, France

4Univeristy Jean Monnet, [email protected], [email protected],

[email protected], [email protected]

Abstract. Multi-target tracking (MTT) is an active and challenging researchtopic. Many different approaches to MTT problem exist, yet there are stillfew satisfactory methods of solving multi-target occlusion problem, whichoften appears in multi-target tracking task. The application of multi camerasin most existing researches for multi-target occlusion requires cameracalibration parameters in advance, which is not practical in the case ofoutdoor video surveillance. Most of the proposed solutions for this problemrequire camera calibration parameters that make them impractical for outdoorvideo surveillance applications. To address this problem we propose in thispaper a probabilistic approach, the foremost consideration of which is toreduce the dependency on camera calibration for multiple cameracollaboration. More robustness on target representation and object trackinghas been ensured by combining multiple cues such as border information ofthe object with color histogram, while Gale-Shapley algorithm (GSA) hasbeen used for finding the stable matching between objects of two or morecamera views. Efficient tracking of object ensures proficient recognition oftarget depicting parameters (i.e. apparent color, height and width informationof the object) as a consequence provides better camera collaboration. Initialsimulation results prove the validity of the proposed approach.

Keywords: Surveillance, multi-camera tracking, multi-target tracking,particle filter.

1 INTRODUCTION

Object tracking has received tremendous

attention in the video processing community dueto its numerous potential applications in videosurveillance, human activity analysis, trafficmonitoring, insect / animal tracking and so on.Over the last couple of years, many algorithms

and results have been presented with regard tothe problem of object tracking and recently thefocus of the community has been concentratedon multi-target tracking (MTT) that requiresdetermining the number as well as the dynamicsof the targets. Even though single objecttracking is considered as a mature field of

InternationalJournalonNewComputerArchitecturesandTheirApplications(IJNCAA)2(1):2333

TheSocietyofDigitalInformationandWirelessCommunications,2012(ISSN:22209085)


2/11

24

research, due to a combination of severalfactors, reliable multi-target tracking stillremains a challenging domain of research. Theunderlying difficulties behind multi-targettracking are founded mostly upon the apparent

similarity of targets and then multi-targetocclusion. MTT for targets with distinctiveappearance is comparatively easier, since it canbe solved reasonably well by using multipleindependent single-target trackers. But MTT fortargets with similar appearance, such aspedestrians in crowded scenes, is a much moredifficult task. In addition, MTT must deal withmulti-target occlusion, namely, the tracker mustseparate the targets and assign them correctlabels. This is a particular problem in a multi-

object scenario, where object-object interactionsand occlusions are likely to occur. It is notpossible to reliably maintain a targets identityusing single clue such as continuity ofappearance or motion; besides, computationalcomplexity also plays an important role, asmany applications require the tracking to bedone in real time. All these issues make targettracking or multi-object tracking a sturdy taskeven now.

The proposed methodology in this paper is an

enhancement of the scheme proposed in [33],where the main concentration was to use colorand size information of the object for multicamera collaboration in order to reduce thedependency on camera calibration. Here morerobustness on object tracking has been ensuredby combining the centroid informationcalculated from the detected border of theobject.

2 LITERATURE SURVEY

Most of the early works for MTT were basedon monocular video [1]. A widely acceptedapproach that addresses many problems of MTTis based on a joint state-space representation andinfers the joint data association [2, 3]. A binaryvariable has been used by MacCormick and

Blake [4] to identify foreground objects andthen a probabilistic exclusion principle has beenused to penalize the hypothesis where twoobjects occlude. In [5], the likelihood iscalculated by enumerating all possible

association hypotheses. Zhao and Nevatia [6, 7]used a different 3D shape model and jointlikelihood for multiple human segmentation andtracking. Tao et al. [8] proposed a sampling-based multiple-target tracking method usingbackground subtraction. Khan et al. in [9]proposed an MCMC-based particle filter whichuses a Markov random field to model motioninteraction. Smith et al. presented a differentMCMC-based particle filter to estimate themulti-object configuration [10]. McKenna et al.

[11] presented a color-based system for trackinggroups of people. Adaptive color models areused to provide qualitative estimates of depthordering during occlusion. Although the abovesolutions, which are based on a centralizedprocess, can handle the problem of multi-targetocclusion in principle, they require atremendous computational cost due to thecomplexity introduced by the highdimensionality of the joint-state representationwhich grows exponentially in terms of thenumber of objects tracked.

