您好,欢迎访问三七文档
当前位置:首页 > 商业/管理/HR > 管理学资料 > 一种分层Mean Shift目标跟踪算法
354Vol.35,No.420094ACTAAUTOMATICASINICAApril,2009MeanShift1;21121Meanshift(MS),Meanshift(Hierarchicalmeanshift,HMS).MS,,,.,,,,,,HMS.,MS,HMS,,.,Meanshift,,TP391AHierarchicalMeanShiftAlgorithmforObjectTrackingXUHai-Xia1;2WANGYao-Nan1YUANXiao-Fang1ZHOUWei2ZHUJiang1AbstractWeproposeahierarchicalmeanshift(HMS)algorithmforobjecttracking.Firstly,clustermodalpointsareobtainedbymean-shiftiterativelyprocessingallthedatapointsintheregionsothattheycanrepresentforegroundobjectinasuccinctmanner.Thetargetmodelandthetargetcandidatemodelaredescribedbytheclustermodalpoints,andmatchprocessesofclusteredblocksareperformed.Then,onthebasisofclusterblocksmatch,similaritymeasurefunctionissetuptomatchbetweentargetmodelandtargetcandidateatpixellevel.Andthepixelshiftvectoroftargetiscalculatedwiththeintroductionoftheneighborhoodconsistencyconcept.So,thecentroidoftrackingobjectisgotlayerbylayerintheconsecutiveframes,andtheHMSmatchiterationforobjecttrackingispresented.ExperimentalcomparisonswithothertwoMSalgorithmsdemonstratethevalidityandperformanceoftheproposedalgorithm.KeywordsObjecttracking,hierarchicalmeanshift(HMS),clustermodalpoint,match.KalmanParticleMeanshift(MS).MStrackComaniciu[1]2000[2],.,,BhattacharyyaKullbackLeibler,MS.MS:1)2007-11-062008-06-26ReceivedNovember6,2007;inrevisedformJune26,2008(60835004),(863)(2007AA04Z244,2008AA04Z214),(07C073)SupportedbyNationalNaturalScienceFoundationofKeyProgramofChina(60835004),NationalHighTechnologyRe-searchandDevelopmentProgramofChina(863Program)(2007AA04Z244,2008AA04Z214),andFoundationofHunanEducationalCommittee(07C073)1.4100822.4111051.SchoolofElectricalandInformationEngineering,HunanUniversity,Changsha4100822.SchoolofInformationEngi-neering,XiangtanUniversity,Xiangtan411105DOI:10.3724/SP.J.1004.2009.00401MS,.[3],,;[4],,,;[5¡6],,,.2)()MS,,,MS[7¡8].MS(Probabilitydensityfunction,PDF),PDF,,,.3)MS,[9],[10],MS[11¡12],(Particleswarmoptimization,PSO)MS[13],(Meanshiftbeliefpropagation,MSBP)[14]MS(Beliefpropagation,BP),BP,40235MS.[15],..MS,.MS:,,[6].(),1),;2),,;3),,.,MS,.,MS,,,.,,,,.,,,.,.1MS(HMS)1.1,,,.,MS[16¡17],,,,.,,.[18]:.,,,,,.[17],MS(1),fxxxngNn=12R5,xxxxxxxxx.f¾;s(xxx)=mXj=1Xxxxn2setj!ð°°°xxxn¡xxx¾;s°°°°2!xxxnmXj=1Xxxxn2setj!ð°°°xxxn¡xxx¾;s°°°°2!(1)(1),xxxn=fi;j;r;g;bgfr;g;bgfi;jg5D;Nmsetj(j=1;¢¢¢;m),m;!(¢)¾,¾;s,.5(1),d,CCCk=fCCCxxxk;CCCuuukg,k=1;¢¢¢;d,CCCxxxk,CCCuuuk,nk.,.1RGB:,5.,,.,¾38,10.,,,14,84,(a)(a)Playerregion(b)(b)5Dclusteredplayer(c)(c)Faceregion(d)(d)5Dclusteredface15DFig.1Twoobjectmodelregionswithdistinctcolordistributionsandtheir5Dclusteringanalysis4:MeanShift403.1.2MS(fg),.