automatttrajecttt ttttadatafromsky.com/wp-content/uploads/2015/05/64_poster.pdft t adat t t targt t...

Automatic vehicle trajectory extraction from aerial video data Adam Babinec www6datafromsky6com infoWdatafromsky6com Partners5 Faculty of Information TechnologyS Brno University of TechnologyS Czech Republic 1 xbabinD4Wstud6fit6vutbr6cz RCE systems s6r6o6S Svatopluka Čecha qdS Gq4 DD BrnoS Czech Republic 1 adam6babinecWrcesystems6cz Traffic Monitoring Racing Telemetry Car: 68 AF Priaulx Speed: BBfFU kmRh TanF AccF: V UFBU G LatF AccF: fFfW G Split: g UF6UB sec Car: 6UU YF Muller Speed: U8BFU kmRh TanF AccF: V UFU8 G LatF AccF: fFfU G Split: g fF994 sec Improve Intersection Design 64 This poster presents a system for automatic vehicle trajectory extraction from aerial video data6 The input video sequence can be captured by regular action camera mounted on a multicopter drone or balloon flying in altitudes from ND to 4DD m6 The geo1registration of video sequence is based upon transformation estimation of two ORB feature sets extracted from a reference image and video sequence frame6 To provide the robustness of the estimationS RANSAC procedure is employed to guide the algorithm6 Vehicle detection candidates are produced by video sequence analysis in temporal dimension to detect moving objects6 This is carried out by background subtraction algorithm using Gaussian Mixture Models6 Its output is fused together with a road surface mask and currently tracked objects database to produce set of detection candidates6 Geo1registered Image Detection Candidates Detections Multi1Object Tracking Input Preprocessing Detection Tracking GIS Input Video Frame GeoV Registered Reference Image RoadV Reference Mask GeoVRegistered Video Image Image Undistortion GeoV Registration Background Model Classifiers Candidate Generation Strong Vehicle Classifier Weak Vehicle Classifier DetectionVTrack Association Tracks Update Tracks Management The detection candidates are analyzed by two vehicle classifiers 1 strong vehicle classifier which produces confident detection cuesS and weak vehicle classifier which provides tracking attractors to aid multi1target tracking6 The classifiers use Multi1Block Local Binary Pattern features and are constructed by AdaBoost algorithm on dataset of 4DSDDD hand annotated vehicles and road structures6 The tracking of vehicles is carried out in RGBEEdge image space using a set of independent Bayesian bootstrap particle filtersS one for each target6 The transition model of the particle filter is simple velocity model considering the vehicle position as integration of its velocity6 Target model of the vehicle is represented by O4xO4 RGB EEdge template with circular maskS which is extracted from the area initial vehicle detection6 The plasticity of the model is achieved by Vz template update in case of overlapping strong detection andMor when the cues from the weak classifier are strong enough in the area of tracked vehicle6 To prevent target swapsS the update is disabled when multiple targets overlap6 The evaluation of particle is based on its template distance to the target model and the attractor cues produced by the weak classifier6 The initialisation of particle filter for moving object is guided by ,falloff, algorithmS which at the beginning of the target tracking causes the position of the particles of given target to be more affected by random noise than their velocityS so the particles can slowly adapt to the target motion while elevating particlesp velocity effect on their behaviour6 q | Capture Aerial Video 4 | Provide GIS Annotation O | Extract Trajectories N | Improve the World

Upload: others

Post on 03-Jan-2020

0 views

Category:

Documents

0 download

Report

Download

Embed Size (px):

TRANSCRIPT

Page 1: AutomatTTrajectTT TTTtadatafromsky.com/wp-content/uploads/2015/05/64_poster.pdfT T adaT T T targT T T elevatinT particpT TTTT6 qTTTT 4TTTTion OTTTories NTTTT

AutomaticTvehicleTtrajectoryTextractionTfromTaerialTvideoTdataAdamTBabinec

www6datafromsky6cominfoWdatafromsky6com

Partners5

FacultyTofTInformationTTechnologySTBrnoTUniversityTofTTechnologySTCzechTRepublicTT1TxbabinD4Wstud6fit6vutbr6czTRCETsystemsTs6r6o6STSvatoplukaTČechaTqdSTGq4TDDTBrnoSTCzechTRepublicT1Tadam6babinecWrcesystems6cz

TrafficTMonitoring

RacingTTelemetry

Car:t68tAFtPriaulxSpeed:tBBfFUtkmRhTanFtAccF:tVtUFBUtGLatFtAccF:tfFfWtGSplit:tgtUF6UBtsec

Car:t6UUtYFtMullerSpeed:tU8BFUtkmRhTanFtAccF:tVtUFU8tGLatFtAccF:tfFfUtGSplit:tgtfF994tsec

Improve

IntersectionTDesign

ThisT posterT presentsT aT systemT forT automaticT vehicleT trajectoryTextractionTfromTaerialTvideoTdata6TTheT inputTvideoTsequenceTcanTbeTcapturedTbyTregularTactionTcameraTmountedTonTaTmulticopterTdroneTorTballoonTflyingTinTaltitudesTfromTNDTtoT4DDTm6

