视频中旋转与尺度不变的人体分割方法

A Rotation-and Scale-invariant Human Parts Segmentation in Videos

提出了一种旋转与尺度不变的人身体部位所在区域的视频分割方法.方法中不仅考虑到躯干与四肢之间的关系,还考虑到四肢之间的相互关系,通过空间与时间的连续性约束对每帧中各个可能的身体部位进行优化组合,并巧妙地用动态规划对非线性图模型进行优化,且不受运动目标尺度变化与各种翻转运动的影响.该方法首先用动态规划的优化方法得到每一帧中最优的N个身体部位组合,将每一个组合作为图模型中的一个节点,并用动态规划对所有帧中的各个组合所构成的网格状图结构进行优化,最终得到每一帧中最优的身体部位组合.实验结果表明,该视频分割方法不仅适用于行人视频,还适用于具有各种姿势的运动视频,且具有较好的鲁棒性.

This paper proposes a rotation- and scale-invariant method for human body parts segmentation. The proposed method considers not only the relationship between torso and limbs but also between limbs. The method optimizes the candidate assembly of body parts in each frame with spatial and temporal constraints, and uses dynamic programming to optimize a non-linear graph model smartly, which is rotation and scale invariant. First, it generates the best N combinations of human body parts using dynamic programming, each being a node in the graph. Then it optimizes the graph which is the grid made up of all the nodes in each frame, using dynamic programming to get the optimal human body part combination in each frame. Experiments show that this method can robustly get efficient and accurate results both on pedestrian videos and sports videos with any human poses.