搜索
©1994-2010ChinaAcademicJournalElectronicPublishingHouse.Allrightsreserved.:2009201216.:863(No.2005AA778032)10042924X(2010)0120234206Mean2shift1,2,1,1,2(1.,130033;2.,100039):Mean2shift,,,(log2likelihoodimage),,,,,Mean2shift,,,,:,:;Mean2shift;;:TP301.6;TP391:ARobustobjecttrackingbasedonimprovedMean2shiftalgorithmXUEChen1,2,ZHUMing1,CHENAi2hua1,2(1.ChangchunInstituteofOptics,FineMechanicsandPhysics,ChineseAcademyofSciences,Changchun130033,China;2.GraduateUniversityofChineseAcademyofSciences,Beijing100039,China)Abstract:ToovercometheshortcomingsofthetraditionalMean2shiftalgorithmforobjecttrackingsuchasthelocalizationerrorcausedbybackgroundpixelsandthetrackingfailurefromtheobjectocclusion,anim2provedMean2shiftalgorithmisproposed.Firstly,accordingtothedifferenceofcolordistributionbetweentheobjectandthebackgroundintheinitialframe,alog2likelihoodimageissetuptoselectthediscriminativecolorfeaturesforobjectmodeling,andthenthecandidatemodelingisestablishedbythesameway.Byaboveoper2ation,theeffectofbackgroundpixelsontheimagehasreducedgreatly.Secondly,thewholecandidateregionisseparatedintoseveraloverlappedfragments,andeveryfragmentisiteratedbytheMean2shift.Then,theob2jectlocalizationisresetbythelocationoffragmentinthecandidateregion,whichmatchesmostlytothecorre2spondingfragmentintheobjectregion.ExperimentalresultsshowthatthefragmentbasedontheMean2shiftisveryrobusttopartialocclusion.Furthermore,whenobjectisseverelyoccluded,thelinearpredictioncanbeusedtoestimatetheprobablelocationoftheobjectinthenextframe.Theseresultsprovethatthetrackingusing©1994-2010ChinaAcademicJournalElectronicPublishingHouse.Allrightsreserved.[122]Mean2shift,,,,Mean2shift,,,,Mean2shift,,[324](Spatiogram)[526]Mean2shift[728]Mean2shiftKalman[9210],:(1),(log2likelihoodimage),,,,;(2),,Mean2shift,,,,2Mean2shiftMean2shiftnXS,x(xX):m(x)=sSK(s-x)(s)ssSK(s-x)(s),sS,(1):K;,m(x)-xMean2shiftMean2shiftMean2shiftMean2shift,:qu=Cni=1k(xi-x0h2)[b(xi)-u],(2):uu,{xi}i=1,,n,kK,Kronecker,b:R2{1,,m},Cqu=1.,:pu(y)=Chni=1k(y-xih2)[b(xi)-u],(3),Chpu=1qupu(y)Bhattacharyya:[p(y),q]=Mu=1pu(y)qu.(4)BhattacharyyaTaylorMean2shift:y1=ni=1xiwig(y-xih2)ni=1wig(y-xih2),(5):g(x)=-k(x),(6)wi=mu=1qupu(y)[b(xi)-u].(7)3Mean2shiftMean2shift,(),Mean2shift5321,:Mean2shift©1994-2010ChinaAcademicJournalElectronicPublishingHouse.Allrightsreserved.[11],f,HobjHbg(i)i,p(i)q(i):p(i)=Hobj(i)/nobj,(8)q(i)=Hbg(i)/nbg,(9)nobjnbgi:L(i)=logmax(p(i),)max(q(i),),(10),(0.0001),(a)&(b)(a)Object&background(b)Log2likelihoodimage1Fig.1Log2likelihoodimage,(log2likelihoodimage),,(11),tho,,,T(xu)=1,L(i)tho0,otherwise.(11)3.2Mean2shift(8)L(i),sigmoidL(i)(0,1)u,:u=max1-11+exp(Lu),0.1.(12)(2)(3),:qu=1Cni=1k(xi2)u[b(xi)-u],(13)pu(y)=1Chni=1k(y-xih2)uh[b(xi)-u],(14)CChqu=1pu=1Mean2shift4Mean2shift,,,,Mean2shift,,Mean2shift,(Bhattacharyya),,,;,,61,;2,;36,,2,6Bhat2tacharyya,,Bhattacharyya63218©1994-2010ChinaAcademicJournalElectronicPublishingHouse.Allrightsreserved.(2),,Mean2shift,,(a)(b)18(a)Frame1(b)Frame18(c)44(d)70(c)Frame44(d)Frame70(e)98(f)106(e)Frame98(f)Frame1063Fig.3Examplesofcartracking4errori:errori=(xi-xc)2+(yi-yc)2,(15),(xi,yi),(xc,yc),,4Mean2shift4Fig.4Localizationerrors5,35,Bhattacharyya,;,4351,BhattacharyyaTocc,,;,55,Bhattacharyya,Tapp,6Bhattacharyya7321,:Mean2shift©1994-2010ChinaAcademicJournalElectronicPublishingHouse.Allrightsreserved.(a)(b)35(a)Frame1(b)Frame35(c)43(d)51(c)Frame43(d)Frame51(e)55(f)65(e)Frame55(f)Frame655Fig.5Cartrackingthroughocclusion65BhattacharyyaFig.6BhattacharyyacoefficientforFig.56Mean2shift,,,,Mean2shift,:[1]CHENGY.Meanshift,modeseeking,andcluste2ring[J].IEEETrans.onPatternAnalysisandMachineIntelligence,1995,17(8):7902799.[2]COMANICIUD,RAMESHV,MEERP.Kernel2basedobjecttracking[J].IEEETrans.onPatternAnalysisandMachineIntelligence,2003,25(5):5642577.[3]YIMAZA.Objecttrackingbyasymmetrickernelmean2shiftwithautomaticscaleandorientationse2lection[C].IEEEConferenceonComputerVisionandPatternRecognition,IEEE,2007:126.[4]CHENXP,YUSHSH,MAZHL.Animprovedmean2shiftalgorithmforobjecttracking[C].Pro2ceedingsofthe7thWorldCongressonIntelligentControlandAutomation,IEEE,2008:511125114.[5]BIRCHFIELDST,RANGARAJANS.Spatio2gramsversushistogramsforregionbasedtracking[C].IEEEComputerSocietyConferenceonCom2puterVisionandPatternRecognition,IEEE,2005:115821163.[6]OCONAIREC,OCONNORN,SMEATONAF.Animprovedspatiogramsimilaritymeasureforrobustobjectlocalization[C].ICASSP,IEEE,2007:106921072.[7]NUMMIAROK,KOLLER2MEIERE,VANGL.Colorfeaturesfortrackingnon2rigidobject[J].SpecialIssueonVisionSurveillance,2003,29(3):3452355.[8],.MSMC[J].,2008,16(1):1222127.MENGB,ZHUM.Applicati