Human action recognition with topic-relative conditional random field model

Human action recognition is a challenging filed in computer vision. In this paper, a novel probabilistic graphical model, called topic-relative conditional random field(TCRF), is firstly proposed. The model is constructed by adding a topic node and using a triangular-chain structure in the top layer of the linear-chain conditional random field(LCRF) to overcome the drawback of independent and identical distribution in LCRF. Then, we define a dynamic region for each action and the discriminative features are extracted by using a hierarchical energy method. Lastly, two popular probabilistic graphical models, HMM and LCRF, and the proposed TCRF model are evaluated on our database, the experimental results show the effectiveness of the proposed method.