Several researchers proposed decentralizedsolutions for multi-target tracking. In [12] themulti-object occlusion problem has been solvedby using multiple cameras where the camerasare separated widely in order to obtain visualinformation from wide viewing angles andoffers a possible 3D solution. The system needsto pass the subjects identities across cameraswhen the identities are lost in a certain view bymatching subjects across camera views.Therefore, the system needs to match subjects inconsecutive frames of a single camera and alsomatch subjects across cameras in order tomaintain subject identities in as many camerasas possible. Although this cross-viewcorrespondence is related to wide baselinestereo matching, traditional correlation based




3/11

25

methods fail due to the large difference inviewpoint [13].

Yu and Wu [14] and Wu et al. [15] usedmultiple collaborative trackers for MTTmodeled by a Markov random network. This

approach demonstrates the efficiency of thedecentralized method. The decentralizedapproach was carried further by Qu et al. [16]who proposed an interactively distributed multi-object tracking framework using a magnetic-inertia potential model.

However, using multiple cameras raises manyadditional challenges. The most criticaldifficulties presented by multi-camera trackingare to establish a consistent labelcorrespondence of the same target among the

different views and to integrate the informationfrom different camera views for tracking, whichis robust to significant and persistent occlusion.

Many existing approaches address the labelcorrespondence problem by using differenttechniques such as feature matching [17, 18],camera calibration and/or 3D environmentmodel [18, 19], and motion-trajectory alignment[20]. A kalman filter based approached has beenproposed in [13] for tracking multiple object inindoor environment. In addition with apparentcolor, apparent height, landmark modality andhomography, epipolar geometry has been usedfor multi-camera cooperation. Recently, Qu etal. have presented a distributed Bayesianframework for multiple-target tracking usingmultiple collaborative cameras. The distributedBayesian framework avoids the computationalcomplexity inherent in centralized methods thatrely on joint-state representation and joint dataassociation. Epipolar geometry has been usedfor multi-camera collaboration. However, a lotof dependency on epipolar geometry makes thatapproach impractical for outdoor videosurveillance application, since weneed to knowthe angle of view for each camera with respectto others accurately; which can change veryfrequently for outdoor surveillance cameras dueto environmental maladies.

3 PROPOSED FRAMEWORK

3.1 BAYESIAN SEQUENTIAL

ESTIMATION

The Bayesian sequential estimation helps toestimate a posterior distribution noted as:|: or its marginal |: , where: includes a set of all states up to time tand: - set of all the observations up to time taccordingly. The evolution of the statesequences , of a target is given byequation 1; and the observation model is givenby equation 2: , (1) , (2) Where : and : can be both linear and nonlinear,and and are sequences of i.d.d. processnoise and measurement noise respectively; atthe same time and are their dimensions.

In Bayesian context, the tracking problem canbe considered as recursive calculation of somebelief degree in the state at time step t, giventhe data observation : . That is, we need toconstruct a pdf |:. It is assumed that theinitial state of the system (also called prior) is

given. | (3) Then, the posterior distribution |:canbe calculated by the following next two stepsgiven in equations (4) and (5).

Prediction: |: ||: (4)Update:

|:||:

|:

||: ||:

(5)

In equation (5) the denominator is anormalization constant and it depends on thelikelihood function | - as described byequation (2). But presented recursive




4/11

26

propagation is just a conceptual solution andcannot be applied in practice.

Monte-Carlo simulation [26] with sequentialimportance sampling (SIS) technique, allows usto approximate equations 4 and 5 in the discreet

form:Prediction:

|: | (6)Update:|: (7)

Where || | ,: as wellas

1

Nonetheless in order to avoid degeneracy(one of the common problems with SIS) re-sampling of the particles need to be done.Therefore, the main idea is to update theparticles which almost do not make anycontribution to the approximation to |:and pay an attention on the more promisingparticles. It generates a new set, by re-sampling with replacement times from an approximate discreetrepresentation of |:, so that and weights must be reset to (we denote it ).For this project we used the re-sampling schemeproposed in [27].

State estimate:

The mean state (mean particle, meanshape

or weighted sum of particles) has been used.

(8)

3.2 MODELING OF DENSITIES

Modeling the densities for SIS is an importanttask, at the same time plays a critical role in thesystem performance. The models that have beenused in our proposed method are different from[1], and the following subsections will cover theconstruction of each of those densities.

