资源描述：
《光流的经典文献,运动跟踪专业的同学必读 Performance o.pdf》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。
1、1PerformanceofOpticalFlowTechniques121J.L.Barron,D.J.FleetandS.S.Beauchemin12Dept.ofComputerScienceDept.ofComputingScienceUniversityofWesternOntarioQueen'sUniversityLondon,Ontario,N6A5B7Kingston,Ontario,K7L3N6AbstractWhiledi erentoptical
owtechniquescontinuetoappear,thereha
2、sbeenalackofquantitativeevaluationofexistingmethods.Foracommonsetofrealandsyntheticimagesequences,wereporttheresultsofanumberofregularlycitedopti-cal
owtechniques,includinginstancesofdi erential,matching,energy-basedandphase-basedmethods.Ourcomparisonsareprimarilyempirical,
3、andconcentrateontheaccuracy,reliabilityanddensityofthevelocitymeasurementstheyshowthatperformancecandi ersigni cantlyamongthetechniquesweimplemented.1IntroductionWithoutdoubt,afundamentalproblemintheprocessingofimagesequencesisthemea-surementofoptical
ow(orimagevelocity).T
4、hegoalistocomputeanapproximationtothe2-dmotion eld{aprojectionofthe3-dvelocitiesofsurfacepointsontotheimag-ingsurface{fromspatiotemporalpatternsofimageintensity[31,58].Oncecomputed,themeasurementsofimagevelocitycanbeusedforawidevarietyoftasksrangingfrompassivesceneinterpret
5、ationtoautonomous,activeexploration.Ofthese,taskssuchastheinferenceofegomotionandsurfacestructurerequirethatvelocitymeasurementsbeaccurateanddense,providingacloseapproximationtothe2-dmotion eld.Currenttechniquesrequirethatrelativeerrorsintheoptical
owbelessthan10%[10,36].Ve
6、rriandPoggio[58]havesuggestedthataccurateestimatesofthe2-dmotion eldaregen-erallyinaccessibleduetoinherentdi erencesbetweenthe2-dmotion eldandintensityvariations,whileothers(e.g.[4])arguethatthemeasurementofoptical
owisanill-posedproblem.Forthesereasonsithasbeensuggestedtha
7、tonlyqualitativeinformationcanbeextracted.Manymethodsforcomputingoptical
owhavebeenproposed{otherscontinuetoappear.Lacking,however,arequantitativeevaluationsofexistingmethodsanddirectBarron,FleetandBeaucheminIJCV12:1,pp43-77,19942comparisonsonasinglesetofinputs.Kearneyetal.
8、[37]discussedsourcesoferrorwithgradient-basedmethods.LittleandVerri[39]comparedpro
