Action recognition based on hierarchical dynamic Bayesian network

Abstract

In this paper, a novel action recognition method is proposed based on hierarchical dynamic Bayesian network (HDBN). The algorithm is divided into system learning stage and action recognition stage. In the stage of system learning, the video features are extracted using deep neural networks firstly, and using hierarchical clustering and assisting manually, a hierarchical action semantic dictionary (HASD) is built. The next, we construct the HDBN graph model to present video sequence. In the stage of recognition, we first get the representative frames of unknown video using deep neural networks. The features are inputted into the HDBN, and the HDBN inference is used to get recognition results. The testing results show the proposed method is promising.