Robust Tracking Using Local Sparse Appearance Model and K-Selection

RobustTrackingUsingLocalSparseAppearanceModelandK-Selection

BaiyangLiu1 JunzhouHuang11

Rutgers,TheStateUniversityofNewJersey

Piscataway,NJ,08854

baiyang,jzhuang,kulikows@cs.rutgers.edu

LinYang2CasimirKulikowsk1

2

UMDNJ-RobertWoodJohnsonMedicalSchool

Piscataway,NJ,08854

yangli@umdnj.edu

Abstract

Onlinelearnedtrackingiswidelyusedforit’sadap-tiveabilitytohandleappearancechanges.However,itin-troducespotentialdriftingproblemsduetotheaccumula-tionoferrorsduringtheself-updating,especiallyfortheoccludedscenarios.Therecentliteraturedemonstratesthatappropriatecombinationsoftrackerscanhelpbal-ancestabilityand exibilityrequirements.Wehavede-velopedarobusttrackingalgorithmusingalocalsparseappearancemodel(SPT).Astaticsparsedictionaryandadynamicallyonlineupdatedbasisdistributionmodelthetargetappearance.Anovelsparserepresentation-basedvotingmapandsparseconstraintregularizedmean-shiftsupporttherobustobjecttracking.Besidesthesecontri-butions,wealsointroduceanewdictionarylearningal-gorithmwithalocallyconstrainedsparserepresentation,calledK-Selection.Basedonasetofcomprehensiveexper-iments,ouralgorithmhasdemonstratedbetterperformancethanalternativesreportedintherecentliterature.

(a)(b)(c)

(d)(e)(f)

Figure1.Thetargetappearance(a)ismodeledwithadictionary(b)andasparsecodinghistogram(c).Thecon dencemap(e)oftheimage(d)istheinverseofthereconstructionerrorfromthelearnedtargetdictionary.Thetargetcenterisfoundbyvotingandsparseconstraintwithregularizedmean-shiftontheprobabilitymap(f).

1.Introduction

Generativeanddiscriminativemethodsaretwomajorcategoriesusedincurrenttrackingtechniques.Thegen-erativemodelsformulatethetrackingproblemassearch-ingfortheregionswiththehighestlikelihood[5,7,16,18,19,25,28,15],eitherusingasingleappearancemodelormultipleappearancemodels[27].Discriminativemethodsformulatetrackingasaclassi cationproblem[3,4,6,11].Thetrainedclassi erisusedtodiscriminatethetargetfromthebackgroundandcanbeupdatedonline.Grabneret.al.[9]proposedtoupdatetheselectedfeaturesincrementallyusingthecurrenttrackingresult,whichmayleadtopoten-tialdriftingbecauseofaccumulatederrors,aggravatedbyocclusions.

Inordertohandlethedriftingproblem,semi-onlineboosting[10]waslaterproposedtoincrementallyupdate

researchiscompletedwhentheauthorisaresearchassistantin

theUMDNJ-RobertWoodJohnsonMedicialSchool

This

theclassi erusingbothunlabeledandlabeleddata.TheMultipleInstanceLearningboostingmethod(MIL)[4]putsallsamplesintobagsandlabelsthem.Thedriftingprob-lemishandledinthismethodsincethetruetargetincludedinthepositivebagislearnedimplicitly.Recently,ithasbeenshownthatanappropriatecombinationofcomplemen-tarytrackingalgorithmscanhelpalleviatedriftingproblems[12,26,21,20].In[17,13],asparserepresentationmod-elsthedynamicappearanceofthetargetandhandlestheocclusionasasparsenoisecomponent.Allthesemethodsmodelthetargetasasingleentity,andthereforecannothan-dlepartialocclusionverywell.Fragment-basedtrackingin[1]coupledwithavotingmapcanaccuratelytrackthepar-tiallyoccludedtarget.However,thismethodtrackseachtargetpatchwithastatictemplate,whichlimitsitsexpres-sivepower.Itmayfailinadynamicenvironmentwhichexhibitsappearancechangesorposevariations.

Inthepresentpaper,weproposeandtestarobusttrack-ingalgorithmwithalocalsparseappearancemodel(SPT).Thealgorithm’skeycomponentsareastaticsparsedictio-narywhichisusedtolimitdriftingandkeepthe exibil-ityinitslinearlyspannedsubspace;adynamicbasisdistri-