计算机视觉new

计算机视觉new

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\nMarr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2.5D3D\nMultiobjectTrackingasMaximumWeightIndependentSetThispaperaddressestheproblemofsimultaneoustrackingofmultipletargetsinavideo.Wefirstapplyobjectdetectorstoeveryvideoframe.Pairsofdetectionresponsesfromeverytwoconsecutiveframesarethenusedtobuildagraphoftracklets.Thegraphhelpstransitivelylinkthebestmatchingtrackletsthatdonotviolatehardandsoftcontextualconstraintsbetweentheresultingtracks.Weprovethatthisdataassociationproblemcanbeformulatedasfindingthemaximum-weightindependentset(MWIS)ofthegraph.Wepresentanew,polynomial-timeMWISalgorithm,andprovethatitconvergestoanoptimum.Similarityandcontextualconstraintsbetweenobjectdetections,usedfordataassociation,arelearnedonlinefromobjectappearanceandmotionproperties.Long-termocclusionsareaddressedbyiterativelyrepeatingMWIStohierarchicallymergesmallertracksintolongerones.Ourresultsdemonstrateadvantagesofsimultaneouslyaccountingforsoftandhardcontextualconstraintsinmultitargettracking.Weoutperformthestateoftheartonthebenchmarkdatasets.\nOverviewoftheApproachStep1:Weapplydetectorsofasetofobjectclassestoallvideoframes.Eachdetectionischaracterizedbyadescriptorthatrecordsthefollowingpropertiesofthecorrespondingboundingbox:location,size,andthehistogramsofcolor,intensitygradients,andopticalflow.\nStep2:Thebestmatchingdetectionsaretransitivelylinkedacrossvideointodistincttracks,whosetotalnumberisunknownapriori.Thisisdoneunderthehardconstraintthatnotwotracksmaysharethesamedetection,topreventimplausiblevideointerpretations.Inaddition,thelinkingisinformedbyspatiotemporalrelationshipsbetweenthetracks,whichprovideforsoftconstraints.Tothisend,webuildagraph,wherenodesrepresentcandidatematchesfromeverytwoconsecutiveframes,referredtoastracklets;nodeweightsencodethesimilarityofthecorrespondingmatches;andedgesconnectnodeswhosecorrespondingtrackletsviolatethehardconstraints.Giventhisattributedgraph,dataassociationisformulatedasthemaximum-weightindependentset(MWIS)problem.MWISistheheaviestsubsetofnon-adjacentnodesofanattributedgraph.Conveniently,MWISoftheentiregraphisequivalenttoaunionoftheMWISsolutionsofindependentsubgraphs.Thisallowsustoconductmultitargettrackingonline.WepresentanewMWISalgorithmthatisguaranteedtoconvergetoanoptimum.\nStep3:Intrinsictargetpropertiesandpairwisecontext,usedinStep2,arelearnedonline,asthetrackskeepaccumulatingstatisticalevidenceofthetargets.TherelativesignificanceofthesepropertiesforeachtrackislearnedsoastominimizetheMahalanobisdistancesofdetectionswithinthesametrack,andmaximizetheMahalanobisdistancesbetweendetectionsfromdistincttracks.Step4:Toaddresslong-termocclusions,weiterateStep2andStep3tomergeorsplittrackssoastoincreasethetotalweightoftheMWIS,untilconvergence.\nPCA主成分分析（PrincipalComponentAnalysis，PCA）是一种掌握事物主要矛盾的统计分析方法，它可以从多元事物中解析出主要影响因素，揭示事物的本质，简化复杂的问题。计算主成分的目的是将高维数据投影到较低维空间。鲁棒性鲁棒性就是系统的健壮性。它是在异常和危险情况下系统生存的关键。比如说，计算机软件在输入错误、磁盘故障、网络过载或有意攻击情况下，能否不死机、不崩溃，就是该软件的鲁棒性。所谓“鲁棒性”，是指控制系统在一定（结构，大小）的参数摄动下，维持某些性能的特性。\nHOG（梯度方向直方图检测器）HOGdescriptors是应用在计算机视觉和图像处理领域，用于目标检测的特征描述器。这项技术是用来计算局部图像梯度的方向信息的统计值。ImplicitShapeModel(ISM)（绝对形状模型）\n疑问：1.Occlusionn.闭塞?2.MaximumWeightIndependentSet?3.Hard&softconstraints(硬件&软件系统规定参数)?