Mining human movement evolution for complex action recognition

In this paper, a novel and efficient system is proposed to capture human movement evolution for complex action recognition. First, camera movement compensation is introduced to extract foreground object movement. Secondly, a mid-level feature representation called trajectory sheaf is proposed to capture the temporal structural information among low-level trajectory features based on key frames selection. Thirdly, the final video representation is obtained by training a sorting model with each key frame in the video clip. At last, the hierarchical version of video representation is proposed to describe the entire video with higher level representation. Experimental results demonstrate that the proposed method achieves state-of-the-art performance on UCF Sports, and comparable results on several challenge benchmarks, such as Hollywood2 and HMDB51 dataset.