提出一种基于光流特征与序列比对的行为识别算法。首先利用分层光流提取视频序列中的运动信息;然后用光流场的方向直方图构造相应行为的模板库和索引序列库;最后用序列比对方法实现行为识别。实验结果表明,该算法可在线进行人的典型行为识别,对目标尺度变化、小角度倾斜和旋转具有一定程度的鲁棒性。目前以该算法为核心的行为识别实验系统对图像尺寸为320×240的序列平均处理速度达到10fps。
This paper Proposed an action recognition algorithm based on optical-flow feature and sequence alignment. First motion information in the video image sequence extracted using hierarchical optical-flow, then using the direction histogram from optical-flow field, the common spatiotemporal template and the index sequences warehouse generated, finally action recognized via sequence alignment. The experimental result indicates that this algorithm achieve real-time human' s typical action recognition, robust to the object' s size change, certain incline degree and revolves. At present, the action recognition experimental system cored this algorithm will handle 320×240 image sequence average processing speed of 10 fps.