Real-time human action recognition based on shape combined with motion feature

The visual analysis of human motion have become a direction of the leading edge of concern in computer vision. The problem of recognizing human actions in video have proven to be a difficult challenge for computer vision. A common trend is to combine shape and motion feature in Bag-of-word (BoW) framework. The BoW framework needs read the whole video for once recognition. It could not be applicable in real-time system. For solving this problem, this paper proposes a real-time processing human action method. Firstly, the motion information is used to locate the region of interesting area. And then the effective shape and motion description is found for the area. Finally, the SVM classifier is trained for the event recognition. And the recognition results on KTH human action datasets including a variety of person and action show that the accuracy and recall of our method is better than Jhuang and Dollar´s; and the process time is superior to Jhuang´s system.