Abstract. Visual tracking, in essence, deals with non-stationary data streams that change over time. While most existing algorithms are able to track objects well in controlled environments, they usually fail if there is a significant change in object appe

AdaptiveProbabilisticVisualTrackingwithIncrementalSubspaceUpdate

DavidRoss1,JongwooLim2,andMing-HsuanYang3

2UniversityofToronto,Toronto,ONM5S3G4,CanadaUniversityofIllinoisatUrbana-Champaign,Urbana,IL61801,USA3HondaResearchInstitute,MountainView,CA94041,USA

dross@cs.toronto.edujlim1@uiuc.edumyang@honda-ri.com1

Abstract.Visualtracking,inessence,dealswithnon-stationarydatastreamsthatchangeovertime.Whilemostexistingalgorithmsareabletotrackobjectswellincontrolledenvironments,theyusuallyfailifthereisasigni cantchangeinobjectappearanceorsurroundingillumina-tion.Thereasonbeingthatthesevisualtrackingalgorithmsoperateonthepremisethatthemodelsoftheobjectsbeingtrackedareinvarianttointernalappearancechangeorexternalvariationsuchaslightingorviewpoint.Consequentlymosttrackingalgorithmsdonotupdatethemodelsoncetheyarebuiltorlearnedattheoutset.Inthispaper,wepresentanadaptiveprobabilistictrackingalgorithmthatupdatesthemodelsusinganincrementalupdateofeigenbasis.Totrackobjectsintwoviews,weuseane ectiveprobabilisticmethodforsamplinga nemotionparameterswithpriorsandpredictingitslocationwithamaxi-mumaposterioriestimate.Borneoutbyexperiments,wedemonstratetheproposedmethodisabletotrackobjectswellunderlargelighting,poseandscalevariationwithclosetoreal-timeperformance.

1Introduction

Visualtrackingessentiallydealswithnon-stationarydata,boththeobjectandthebackground,thatchangeovertime.Mostexistingalgorithmsareabletotrackobjects,eitherpreviouslyviewedornot,inashortspanoftimeandinawellcontrolledenvironment.Howeverthesealgorithmsusuallyfailtoobservetheobjectmotionorhavesigni cantdriftsaftersomeperiodoftime,eitherduetothedrasticchangeoftheobjectappearanceorlargelightingvariationinthesurroundings.Althoughsuchsituationscanbeamelioratedwithrecoursetoview-basedappearancemodels[1][2],adaptivecolor-basedtrackers[3][4],contour-basedtrackers[5][4],particle lters[5],3Dmodelbasedmethods[6],optimizationmethods[1][7],andbackgroundmodeling[8],mostalgorithmstypicallyoperateonthepremisethatthetargetobjectmodelsdonotchangedrasticallyovertime.Consequentlythesealgorithmsbuildorlearnmodelsoftheobjects rstandthenusethemfortracking,withoutadaptingthemodelstoaccountforchangesoftheappearanceoftheobject,e.g.,largevariationofposeorfacialexpression,orthesurroundings,e.g.,lightingvariation.Suchanapproach,inourview,ispronetoperformanceinstabilityandneedstobeaddressedforbuildingarobusttracker.