资源描述:
《apg_L1_final real time robust L1 tracker using acclerated proximal gradient approach .pdf》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、RealTimeRobustL1TrackerUsingAcceleratedProximalGradientApproachChenglongBao1,YiWu2,HaibinLing2,andHuiJi11DepartmentofMathematics,NationalUniversityofSingapore,Singapore,1190762DepartmentofComputerandInformationSciences,TempleUniversity,Philadelphia,PA,USA,19122{baochenglong,ma
2、tjh}@nus.edu.sg,{wuyi,hbling}@temple.eduAbstractcludingvisualtracking[15,16,11,14,24].Similartosparsity-basedapproachforfacerecognitiondevelopedinRecentlysparserepresentationhasbeenappliedtovi-[22],thesetrackingmethodsexpressatargetbyasparsesualtrackerbymodelingthetargetappear
3、anceusingalinearcombinationofthetemplatesinthetemplatespace,sparseapproximationoveratemplateset,whichleadstoi.e.,thetargetiswellapproximatedbythelinearcombina-theso-calledL1trackersasitneedstosolvean1normtionofonlyafewtemplates.Benefittingfromthestablerelatedminimizationproble
4、mformanytimes.Whiletheserecoverycapabilityofsparsesignalusingthe1normmin-L1trackersshowedimpressivetrackingaccuracies,theyareimization(e.g.[5]),thesetrackershavedemonstratedgoodverycomputationallydemandingandthespeedbottleneckrobustnessinvarioustrackingenvironments.isthesolve
5、rto1normminimizations.ThispaperaimsIntheL1trackerfirstproposedby[15],hundredsof1atdevelopinganL1trackerthatnotonlyrunsinrealtimenormrelatedminimizationproblemsneedtobesolvedforbutalsoenjoysbetterrobustnessthanotherL1trackers.Ineachframeduringthetrackingprocess.Thesolverforthe
6、ourproposedL1tracker,anew1normrelatedminimiza-1normminimizationsusedin[15]isbasedontheinteriortionmodelisproposedtoimprovethetrackingaccuracybypointmethodwhichturnsouttobetooslowfortracking.addingan2normregularizationonthecoefficientsassoci-Aminimalerrorboundingstrategyisint
7、roduced[16]tore-atedwiththetrivialtemplates.Moreover,basedontheac-ducethenumberofparticles,equaltothenumberoftheceleratedproximalgradientapproach,averyfastnumeri-1normminimizationsforsolving.Aspeedupbyfourtocalsolverisdevelopedtosolvetheresulting1normrelatedfivetimesisreporte
8、din[16],butitisstillfarawayfromminimizationproblemwithguarant