资源描述：
《2011-Robust_Tracking_Using_Local_Sparse_Appearance_Model_and_K-Selection.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、RobustTrackingUsingLocalSparseAppearanceModelandK-SelectionBaiyangLiu1JunzhouHuang1LinYang2CasimirKulikowsk11Rutgers,TheStateUniversityofNewJersey2UMDNJ-RobertWoodJohnsonMedicalSchoolPiscataway,NJ,08854Piscataway,NJ,08854baiyang,jzhuang,kulikows@cs.rut
2、gers.eduyangli@umdnj.eduAbstractOnlinelearnedtrackingiswidelyusedforit’sadap-tiveabilitytohandleappearancechanges.However,itin-troducespotentialdriftingproblemsduetotheaccumula-(a)(b)(c)tionoferrorsduringtheself-updating,especiallyfortheoccludedscenario
3、s.Therecentliteraturedemonstratesthatappropriatecombinationsoftrackerscanhelpbal-ancestabilityandflexibilityrequirements.Wehavede-velopedarobusttrackingalgorithmusingalocalsparse(d)(e)(f)appearancemodel(SPT).Astaticsparsedictionaryandadynamicallyonlineup
4、datedbasisdistributionmodeltheFigure1.Thetargetappearance(a)ismodeledwithadictionary(b)andasparsecodinghistogram(c).Theconfidencemap(e)oftargetappearance.Anovelsparserepresentation-basedtheimage(d)istheinverseofthereconstructionerrorfromthevotingmapandsp
5、arseconstraintregularizedmean-shiftlearnedtargetdictionary.Thetargetcenterisfoundbyvotingandsupporttherobustobjecttracking.Besidesthesecontri-sparseconstraintwithregularizedmean-shiftontheprobabilitybutions,wealsointroduceanewdictionarylearningal-map(f)
6、.gorithmwithalocallyconstrainedsparserepresentation,calledK-Selection.Basedonasetofcomprehensiveexper-iments,ouralgorithmhasdemonstratedbetterperformancetheclassifierusingbothunlabeledandlabeleddata.Thethanalternativesreportedintherecentliterature.Multip
7、leInstanceLearningboostingmethod(MIL)[4]putsallsamplesintobagsandlabelsthem.Thedriftingprob-lemishandledinthismethodsincethetruetargetincluded1.Introductioninthepositivebagislearnedimplicitly.Recently,ithasGenerativeanddiscriminativemethodsaretwomajorbe
8、enshownthatanappropriatecombinationofcomplemen-categoriesusedincurrenttrackingtechniques.Thegen-tarytrackingalgorithmscanhelpalleviatedriftingproblemserativemodelsformulatethetrackingproblemassearch-[12,26,21,20].In[17,13],aspars