Target representationSince ellipses can provide more precise shape

information than a bounding box [28], simple5D parametric ellipse was used in order todecrease computational costs, and at the sametime, it is sufficient to represent the trackingresults. Therefore, one state

is defined by

, , , , , where i is targetindex, t - current time, (cx, cy) - coordinates ofthe center of ellipse, (a, b) major and minoraxis, and - orientation angle, as shown inFigure 1.

Fig. 1.Ellipse representation for describing a humanbody.

Local observation modelAs a local likelihood,, single cue -

color histogram like [29] has been used. Colormodels are obtained in the Hue-Saturation-Value (HSV) color space in order to decouple




5/11

27

chromatic information from shading effects. Asfor the bin numbers, we have used Nh=Ns = 8and Nv=4 (Hue, Saturation and Value binnumbers respectively).

State dynamics modelThe better prediction results were achieved by

using motion information. It was decided to useLucas-Kanade (LK) optical flow algorithm [30]when we dealt with single independent tracker.However, this method works well for the onlycase when targets do not interact with eachother.

Interaction modelWhen targets start to interact with each other

(i.e., occlusion), we cannot rely on the motionbased proposal anymore. Hence magnetic-repulsion inertia re-weighting scheme [29]established on random based proposal was usedinstead of the re-weighting scheme [31], since[29] gives better result than [31] in ourexperiment.

3.3 IMPROVED DETECTION OF

MOVING OBJECT

Another important point to consider is how toinclude prediction of a, b and the ellipseparameters for stating dynamics model. Untilnow it was possible to predict only the centercoordinates (cx, cy) of the ellipse based on theLK optical flow [30] calculation. At the sametime parameters a, band propagate according

to the random-based prediction which is notvery effective sometimes since it can lead toover-expanding or over-squeezing of the ellipsebounds. Hence to overcome this problem,objects contour information has beenconsidered for calculating a, band .

Detecting the contour information consists ofthree phases like in [32]; in the first phase

gradient information has been extracted since itis much less affected by the quantization noiseand abrupt change of illumination. Sobeloperators have been used instead of Robertsoperators or Prewitt operators, as they are

generally less computationally expensive andmore suitable for hardware realizations [32].

,,|,, | |,, | (9)

where 1 0 1 2 0 2 1 0 1and

14 1 2 1 0 0 0 1 2 1Where, f(x, y, t) denotes a pixel of a gray-

scale image, fE (x, y, t) denotes the gradient ofthe pixel, HX and HY denote the horizontaland the vertical transform matrix,respectively.

Color image has been converted to grayimage using the following equation

,, 0.587 , , 0.114 ,, 0.29 , , (10)In the second phase three frame differencing

scheme [32] has been used instead of commonlyused two frame differencing method for betterdetection of the contours of moving objects. Theoperation is detailed in the following equation,, |,, , , 1| |,, , , 1| (11)

The background contour in an imageresulted from such a gradient-based three-frame differencing can be largely eliminated,whereas the contours of moving objects can beremarkably enhanced.

In the third phase a step of eight-connectivitytesting is appended with three-framedifferencing for the sake of noise reduction.




6/11

28

Once the contours of moving objects havebeen detected using the above mentionedprocedure, then (cx,cy) detected by particle filterhas been used to calculate , , . A point(

,

), that is the farthest from (cx,cy) within

a certain perimeter defined by (cx, cy, a, b+ ,) has been calculated. Then =tanhas been calculated. Now another farthest point( , ), which makes 90 5 angle withrespect to line joining (cx,cy) and , hasbeen calculated.

Once ( , ) has been detected then(, ) has been calculated in the directionthat makes 180 5angle with respect to linejoining (cx,cy) and

,

. Likewise

( , ) has also been calculated in thedirection that makes 180 5 angle withrespect to line joining (cx,cy) and , .Now , , , has been calculated asfollowing

, , (12) , , (13)

where

, is the length of the line joining

(cx,cy) and , Likewise ,

(14)Now the ellipse that tracks the object(s) has

been defined by the parameters, , , , where 0.4 0.6 , 0.4 0.6 , 0.4 0.6 , 0.4 0.6 and 0.4

0.6

If for any reason (shadow, occlusion etc; Fig.7), the calculated becomes unreliable while is being calculated, the weighting factors(0.5, 0.5) have been used instead of (0.4, 0.6);and the last used (from known previous frame)value of will be used; if the value of istaken from a frame which is not more than 3frames before the current frame. These factors

are (0.6, 0.4) if the value of has been takenfrom a frame which is more than 3 frames butless than 9 frames before from the currentframe; otherwise (1, 0). Same is true for

,

,

,

.

