基于ROI特征匹配融合的图像多目标跟踪算法

13320083JournalofImageandGraphicsVol.13,No.3Mar.,2008:2007207230;:2007212203:(1980),2003,E2mail:[e-mail](,410073),,(standarddeviationmap,STDM),(regionofinterest,ROI);;,,:TP391141:A:100628961(2008)0320580206Multi2targetTrackingBasedonFusionofROIFeatureMatchingLEILin,LIZhi2yong,SUYi(SchoolofElectronicScienceandEngineering,NationalUniversityofDefenseTechnology,Changsha410073)AbstractTherearetwoproblemswhentrackingmultipletargetsinsequenceimages.Firstly,thetargetpskinematicstatecannotbeestimatedaccuratelywhentargetpsabruptmaneuverhappens.Secondly,theimagematchingmethodcannotdiscriminatethetargetswhichbelongtothesamecategory.Toresolveaboveproblems,anovelmulti2targettrackingmethodbasedonfusionoftargetpsfeaturematchingisproposed.Onbaseoftargetdetectioninlocalstandarddeviationmap(STDM),theregionofinterestps(ROI)imagematchingresultisusedtoreducetheerrorofstateestimationmatching,andthestateestimationmatchingisusedtoreducetheambiguityofimagematching.Undertheglobaloptimalassociationcostframe,fusionofabovetwomatchesisrealizedtoimprovetheaccuracyoftracking,whichcanresolvethecomplexmulti2targettrackingproblemavailably.Keywordsmulti2targettracking,targetdetection,associationcost,matching1,,,,,,,,3[1]:(1),,;(2)[2,3],,;(3)[4],Kalman,,,,,,,[5],;,3:ROI581,11Fig11Twosamplemulti2targettrackingscenarios1(a),2,,,,50%,1(b),3,,,,50%,,,,;;,,221132:,(standarddeviationmap,STDM)(ROI),ROI,ROI;,2;,2,,,2ROIFig12Flowchartofmulti2targettrackingbasedonfusionofROIfeaturematching212,2[6]kM(k)Z(k)={Zm(k),m=0,1,,M(k)},k-1N(k-1)T(k-1)={T(n)(k-1),n=0,1,,N(k-1)}()a(k,m,n)=1Zm(k)T(n)(k-1)0a(k)={a(k,m,n);m=0,1,,M(k);n=0,1,,N(k-1)},m=0,n=0,a(k,0,n)T(n)(k-1),kn,a(k,m,0)mZm(k)T(k-1),kmROIZm(k)={zm(k),m(k)},zm(k),m(k)ROI,Zm(k)T(n)(k-1)(state,S)S(k,m,n)(image,I)I(k,m,n)58213(k,m,n)=I(k,m,n)S(k,m,n)(1)a(k)C(k|a(k))=M(k)m=0N(k-1)n=0a(k,m,n)c(k,m,n)(2)a(k)=argmina(k)C(k|a(k))(3),c(k,m,n)=-ln(k,m,n)(4),3ROI,,,(automatictargetdetection,ATD),ROI[7]Snorrason,,,;STDMROI,STDM;ROI311ROI,,;NNSTDM,(x,y):STDM(x,y)=1N2x+Mi=x-My+Mj=y-Mf(i,j)2-1N2x+Mi=x-My+Mj=y-Mf(i,j)2(5),f(i,j)(i,j),M=õN/2STDM,,[8],ROI312ROI,,ROI3:,ROI;,,;,3ROI34STDM31Fig13Sequenceimages1anddetectionresults42Fig14Sequenceimages2anddetectionresults4411[3],,3:ROI583,,,r=[x,y]Tf2,r^Ug,r^=U.r=Ar+b=a1a2a3a4xy+b1b2(6),f(r)=g(r^)+e+n,e,n02I,2I,fgL(U)=p[g(r^)|f(r)]=1(2I)K1K2exp-122IrK(f(r)-f-)-(g(r^)-g(U))2(7)f-g-(U)fgK[1,K1][1,K2],L(U),L(U)f,g(U)=rK(f(r)-f-)(g(r^)-g(U))rK(f(r)-f-)21/2^rK(g(r^)-g(U)21/2(8),rKf(r)=0,rKf(r)2=1,kmROIk-1nROII(