城区交通场景中运动目标视觉跟踪方法研究

重庆大学硕士学位论文城区交通场景中运动目标视觉跟踪方法研究姓名：詹建平申请学位级别：硕士专业：控制理论与控制工程指导教师：黄席樾2010-04BhattacharyyaABSTRACTAutomobileasanimportantvehicleinthemodernworld,itbringssomuchconvenienttopeople.Justaseverycornhastwosides,thetrafficaccidentsalsothreatenthepeople’sliviesandpropertysafety.Decreasingthetrafficaccidentandraisingthesafetylevelofroadtrafficbecomethecryingrequestofthesociety.SoStudyingonvisiontrackingmethodofmovingobjectunderurbantrafficscenesandcoretechnologytodecreasethetrafficaccidenthasanimportantacademicsignificanceandsocialvalues.Urbanvehicletargettrackingisanimportantpartofthevehicleactivesafetyearlywarningsystem.Meanwhile,themovingtargettrackingisanimportantissueofcomputervisionresearch.Thisdissertationwhoseresearchobjectsarevehicletargetsundertheurbantrafficscenariosisbasedontheintelligentvehicleactivesafetyforewarningsystem.Inthisdissertation,weresearchedtheapplicationofMeanshiftParticlefilterinthefieldofobjecttracking.First,amethod,whichcombinesMeanshiftandKalmanfilter,isproposed.Atthesametime,anocclusioncoefficientisdesignedastheevidenceoftargetocclusion.Second,histogram-basednuclearparticlefilteralgorithm,putforwardanadaptivetemplateupdatingmethod,effectivelyraisingthereal-timealgorithm,whileavoidingexcessiveupdateofthetargetmodelaswellasfrequentupdate.Mainstudycontenthasbeensummarizedasfollows:Describeourcurrenttrafficconditions,andoutlinethemainreasonsofroadtrafficaccidents.Highlightthesignificancetoachievethehumanvisualsystemfunctionofcomputervisiontechnologywhichintegratesthecomputerscience,machinevision,imageengineering,patternrecognition,artificialintelligenceandotheradvancedtechnologies.Analysisthemaintargettrackingmethodsandtechnicalchallengescurrentlyfacing.AfteranalyzingthetheoreticlimitationoftheMeanshifttotrackfastmovingtargetsincomplexbackground,amethod,whichcombinesMeanshiftandKalmanfilter,isproposed.Firstly,theinitialpositionofMeanshiftispredictedbyKalmanfilteratpresent,andthentheMeanshiftisutilizedtotrackthetargetpositionaroundtheinitialposition.Meanwhile,theBhattacharyyacoefficientisadoptedtomeasurethecomparabilitybetweenthetargetmodelandthecandidatemodel.Anocclusioncoefficientisproposedastheevidenceofocclusion.Experimentsbasedonthevehicleobjectsinthecityarecarriedout,andthesimulationresultsshowthatthetrackingstabilityforthefastmovingtargets,eveninocclusion,isimprovedsignificantlywiththeproposedmethod.Discussthehistogram-basednuclearparticlefilter,andanalysistheadvantagesofparticlefilterappliedtoobjecttrackingrelativetoMeanshiftandKalmanfilteralgorithm.AfterhighlightingParticlefilteralgorithmmodelingissues,includingthedynamicmodelandobservationmodel,agoalofadaptivetemplateupdatingmethodisproposed.Describethechoiceofkernelfunction,bandwidthdeterminationandthedeterminationoftargetlocation,then,showthestructureandflowchartofthealgorithminthischapter.Finally,testthealgorithmofthischapterbyexperimentsandgetaverygoodeffort.Finally,theconclusionofwholeresearchworkinthepaperisgiven.Furthermore,thefurtherworkandresearchprospectsareintroduced.KeywordsComputerVision,TargetTracking,Meanshift,ParticleFilter,UrbanTrafficScenes.1111.12009916[1]2009818001851271856993923875711.8ITS(IntelligentTransportationSystem)65%57%95%94%90%[2]121.2,2009720[3]2009107193298661283364.11012ITS,[4]1390%CPUITSITSITS190%141.3[5-6]ComputerVision[7][8]VisionTracking[9]Multi-TargetsVisualTracking1997[10]VisualSurveillanceAndMonitoring,VSAM911[11]VideoImageProcessing15System,VIPS[12][13],[14],,19821986CCD16()()P2OT()[15][16][17][18]1TT2173Sobel4FFT,,[19]1Multi-PointCorrelation[20]182RegionTemplateCorrelation[21]3WaveletTransformCorrelation[22-24]20801,,1t-t2,ab-abg--MonteCarlo[25]kalman,[26]19123451.41.4.1,,,1101.4.2,21122.1MeanshiftFukunaga1975201995YizongChengYizongChengComaniciuMeer[27-29]Comaniciu,[29]ComaniciuComaniciuChangjiangYangNummiaroCollins2.22122.2.1X(|)fq⋅qqˆq,ˆqˆ(|)fq⋅ˆf()Nmsm2s{}1...=iiNXx=ˆmˆs1ˆxXxNm∈=∑(2.1)1ˆˆ()xXxNdm∈=-∑(2.2)2.2.22040502135060binbinUniform(Trangle)(Epanechikov)(Biweight)(Gaussian)(Cosinusarch)(DoubleExponential)(DoubleEpanechnikov)2.2.3[30-34][31]1/iwn=[30]1/iwn=[34][31,32][33][34][34]1,2,...{}jjy=1y,1,2,...ˆ{()}hkjfj=1,2,...{}jjy=,1,2,...ˆ{()}hkjfj=[34]()Kx1,2,...{}jjy=,1,2,...ˆ{()}hkjfj=214,1,2,...ˆ{()}hkjfj=2.3n,1,...diXRin∈=X11ˆ()()nHiifxKxxn==-∑(2.3)1/21/2()||()HKxHKHx--=,Hdd×[56]d()KX||||()1lim||||()0()0()dddRdxRTKRKxdxxKxxKxdxxxKxdxcI→∞⎧=⎪⎪=⎪⎨=⎪⎪⎪=⎩∫∫∫(2.4)KcH2HhI=()Kx2,()(||||)kdKxckx=(2.5),,kdc,()Kx1,(2.3),2,()1ˆ(||||)nkdihKxdicxxfknhh=-=∑(2.6)2,d',(),()212ˆˆ()nKihKxhKxidicxxffxxKnhh+=⎛⎞-∇≡∇=-⎜⎟⎜⎟⎝⎠∑(2.7)()Kx[)0,x∈∞'()()gxkx=-(2.8)()gx()Gx2,()()gdGxcgx=(2.9),gdc()Kx()Gx215()gx(2.7)()2,,21221,22112ˆ()2nKdihKidiniiniKdidniiicxxfxxxgnhhxxxghcxxgxnhhxxgh+==+==⎛⎞-∇=-⎜⎟⎜⎟⎝⎠⎡⎤⎛⎞-⎢⎥⎜⎟⎜⎟⎡⎤⎛⎞-⎢⎥⎝⎠=-⎢⎥⎜⎟⎢⎥⎜⎟⎛⎞-⎢⎥⎝⎠⎣⎦⎢⎥⎜⎟⎜⎟⎢⎥⎝⎠⎣⎦∑∑∑∑(2.10)21niixxgh=⎛⎞-⎜⎟⎜⎟⎝⎠∑xG2,,1ˆ()ngdihGdicxxfxgnhh=⎛⎞-=⎜⎟⎜⎟⎝⎠∑(2.11)21,21niiihGniixxxghmxxxgh==⎛⎞-⎜⎟⎜⎟⎝⎠=-⎛⎞-⎜⎟⎜⎟⎝⎠∑∑(2.12)g()1x=11()()niimxxxn==-∑(2.13)2.1Fig.2.1Meanshiftvectordiagram2.1(2.13)x(2.12)216()gxixx,()hGmxx(2.13)ix,()hGmxxx(2.12)2()ixxgh-(2.12)()gx2.4Epa