3.4 MULTI-CAMERA DATA FUSION

In order to correctly associate correspondingtargets (assign the same identity of objectsirrespective of camera views) Gale-Shapleyalgorithm (GSA) [24] has been applied, whichuses the color, height and width information ofthe detected object in 2 camera views. Eachtime a new object appears in the camera view(s),its normalized color histogram (normalized byapparent height and width; we will refer it asinitialized histogram) has been stored in thecamera view(s), at the same time thecorresponding labeling of objects with respect toone reference camera has been done by usingGSA that uses the histogram distance betweenobjects of two different camera views togenerate the preference lists for objects. For thesystem using more than 2 cameras that labelingshould be done for all the cameras with respectto one reference camera. After labeling has beendone each camera can work independently andtrack individually. Once an occlusion occurs inthe reference camera, the histogram distance ofobjects has been calculated by considering thenormalized color histograms of objects (of thatframe) with respect to its initialized histogramon that camera view; and then occluded objecthas been identified by comparing the histogramdistances of objects that are interacting. If an (ormore) occlusion has been detected, the idea is tofigure out which camera can be suggested forthat object. For that perspective the referencecamera will look for that object in the cameraright to it. If that object is also occluded therethen the reference camera will look for thatobject in the camera just left to it to checkwhether that object is occluded on that cameraview or no and so on.




7/11

29

4 EXPERIMENTAL RESULT

4.1 EXPERIMENTAL SETUP

Two USB Logitec web cameras have beenused. All the video sequences with people wererecorded from these cameras. For the purpose tomake synchronized recordings, softwaredeveloped by Nils Fjelds at HIG was used fortwo cameras. For the initialization of the targetsthe code implemented in [31] has been used.

Several experimental camera set-ups wereorganized, where people number, their activities(hand-shaking, walking and occluding eachother) and also illumination (daylight, artificialroom light) varied. Original videos wererecorded with the frame size of 640 480. Forour tests it was decided to decrease the framesize to 320 240 to lower computational costsneeded for processing one frame. For all thesequences, we have used 50 particles for eachtarget.

4.2 EXPERIMENTAL ANALYSIS

For tracking individual object the modelproposed here that combines border information

of the object with particle filter informationgives better result than using only particle filterinformation (see figure 2, 3; in figure 2 thetracker wraps considerable amount ofbackground data). Even though the extraprocessing for combining border takes someextra time (15 ms per frame [32]), that time isvery negligible compared to the time taken byparticle filter but this operation ensures goodtracking and overall good performance of theproposed framework.

The proposed methodology guaranteesquality tracking and gives constant labeling ofobjects irrespective of camera views. For all thetested video sequences (15 video sequences, 2camera views with overlapping field of view)the tracking and labeling of objects wereappropriate, unless the objects were too far(more than 8 meters) from the camera.

Fig. 2.Tracking of objects using partilcle filter only.

Fig. 3. Tracking of objects using augmented particle filteraugmented with border information.




8/11

30

Fig. 4.Video sequences obtained from two camera viewsand tracking of objects (where the color of ellipses

ensures that the objects are labeled correctly irrespectiveof camera views).

Fig. 5.Video sequences obtained from two camera viewsand tracking of objects.

Fig. 6.Video sequences where the yellow tracker loses itstrack.

#Camera:1Frame:241

#Camera:2Frame:241

#Camera:1Frame:169

#Camera:2Frame:169




9/11

31

Fig. 7.Video sequence where detected border of theobject is not efficient.

5 CONCLUSION AND DISCUSSION

Provided that epipolar geometry is not asapposite solution for multi-camera data fusion, amethodology for multi-camera data fusion hasbeen proposed based on Gale-ShapleyAlgorithm. The idea to use GSA here is not tocheck the correctness of the coordinates oftracked objects in the image coordinate system(like most of the systems do), but to maintaincorrect identities of targets among the differentcamera views and gives more reliable results.

While GSA based on color histogram(normalized with respect to height and width) isapplied, it is very important that the trackedellipse should contain mostly the object.Experiment results indicate that the proposedmethodology ensures better tracking of objects.As a consequence, the overall accuracy of GSAincreases as well. At the same time the efficient

tracking of objects by the proposedmethodology guarantees better identification ofocclusion in camera view(s). The framework isgeneral and can be easily adapted for trackingany class of objects.