k,m,n)=rKfn(r)gm(Ar+b)(m0,n0)(9)412,U,,,,,[9],=rKr^Kf(r)g(r^)exp(-Ar+b-r^2/D)(10):D0,=rKr^Kf(r)g(r^)exp(-Ar+b-r^2/D)rKr^Kf(r)g(r^)(Ar+b-r^)=rKf(r)g(Ar+b),0=9/9Af(r)g(r^)r(Ar+b-r^)Texp(-Ar+b-r^2/D)0=9/9bf(r)g(r^)(Ar+b-r^)exp(-Ar+b-r^2/D)(11),[9],(11),413,xk|k-1=Fk(xk-1|k-1,wk)xk|k-1,zk=Hk(xk|k-1,vk)xk|kd(k)=z(k)-H(k)^x(k|k-1)=z(k)-^z(k|k-1)R,R,p[d(k)]=e(-2/2)(2)R/2|S|(12),(k)=dT(k)S-1(k)d(k),Smn[10]S(k,m,n)=p[dm,n(k)]m0,n0PFm0,n=01-PDm=0,n0(13)58413,dm,n(k)=zm(k)-z^n(k|k-1)kmk-1nPD,PF5,,51154,5(b)5(c),,,65Fig15Originalimagesandpositionsoftracking6Fig16Movingtargetsptrajectories512,44{1,1,2,3},7,41,0,200s,2s5132,2,0,2S82S=2125m,,Singer74Fig17Truetrajectoriesoffourtesttargets84Fig18Simulatedobservedtrajectoriesoffourtesttargets,[0,255],[-/9,/9],e[-10,10],[019,111],0,2I=100,4(a)4ROI,100ROI,ROI51410099,,,4,12,ROI,,,,515MonteCarlo,,3:ROI5859Fig19Resultsofdifferentassociationalgorithmstosinglesimulation,=[i,j],ij,,,i,i1,ij,i,j=0,50MonteCarlostatefusionstate=014795013107011452010646012874014058012448010619011776012204013740012280010555010631012359016455fusion=018894010575010247010284010514018621010497010368010371010421018655010553010221010383010601018795,25%50%,6,,,:(1)ROI,;(2),,,,;(3)ROI,,,(References)1ShaoWen2kun,HuangAi2min,WeiQing.Researchonobjecttracking[J].ImageTechnology,2006,(1):1720.[,,.[J].,2006,(1):1720.]2TianYuan,TanTie2niu,SunHong2zan.Anovelrobustalgorithmforreal2timeobjecttracking[J].ACTAAutomaticaSinica.2002,28(5):851853.[,,.[J].,2002,28(5):851853.]3WuQX.Acorrelation2relaxation2labelingframeworkforcomputingopticalflow2templatematchingfromanewperspective[J].IEEETransactionsonPatternAnalysisandMachineIntelligence,1995,17(8):843851.4ComaniciuDorin,MeerPeter.Kernel2basedobjecttracking[J].IEEETransactionsonPatternAnalysisandMachineIntelligence,2003,25(5):564577.5NguyenDuyH,KayJohnH,OrchardBradleyJ,etal.Classificationandtrackingofmovinggroundvehicles[J].LincolnLaboratoryJournal,2002,13(2):275308.6Bar2ShalomY,KirubarajanT,GokberkCenk.Trackingwithclassification2aidedmultiframedataassociation[J].IEEETransactionsonAerospaceandElectronicSystems,2005,41(3):868878.7GreenbergShlomo,RotmanStanleyR,GutermanHugo,etal.Region2of2interest2basedalgorithmforautomatictargetdetectionininfraredimages[J].OpticalEngineering,2005,44(7):110.8MandalAchintyaK,PalSrimanta,DeArunK,etal.Novelapproachtoidentifygoodtracercloudsfromasequenceofsatelliteimages[J].IEEETransactionsonGeoscienceandRemoteSensing,2005,43(4):813818.9WakaharaToru,KimuraYoshimasa,TomonoAkira.Affine2invariantrecognit