一种基于实例的单视频人体姿态分析方法.pdfVIP

188 EGVR’2009中国江苏淮安 一种基于实例的单视频人体姿态分析方法掌 李毅，孙正兴+ (南京大学计算机软件新技术国家重点实验室，南京，210093) An toHumanPoseEstimationfromMonocular Example-basedApproach Images Li Yi，Sun Zhengxing ofNovelSoftware (StateLaboratory Technology，NanjingUniversity。Nanjing210093) Key +Corresponding Abstract：Humanestimationfrommonocular isa workin vision．Based pose imagesequenceschallengecomputer humanestimation olla databaseconstructedmotion data。an to example using capture example-basedapproach pose context thehumansilhouette frommonocular is USe todescribe imagesequencesproposed．Firstly，weshape the database detectedfromvideoframes。andcandidatefrom silhouette get poses example by matching；Secondly。 webuild andstatisticalmodelof out estimationfromthese probability motion，andcarrypose candidates；Finally， Kernel isusedtosmooththemotion．The method 3D fromvideosand Regression proposedeffectivelyanalyzepose method couldsolvetheorientation itisinvarianttoview effectivenessofthis ambiguitiesproblem，also points．The isverifiedonvideosof and human walking，runningjumpingposes． method。silhouette context Key estimation，example-basedmatching，shape words：pose 摘要：单视频人体姿态分析是计算机视觉中的难点问题．利用已有的运动捕获数据构建姿态实例库，提出了一 种基于实例的单视频人体姿态分析方法．该方法首先采用基于形状上下文的轮廓匹配方法得到人体姿态的侯选 集；其次，通过统计建模和转移概率建模得到人体姿态的参教估计；最后，使用核规约方法平滑运动得到人体连 续姿态．该方法能