单目视频人体三维运动高效恢复  

Efficient 3D Recovery of Human Motion in Monocular Video

作　　者：陈成[1,2] 肖俊[1,2] 庄越挺[1,2] 

机构地区：[1]浙江大学计算机科学与技术学院,杭州310027 [2]浙江大学人工智能研究所,杭州310027

出　　处：《计算机辅助设计与图形学学报》2009年第8期1118-1126,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(60525108;60533090);国家科技支撑计划课题(2006BAH02A13-4);中国博士后科学基金(20080431327);浙江省教育厅科研计划项目(Y200803033)

摘　　要：为解决计算机图形学和视觉领域的人体数据运动获取问题,提出一种从无标记点的单目视频恢复三维人体运动的方法.首先对人体侧影进行分析,获取躯干和末端节点位置信息;然后进行三维姿态优化.根据人体骨架特点,提出一个有效且计算简单的目标函数以及一种迭代优化策略,极大地减少了优化过程的计算量;设计了一个新颖的姿态序列恢复流程,克服了误差累积等传统跟踪方法的缺点.实验结果表明,文中方法可以准确地对视频中的复杂人体运动进行三维恢复.A new method of recovering 3D human motion from monocular videos is presented. First, extracted silhouettes are analyzed to derive 2D positions of spine and end sites. Then, 3D poses are recovered by optimizing an object function that encodes the correspondence between the analyzed silhouettes and a pose-parameterized 3D human skeleton. In order to reduce the computational complexity, an effective and computationally efficient object function is devised. A novel iterative optimization process that exploits the human skeleton structure is also proposed to boost the optimization. In addition, we present a new strategy to recover pose sequence that is free from error accumulation. Experimental results show that complex motions of a large variety of types can be recovered by the proposed method.

关 键 词：三维恢复 运动跟踪 姿态恢复 视频理解 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]