结合模型匹配与特征跟踪的人体上半身三维运动姿态恢复方法  

Combining Model Matching with Feature Tracking for 3D Human Upper Body Pose Recovering

作　　者：陈姝[1,2] 彭小宁[3,4] 

机构地区：[1]湘潭大学信息工程学院,湘潭411105 [2]智能计算与信息处理教育部重点实验室,湘潭411105 [3]中南大学信息科学与工程学院,长沙410083 [4]怀化学院计算机科学与技术系,怀化418008

出　　处：《计算机辅助设计与图形学学报》2012年第11期1455-1463,1470,共10页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61100139;61040009)

摘　　要：人体姿态空间的高维性及单目视频深度信息丢失,导致从单目视频恢复人体三维运动姿态非常困难,为此,利用特征跟踪的快速性及模型匹配的鲁棒性,提出一种无标记人体上半身三维运动跟踪方法.该方法利用匹配SIFT特征,并根据长度不变性约束建立优化目标函数,再采用迭代优化算法得到全局运动位姿;其他关节的姿态先根据逆运动学计算初始估计值,并通过模型匹配验证其可信度,当初始姿态估计错误时,则使用局部搜索获得关节姿态.实验结果表明,文中方法可以准确地恢复单目视频中人体上半身三维运动姿态.Recovering 3D human pose from monocular video sequences is a challenging problem due to the highly articulated structure and inherent depth ambiguity. By take fully advantage of the both benefits of feature tracking and model matching, a new method of markerless 3D human upper body pose tracking is presented. First, an effective object function is derived from the matched SIFT correspondences based on the length invariability constraint. Then, the global pose is obtained by an iterative optimization process. The pose of other joints is initially estimated by inverse kinematics, then, testified by model matching. A local search strategy is followed to estimate the optimal pose if the initial estimated pose is inaccurate. The experiment results show that the proposed algorithm can reconstruct 3D human upper body pose from monocular video sequences.

关 键 词：人体运动跟踪 梯度方向直方图 SIFT特征匹配 逆运动学 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]