资源描述:
《Multilayer Adaptive Linear Predictors for Real-Time Tracking》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、IEEETRANSACTIONSONPATTERNANALYSISANDMACHINEINTELLIGENCE,VOL.35,NO.1,JANUARY2013105MultilayerAdaptiveLinearPredictorsforReal-TimeTrackingStefanHolzer,StudentMember,IEEE,SlobodanIlic,Member,IEEE,andNassirNavab,Member,IEEEAbstract—Enlargingorreducingthetemp
2、latesizebyaddingnewpartsorremovingpartsofthetemplateaccordingtotheirsuitabilityfortrackingrequirestheabilitytodealwiththevariationofthetemplatesize.Forinstance,real-timetemplatetrackingusinglinearpredictors,althoughfastandreliable,requiresusingtemplateso
3、fafixedsizeanddoesnotallowonlinemodificationofthepredictor.Tosolvethisproblem,weproposetheAdaptiveLinearPredictors(ALPs),whichenablefastonlinemodificationsofprelearnedlinearpredictors.Insteadofapplyingafullmatrixinversionforeverymodificationofthetemplate
4、shape,asstandardapproachestolearninglinearpredictorsdo,wejustperformafastupdateofthisinverse.ThisallowsustolearntheALPsinamuchshortertimethanstandardlearningapproacheswhileperformingequallywell.Additionally,weproposeamultilayerapproachtodetectocclusionsa
5、nduseALPstoeffectivelyhandlethem.Thisallowsustotracklargetemplatesandmodifythemaccordingtothepresentocclusions.Weperformedanexhaustiveevaluationofourapproachandcomparedittostandardlinearpredictorsandotherstate-of-the-artapproaches.IndexTerms—Templatetrac
6、king,linearpredictorsÇ1INTRODUCTIONTEMPLATEtrackinghasbeenstudiedextensivelyandusedAsaresult,thesetofinitiallytrackedtemplatesevolvesinmanycomputervisionapplicationssuchasvision-towardarelativelysmallnumberofcomparablylarge,basedcontrol,human-computerint
7、erfaces,surveillance,optimallyshapedtemplates,yieldingincreasedrobustness.medicalimaging,andreconstruction.Currentlearning-basedtrackingapproaches,like[11],Whiletherearemanytemplatetrackingapproachesbasedusetemplatesofafixedsizebecausethecomputationofthe
8、ontheanalyticalderivationoftheJacobian[1],[2],[3],[4],[5],linearpredictorsrequiresthecostlyinversionofalarge,[6],[7],[8],[9],[10],learning-basedmethods[11],[12],[13],template-specificmatrix.Sincethisisthecomputationally[14