【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.更多还原

【Abstract】 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.更多还原