資源描述:
《Robust Deformable and Occluded Object》由會(huì)員上傳分享,免費(fèi)在線閱讀,更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。
1、IEEETRANSACTIONSONIMAGEPROCESSING,VOL.23,NO.12,DECEMBER20145497RobustDeformableandOccludedObjectTrackingWithDynamicGraphZhaoweiCai,LongyinWen,ZhenLei,Member,IEEE,NunoVasconcelos,SeniorMember,IEEE,andStanZ.Li,Fellow,IEEEAbstractWhilesomeeffortshavebeenpaidtohandl
2、eocclusionareubiquitousintrackingproblems.Whilemanydeformationandocclusioninvisualtracking,theyarestillgreattrackers[2],[15],[29]haveconsideredocclusion,onlyachallenges.Inthispaper,adynamicgraph-basedtracker(DGT)fewworks[20],[40]haveaddressedtheproblemofshapeisp
3、roposedtoaddressthesetwochallengesinauni?edframe-deformation.Inthispaper,weapproachthetwochallengesinwork.Inthedynamictargetgraph,nodesarethetargetlocalpartsencodingappearanceinformation,andedgesaretheauni?edframework,dynamicgraphbasedtracking,becauseinteraction
4、sbetweennodesencodinginnergeometricstructuregraphrepresentationintuitivelyownsthepowertorecognizeinformation.Thisgraphrepresentationprovidesmuchmorethegeometricdeformabletarget,andtoaccuratelylocalizetheinformationfortrackinginthepresenceofdeformationandoccluded
5、targetwiththeotherunoccludedpartsofthetarget.occlusion.ThetargettrackingisthenformulatedastrackingRecenttrackersachievehightrackingaccuracyandrobust-thisdynamicundirectedgraph,whichisalsoamatchingprob-lembetweenthetargetgraphandthecandidategraph.Thenessmainlythr
6、oughthreeaspects:feature,appearancemodellocalpartswithinthecandidategraphareseparatedfromtheandstructureinformation.Featurescommonlyusedwithdif-backgroundwithMarkovrandom?eld,andspectralclusteringferentpropertiesincludepixelvalues[23],color[2],[8],isusedtosolvet
7、hegraphmatching.The?naltargetstate[30],[40],andtexturedescriptors[3],[14].Theappear-isdeterminedthroughaweightedvotingprocedureaccordingancemodelisusedtocharacterizethetarget,suchascolortothereliabilityofpartcorrespondence,andre?nedwithrecoursetoaforeground/back
8、groundsegmentation.Aneffectivedistribution[8],[30],subspaces[23],[42],SupportVectoronlineupdatingmechanismisproposedtoupdatethemodel,Machine[37],Boosting[3],[14],[19]