Due to the design of the approach itself, thereis a limitation of the performance. Like anycolor based detection methods, for objects ofapparent similarity (objects wearing similarcloths and having same height and width) thesystem may not work properly. Besides, it hasbeen observed from the experiment that whenthe object becomes too far from the camera(more than 8 meters) the tracker loses its track(see figure 6). Even though HSV color spacehas been used to ignore the lightness effect on

objects with respect to the camera distance, itdoes not give very robust result in reality. Hencemore investigation on advanced color spacescan be done in future.

There is room for improvement in themodeling of the appearance of the targets. Allellipse paramenters are generated based on thecentroid of the ellipse. To get more precisedescription of targets, geometric formula can beapplied for more robust detection of centroids infuture.

REFERENCES

[1] Qu, W., Schonfeld, D., & Mohamed, M.:Distributed Bayesian multiple-target tracking incrowded environments using multiplecollaborative cameras. EURASIP Journal onApplied Signal Processing, Issue 1, 2121 (2007)

[2] Y. Bar-Shalom and A. G. Jammer: Tracking andData Association. Academic Press, San Diego,Calif, USA (1998)

[3] C. Hue, J.-P. L. Cadre, and P. Perez: SequentialMonte Carlo methods for multiple target trackingand data fusion. IEEE Transactions on SignalProcessing, vol. 50, no. 2, pp. 309325 (2002)

[4] J. MacCormick and A. Blake: A probabilisticexclusion principle for tracking multiple objects.International Journal of Computer Vision, vol.39, no. 1, pp. 5771 (2000)

[5] N. Gordon: A hybrid bootstrap filter for targettracking in clutter. IEEE Transactions onAerospace and Electronic Systems, vol. 33, no. 1,

pp. 353358 (1997)

#Frame:149




10/11

32

[6] T. Zhao and R. Nevatia: Tracking multiplehumans in crowded environment. In Proceedingsof the IEEE Computer Society Conference onComputer Vision and Pattern Recognition (CVPR04), vol. 2, pp. 406413, Washington, DC, USA,June-July (2004)

[7] T. Zhao and R. Nevatia: Tracking multiplehumans in complex situations. IEEE Transactionson Pattern Analysis and Machine Intelligence,vol. 26, no. 9, pp. 12081221 (2004)

[8] H. Tao, H. Sawhney, and R. Kumar: A samplingalgorithm for detection and tracking multipleobjects. In Proceedings of IEEE InternationalConference on Computer Vision (ICCV99)Workshop on Vision Algorithm, Corfu, Greece,September (1999)

[9] Z. Khan, T. Balch, and F. Dellaert: An MCMC-based particle filter for tracking multipleinteracting targets, In Proceedings of 8thEuropean Conference on Computer Vision

(ECCV 04),vol. 4, pp. 279290, Prague, CzechRepublic, May (2004)

[10] K. Smith, D. Gatica-Perez, and J.-M. Odobez:Using particles to track varying numbers ofinteracting people. In Proceedings of the IEEEComputer Society Conference on ComputerVision and Pattern Recognition (CVPR 05), vol.1, pp. 962969, San Diego, Calif, USA, June(2005)

[11] S. J. McKenna, S. Jabri, Z. Duric, A. Rosenfeld,and H. Wechsler: Tracking groups of people.Computer Vision and Image Understanding, vol.80, no. 1, pp. 4256 (2000)

[12] L. Lee, R. Romano and G. Stein.: Monitoringactivities from multiple video streams:Establishing a common coordinate frame. IEEETransactions on Pattern Analysis and MachineIntelligence, Special Issue on Video Surveillanceand Monitoring: 758767 (2000)

[13] Ting-hsun Chang , Shaogang Gong. TrackingMultiple People with a Multi-Camera System.IEEE Workshop on Multi-Object Tracking(2001)

[14] T. Yu and Y. Wu: Collaborative tracking ofmultiple targets. In Proceedings of the IEEEComputer Society Conference on Computer

Vision and Pattern Recognition (CVPR 04), vol.1, pp. 834841, Washington, DC, USA, June-July(2004)

[15] Y. Wu, G. Hua, and T. Yu. Tracking articulatedbody by dynamic Markov network. InProceedings of 9th IEEE InternationalConference on Computer Vision (ICCV 03), vol.2, pp.10941101, Nice, France, October (2003)

