A multi-scale Conditional Random Field model for human action recognition

Human action recognition is an important issue in the field of computer vision. But the existing models put more emphasis on the single scale and little attention on multi-scale and multi-action mode in the motion. With an aim at this problem, this paper presents a human motion recognition method using multi-scale condition random field model. At the first, the trajectory of human movement, the human body posture characteristics as well as the limb movement are considered in order to extract the multi-scale feature. Then a multi-scale Conditional Random Fields model is proposed for human action recognition. The model can take full advantage of the context information of the action sequences, as well as mutual constraints and impact of information between different scales, and can solve the problem that people have multi-action at the same time.