单目视频中无标记的人体运动跟踪  

Markerless Human Motion Tracking from Monocular Videos

作　　者：陈坚[1] 王文成[1] 吴恩华[1] 

机构地区：[1]中国科学院软件研究所计算机科学重点实验室

出　　处：《计算机辅助设计与图形学学报》2005年第9期2033-2039,共7页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(60173022;600333010);国家重点基础研究发展规划项目(2002CB312102)

摘　　要：提出一种人体运动跟踪算法,从无关节标记的单目视频中获取人体运动.利用一个带外观模板的人体关节模型,通过学习得到的运动模型及基于外观模型的相似性计算,巧妙地利用粒子滤波的概率密度传播策略鲁棒地跟踪普通单目视频中的人体运动.当出现跟踪丢失时,能在后续序列中自动恢复正确跟踪,且能较好地处理遮挡和自遮挡问题.实验表明,该算法鲁棒性好,跟踪结果令人满意.In this paper, a novel approach is proposed for tracking markerless human motion ill monocular videos to capture the articulate motion data. With an articulated human model constructed, the new approach uses the probability density propagation of the particle filters through the learnt motion model and likelihood computing with the appearance models to track the human motion. The method is capahle of automatically recovering from tracking failures. It can also process the occlusion and auto-occlusion problem correctly,Experimental results from real monocular videos show that the new approach is robust aud the tracking results are satisfactory.

关 键 词：人体运动跟踪 人体模型 粒子滤波 运动模型 外观模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]