TheT geo1registrationT ofT videoT sequenceT isT basedT uponTtransformationT estimationT ofT twoT ORBT featureT setsT extractedT fromTaT referenceT imageT andT videoT sequenceT frame6T ToT provideT theTrobustnessT ofT theT estimationST RANSACT procedureT isT employedT toTguideTtheTalgorithm6

VehicleT detectionT candidatesT areT producedT byT videoT sequenceTanalysisT inT temporalT dimensionT toT detectT movingT objects6T ThisT isTcarriedT outT byT backgroundT subtractionT algorithmT usingT GaussianTMixtureT Models6T ItsT outputT isT fusedT togetherT withT aT roadT surfaceTmaskT andT currentlyT trackedT objectsT databaseT toT produceT setT ofTdetectionTcandidates6

Geo1registeredTImage DetectionTCandidates

Detections Multi1ObjectTTracking

InputtPreprocessing Detection Tracking

GIS

InputtVideotFrame

GeoVRegisteredtReferencet

Image

RoadVtReferencet

Mask

GeoVRegisteredtVideotImage

ImagetUndistortion

GeoVRegistration

BackgroundtModel

Classifiers

CandidatetGeneration

StrongtVehicletClassifier

WeaktVehicletClassifier

DetectionVTracktAssociation

TrackstUpdate

TrackstManagement

TheT detectionT candidatesT areT analyzedT byT twoT vehicleT classifiersT 1TstrongT vehicleT classifierT whichT producesT confidentT detectionT cuesSTandTweakTvehicleTclassifierTwhichTprovidesTtrackingTattractorsTtoTaidTmulti1targetT tracking6T TheT classifiersT useT Multi1BlockT LocalT BinaryTPatternT featuresT andT areT constructedT byT AdaBoostT algorithmT onTdatasetTofT4DSDDDThandTannotatedTvehiclesTandTroadTstructures6

TheT trackingT ofT vehiclesT isT carriedT outT inT RGBEEdgeT imageT spaceTusingTaTsetTofT independentTBayesianTbootstrapTparticleT filtersSToneTforTeachTtarget6TTheTtransitionTmodelTofTtheTparticleTfilterT isTsimpleTvelocityTmodelTconsideringTtheTvehicleTpositionTasTintegrationTofTitsTvelocity6TTargetTmodelTofT theTvehicleT isT representedTbyTO4xO4TRGBEEdgeTtemplateTwithTcircularTmaskSTwhichTisTextractedTfromTtheTareaTinitialTvehicleTdetection6TTheTplasticityTofT theTmodelT isTachievedTbyTVzTtemplateTupdateTinTcaseTofToverlappingTstrongTdetectionTandMorTwhenT theT cuesT fromT theT weakT classifierT areT strongT enoughT inT theTareaT ofT trackedT vehicle6T ToT preventT targetT swapsST theT updateT isTdisabledTwhenTmultipleTtargetsToverlap6

TheTevaluationTofTparticleT isTbasedTonT itsT templateTdistanceTtoTtheTtargetT modelT andT theT attractorT cuesT producedT byT theT weakTclassifier6T TheT initialisationT ofT particleT filterT forT movingT objectT isTguidedTbyT,falloff,TalgorithmSTwhichTatTtheTbeginningTofTtheTtargetTtrackingT causesT theTpositionTofT theTparticlesTofTgivenT targetT toTbeTmoreTaffectedTbyTrandomTnoiseTthanTtheirTvelocitySTsoTtheTparticlesTcanT slowlyT adaptT toT theT targetT motionT whileT elevatingT particlespTvelocityTeffectTonTtheirTbehaviour6

qT|TCaptureTAerialTVideo

4T|TProvideTGISTAnnotation

OT|TExtractTTrajectories

NT|TImproveTtheTWorld

NATU RLD IGEB S - UConn CT ECOcteco.uconn.edu/maps/town/basin/basin_Eastford.pdf · t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t

ice u Skut Zm na . 1 ÚP RANÁ - Hlinskot t t t t t t t t t t t t t t t t t tt t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t

10 · h e r c u l e s - s t r. t t t t l t t t t s t l t l t t l t t t t t t t l l l t t t t k l l t t t t t t t t t t t t t t t t t t l t t t t m t t t e s s e e e e t e e

T t - T t - T t - T t - T t - T t...Topolino no estés triste Topo bueno toma el tren, Toma el carro, ten la moto Y este cohete también. T t - T t - T t - T t - T t - T t Ta - Te

F T E P FF or TF H - xfer.services.ncdot.gov · t t t t t t t t t t t t t t t t t t t t t t t t t t t t tt t t t t t t t t ttttttttttttttttttttttttttttttttttt t ttttttt t t ttt t

12/5vp/1os - Oikos Engineering...T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T

dtt - samfunnsokonomene.no · hjbn. dtt nr r nd. jr. nt hr n vrt vr dtt nt. Frøvr vl n dtt nr fnn n rntrn løn fnn Høtnfrn r v tbln .2. nvbr, n tl v pblrt rnd fr løn ntttt, ttntt,

T T T T T T T T T T T T T T T T T T