2)以聚类块模式点描述目标模型,利用邻域一致性计算目标运动平移量,给出分层的MS匹配迭代跟踪算法.将N个数据点收敛于d个模式点,再收敛于目标质心模式点.与经典方法相比,分层处理提高了目标模型表达与匹配的鲁棒性,无需逐点计算高斯核,大大节省了计算量,但是增加了聚类的运算.
本文算法存在的主要问题是:在跟踪过程中,目标(参考)模型描述没有在线更新;环境的变化可能导致目标的前景颜色发生较大变化,而本文的匹配误差允许阈值不能设定的太大,这样可能使目标丢失而跟踪失败.因此将来的工作是着重解决如何在线更新目标(参考)模型描述的问题,以获得更为鲁棒的跟踪,以及研究效率更高的聚类算法.
References
1ComaniciuD,RameshV,MeerP.Real-timetrackingofnonrigidobjectsusingmeanshift.In:ProceedingsofIEEEComputerSocietyConferenceonComputerVisionandPat-ternRecognition.HiltonHeadIsland,USA:IEEE,2000.142 1492FukanagaK,HostetlerLD.Theestimationofthegradi-entofadensityfunctionwithapplicationsinpatternrecog-nition.IEEETransactionsonInformationTheory,1975,21(1):32 403ComaniciuD,RameshV,MeerP.Kernel-basedobjecttracking.IEEETransactionsonPatternAnalysisandMa-chineIntelligence,2003,25(5):564 5754BircheldST,RangarajanS.Spatiogramsversushistogramsforregion-basedtracking.In:ProceedingsofIEEECom-puterSocietyConferenceonComputerVisionandPatternRecognition.SanDiego,USA:IEEE,2005.1158 11635LiPH.Aclusteringbasedcolormodelandfastalgorithmforobjecttracking.In:Proceedingsofthe18thInternationalConferenceonPatternRecognition.HongKong,China:IEEE,2006.671 6746LiPei-Hua.Animprovedmeanshiftalgorithmforobjecttracking.ActaAutomaticaSinica,2007,33(4):347 354(李培华.一种改进的MeanShift跟踪算法.自动化学报,2007,33(4):347 354)
7YangCJ,DuraiswamiR,DavisLS.E cientMean-Shifttrackingviaanewsimilaritymeasure.In:ProceedingsofIEEEComputerSocietyConferenceonComputerVision
一种分层Mean Shift目标跟踪算法
andPatternRecognition.SanDiego,USA:IEEE,2005.176 183
8YangCJ,DuraiswamiR,ElgammalA,DavisLS.Real-timeKernel-basedTrackinginJointFeature-spatialSpaces,TechnicalReportCS-TR-4567,ComputerScienceDepart-ment,UniversityofMaryland,USA,20049ElgammalA,DuraiswamiR,DavisLS.Probabilistictrack-inginjointfeature-spatialspaces.In:ProceedingsofIEEEComputerSocietyConferenceonComputerVisionandPat-ternRecognition.Madison,USA:IEEE,2003.781 78810LeungAP,GongSG.Mean-Shifttrackingwithrandom
sampling.In:Proceedingsofthe17thBritishMachineVi-sionConference.Edinburgh,British:UniversityTrierPress,2006.729 73811NaeemA,MillsS,PridmoreT.Structuredcombinationof
particle lterandkernelMean-Shifttracking.In:Proceed-ingsofthe21stInternationalConferenceonImageandVi-sionComputing.GreatBarrierIsland,NewZealand:IEEE,2006.1 612HanB,JooSW,DavisLS.Probabilisticfusiontrackingus-ingmixturekernel-basedBayesian ltering.In:Proceedingsofthe11thInternationalConferenceonComputerVision.Janeiro,Brazil:IEEE,2007.1 813ThomasM,KambhamettuC.AnapproximationtoMean-Shiftviaswarmintelligence.In:Proceedingsofthe18thIn-ternationalConferenceonToolswithArti cialIntelligence.WashingtonD.C.,USA:IEEE,2006.583 59014ParkM,LiuYX,CollinsRT.E cientmeanshiftbelief
propagationforvisiontracking.In:ProceedingsofIEEEConferenceonComputerVisionandPatternRecognition.Alaska,USA:IEEE,2008.1 815HanB,ComaniciuD,ZhuY,DavisLS.Sequentialkernel
densityapproximationanditsapplicationtoreal-timevi-sualtracking.IEEETransactionsonPatternAnalysisandMachineIntelligence,2008,30(7):1186 119716MiguelA,´Carreira-Perp n´an.Fastnonparametricclustering
withGaussianblurringMean-Shift.In:Proceedingsofthe23rdInternationalConferenceonMachineLearning.Pitts-burgh,USA:IEEE,2006.153 16017ZhangK,TangM,KwokJT.Applyingneighborhoodcon-sistencyforfastclusteringandkerneldensityestimation.In:ProceedingsofIEEEComputerSocietyConferenceonComputerVisionandPatternRecognition.SanDiego,USA:IEEE,2005.1001 100718ComaniciuD,RameshV,MeerP.Thevariablebandwidth
meanshiftanddata-drivenscaleselection.In:Proceedingsofthe8thIEEEInternationalConferenceonComputerVi-sion.Vancouver,Canada:IEEE,2001.438 44519CollinsRT,LiuYX,LeordeanuM.On-lineselection
ofdiscriminativetrackingfeatures.IEEETransactionsonPatternAnalysisandMachineIntelligence,2005,27(10):1631 164320ParzenE.Onestimationofaprobabilitydensityfunction
andmode.AnnalsofMathematicalStatistics,1962,33(3):1065 107621RangarajanA,ChuiH,BooksteinFL.Thesoftassignpro-crustesmatchingalgorithm.In:Proceedingsofthe15thIn-ternationalConferenceonInformationProcessinginMedi-calImaging.Poultney,USA:IEEE,1997.29 42
许海霞
湖南大学电气与信息工程学院
自动化重点实验室博士研究生.主要研究方向为计算机视觉,目标跟踪和摄像机标定.本文通信作者.E-mail:xhxia2002@(XUHai-Xia
Ph.D.candidateat
theKeyLaboratoryofAutomation,In-stituteofElectricalandInformationEngineering,HunanUniversity.Herresearchinterestcoverscomputervision,objecttracking,andcameracalibration.Correspondingauthorofthispaper.)
王耀南湖南大学电气与信息工程学院
教授.主要研究方向为机器视觉,智能图像处理和智能机器人.E-mail:yaonan@(WANGYao-Nan
Professoratthe
InstituteofElectricalandInformationEngineering,HunanUniversity.Hisre-searchinterestcoverscomputervision,intelligentimageprocessing,andintelligentrobot.)
袁小芳湖南大学电气与信工程学院自