[16] W. Qu, D. Schonfeld, and M. Mohamed: Real-time interactively distributed multi-object

tracking using a magnetic-inertia potentialmodel. In Proceedings of 10th IEEEInternational Conference on Computer Vision(ICCV 05),vol. 1, pp. 535540, Beijing, China,October (2005)

[17] Q. Cai and J. K. Aggarwal.: Tracking humanmotion in structured environments using adistributed-camera system. IEEE Transactions onPattern Analysis and Machine Intelligence, vol.21, no. 11, pp. 12411247 (1999)

[18] P. H. Kelly, A. Katkere, D. Y. Kuramura, S.Moezzi, S. Chatterjee, and R. Jain: Anarchitecture for multiple perspective interactivevideo. In Proceedings of 3rd ACM InternationalConference on Multimedia (ACM Multimedia95), pp. 201212,San Francisco, Calif, USA,

November (1995)[19] J. Black and T. Ellis.: Multiple camera image

tracking. In Proceedings of 2nd IEEEInternational Workshop on Performance

Evaluation of Tracking and Surveillance (PETS01), Kauai,Hawaii, USA, December (2001)

[20] L. Lee, R. Romano, and G. Stein: Monitoringactivities from multiple video streams:establishing a common coordinate frame. IEEETransactions on Pattern Analysis and MachineIntelligence, vol. 22, no. 8, pp. 758767 (2000)

[21] Carin Hue, Jean-Pierre Le Cadre, and PatrickPerez.: Sequential monte carlo methods formultiple target tracking and data fusion. IEEETransactions on Signal Processing, 50(2):309325, February (2002)

[22] P. P erez, C. Hue, J. Vermaak, and M. Gangnet.:

Color-based probabilistic tracking. In EuropeanConference on Computer Vision, number 2350 inLecture Notes in Computer Science, pages 661675 (2002)

[23] Gale, D. & Shapley, L. S.: College admissionsand the stability of marriage. AmericanMathematical Monthly, 69, 914 (1962)

[24] http://en.wikipedia.org/wiki/Stable_marriage_problem

[25] Guraya, F. F. E., Bayle, P.-Y., & Cheikh, F. A.:People tracking via a modified CAMSHIFTalgorithm. (2009)

[26] Maskell, S. & Gordon, N.: A tutorial on particle

Filters for on-line nonlinear/non-GaussianBayesian tracking. IEEE Transactions on SignalProcessing 2001, 50, 174188 (2001)

[27] Kitagawa, G. Monte Carlo: Filter and smootherfor non-Gaussian nonlinear state space models.Journal of Computational and GraphicalStatistics, 5, No. 1, 125, 1996

[28] Tarng Chen, Yu-Ching Lin, Wen-Hsien Fang: AVIDEO-BASED HUMAN FALL DETECTIONSYSTEM FOR SMART HOMES. YieJournal of




11/11

33

the Chinese Institute of Engineers, Vol. 33, No. 5,pp. 681-690 (2010)

[29] Nummiaro, K., Koller-Meier, E., Gool, L. V., &Gaal, L. V.: Object tracking with an adaptivecolor-based particle Filter.http://www.koller-meier.ch/esther/dagm2002.pdf.(2002)

[30] Bouguet, J. Y.: Pyramidal implementation of theLucas Kanade feature tracker: Description ofthe algorithm. Intel Corporation MicroprocessorResearch Labs,http://robots.stanford.edu/cs223b04/algo_tracking.pdf. (2002)

[31] Blake, A. & Isard, M. : The Condensationalgorithm - conditional density propagation andapplications to visual tracking. In Advances in

Neural Information Processing Systems(NIPS)1996, 361, Denver, CO, USA. The MITPress. December 2-5 (1996)

[32] Zhao, S., Zhao, J., Wang, Y., & Fu, X.: Moving

object detecting using gradient information,three-frame-differencing and connectivity testing.AI 2006: Advances in Artificial Intelligence,4304, 510518 (2006)

[33] Victoria Rudakova, Sajib Kumar Saha, FaouziAlaya Cheikh,: Multiple collaborative camerasfor multi-people tracking using color-based

particle filter and border information of theobjects. DICTAP 2011: International Conferenceon Digital Information and CommunicationTechnology and its Application, 315-326(2011)



multi-people tracking across multiple cameras

Documents