资源描述：
《计算机视觉new》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、计算机视觉*使计算机通过二维图像认知三维环境信息1>.感知三维环境中物体的几何信息：形状，位置，姿态，运动2>.描述，存储，识别，理解Marr的计算视觉理论三个阶段：三个层次：计算理论表达与算法硬件实现三个阶段：基元图2.5D3DMultiobjectTrackingasMaximumWeightIndependentSetThispaperaddressestheproblemofsimultaneoustrackingofmultipletargetsinavideo.Wefirstapplyobje
2、ctdetectorstoeveryvideoframe.Pairsofdetectionresponsesfromeverytwoconsecutiveframesarethenusedtobuildagraphoftracklets.Thegraphhelpstransitivelylinkthebestmatchingtrackletsthatdonotviolatehardandsoftcontextualconstraintsbetweentheresultingtracks.Weproveth
3、atthisdataassociationproblemcanbeformulatedasfindingthemaximum-weightindependentset(MWIS)ofthegraph.Wepresentanew,polynomial-timeMWISalgorithm,andprovethatitconvergestoanoptimum.Similarityandcontextualconstraintsbetweenobjectdetections,usedfordataassociat
4、ion,arelearnedonlinefromobjectappearanceandmotionproperties.Long-termocclusionsareaddressedbyiterativelyrepeatingMWIStohierarchicallymergesmallertracksintolongerones.Ourresultsdemonstrateadvantagesofsimultaneouslyaccountingforsoftandhardcontextualconstrai
5、ntsinmultitargettracking.Weoutperformthestateoftheartonthebenchmarkdatasets.OverviewoftheApproachStep1:Weapplydetectorsofasetofobjectclassestoallvideoframes.Eachdetectionischaracterizedbyadescriptorthatrecordsthefollowingpropertiesofthecorrespondingboundi
6、ngbox:location,size,andthehistogramsofcolor,intensitygradients,andopticalflow.Step2:Thebestmatchingdetectionsaretransitivelylinkedacrossvideointodistincttracks,whosetotalnumberisunknownapriori.Thisisdoneunderthehardconstraintthatnotwotracksmaysharethesame
7、detection,topreventimplausiblevideointerpretations.Inaddition,thelinkingisinformedbyspatiotemporalrelationshipsbetweenthetracks,whichprovideforsoftconstraints.Tothisend,webuildagraph,wherenodesrepresentcandidatematchesfromeverytwoconsecutiveframes,referre
8、dtoastracklets;nodeweightsencodethesimilarityofthecorrespondingmatches;andedgesconnectnodeswhosecorrespondingtrackletsviolatethehardconstraints.Giventhisattributedgraph,dataassociationisformulatedasthemaximum-weight