(bg),MS.,,CCCk(k=1;¢¢¢;d).,,.,,,:1)xxxc=xxx¤(xxx¤);2),,CCCk,fgbg,(CCCk2fg;jCCCxxxk¡xxxcjTdCCCk2bg;(2)xxxcbg,Td.3),,.,2),fgCCCk,L(k)[19],Lfgbg,(CCCk2fg;L(k)0CCCk2bg;L(k)·0(3)4),CCCkxxxc=XCCCk2fgCCCxxxk¢H(CCCuuuk)XCCCk2fgH(CCCuuuk)(4)H(CCCuuuk)HSVH,.5)2)»4)xxxc,.,,.,,ddfdb(d=df+db).,:SSSxxxk=fCCCxxxk;CCCuuukg,k=1;¢¢¢;df.21,.,,.(a)(a)Playermodel(b)(b)Facemodel2Fig.2Objectmodelinthedescriptionofclustermodalpoints1.3SSSx,(PDF),.SSSx=fxxxi;uuuigNi=1PDFP(xxx;uuu),xxxi,uuui=u(xxxi)xxxifr;g;bg,P(xxx;uuu)[20]^P(xxx;uuu)=1NNXi=1!ð°°°xxx¡xxxi¾°°°°2!kð°°°uuu¡uuuih°°°°2!(5)!(¢),k(¢)¾,h,!¾(¢),kh(¢).,,,^P(xxx;uuu)=1NfdfXk=1nkXxxxi2CCCk!ð°°°xxx¡xxxi¾°°°°2!kð°°°uuu¡uuuih°°°°2!(6)Nf=Pdfk=1nk.,SSSy=fyyyj;vvvjgMj=1,d0,CCCl=fCCCyyyl;CCCvvvlg,l=1;¢¢¢;d0,df0,db0,n0l.40435:SSSyyyl=fCCCyyyl;CCCvvvlg,l=1;¢¢¢;df0.,,,dfdf0.HSV,,CCCkCCCl.1)CCCuuuk,CCCvvvlH;2)k,CCCuuukCCCvvvlH,argminljCCCvvvl¡CCCuuukjjCCCvvvl¤¡CCCuuukjTc,l.l,CCCkCCCl;l,.1kdf,1ldf0,Tc.CCCkCCCl,l=k=1;¢¢¢;d,3(a).,.,MS(3(b)).,,,,(7).(7),±(¢)delta,CCClCCCk1,0;Np;n0l;nkl0;k.,,,d.(a)(a)Pairsofmatchundercluster(b)(b)Pointbypointmatchundernon-cluster3Fig.3Clusterandnon-clustermatches1.4MS,.xxx¤,yyy,,(7)(8).J,.MS(8).yyyrL(yyy)=¡rJ(yyy)J(yyy)(9)rJ(yyy)(10).(10),4yyyj=(yyyj¡yyy),4xxxi=(xxxi¡xxx¤),!0(¢).(11).(11)yyyJ(SSSxxxk;SSSyyyl)=1Npdf0Xl=1dfXk=1·±(CCCl;CCCk)µn0lXyyyj2CCClnkXxxxi2CCCk!¾(yyyj¡xxxi)kh(vvvj¡uuui)¶¸(7)J(SSSxxxk;SSSyyyl)=1Npdf0Xl=1dfXk=1·±(CCCl;CCCk)µn0lXyyyj2CCClnkXxxxi2CCCk!¾((yyyj¡yyy)¡(xxxi¡xxx¤))kh(vvvj¡uuui)¶¸(8)rJ(yyy)=df0Xl=1dfXk=1·±(CCCl;CCCk)µn0lXyyyj2CCClnkXxxxi2Ck(4xxxi¡4yyyj)!0¾(4yyyj¡4xxxi)kh(vvvj¡uuui)¶¸(10)ms(yyy)=df0Xl=1dfXk=1·±(CCCl;CCCk)µn0lXyyyj2CCClnkXxxxi2CCCk(yyyj¡xxxi)!¾(4yyyj¡4xxxi)kh(vvvj¡uuui)¶¸df0Xl=1dfXk=1·±(CCCl;CCCk)µn0lXyyyj2CCClnkXxxxi2CCCk!¾(4yyyj¡4xxxi)kh(vvvj¡uuui)¶¸+xxx¤(11)4:MeanShift405,,FGF[7],[17],,:,,MSCCCxxxk=(1=nk)Pnkxxxi2CCCkxxxi,CCCyyyl=(1=n0l)Pn0lyyyj2CCClyyyj.,,n0lXyyyj2CCClnkXxxxi2CCCkyyyj!¾((yyyj¡yyy)¡(xxxi¡xxx¤))kh(vvvj¡uuui)¼nkXxxxi2CCCkn0l¢CCCyyyl¢!¾((CCCyyyl¡yyy)¡(xxxi¡xxx¤))kh(vvvCCCyyyl¡uuui)¼nk¢n0l¢CCCyyyl¢!¾((CCCyyyl¡yyy)¡(CCCxxxk¡xxx¤))£kh(vvvCCCyyyl¡uuuCCCxxxk)¼nk¢n0l¢CCCyyyl¢!¾(¢)kh(¢)(12)n0lXyyyj2CCClnkXxxxi2CCCkxxxi!¾((yyyj¡yyy)¡(xxxi¡xxx¤))kh(vvvj¡uuui)¼n0l¢nk¢CCCxxxk¢!¾(¢)kh(¢)(13),¼nk¢n0l¢!¾(¢)kh(¢)(CCCyyyl¡CCCxxxk),(11)HMS(14).,,[7,21]J(SSSxxxk;SSSyyyl)=1Npdf0Xl=1dfXk=1·±(CCCl;CCCk)£µn0lXyyyj2CCClnkXxxxi2CCCk!µ°°°°µ4yyyjps¾¶¡µps4xxxi¾¶°°°°2¶kh(vvvj¡uuui)¶¸(15),s.,Jyyys,yyys(16)(17),,!¾
本文标题:一种分层Mean Shift目标跟踪算法
链接地址:https://www.777doc.com/doc-633556 .html