Real-time action recognition based on a modified Deep Belief Network model

This paper presents a real-time human action recognition method based on a modified Deep Belief Network (DBN) model. To recognize human actions, the positions of human joints are taken into account. Each action is made of a sequence of human joint positions. Since the classic DBN cannot deal with temporal information, the proposed method employs the conditional Restricted Boltzmann Machine (cRBM) to handle the human joint sequence. To verify the effectiveness of the proposed method, two skeletal representation datasets are used for testing. Experimental results show that the proposed method is able to achieve real-time human action recognition, and the recognition accuracy is comparable to state-of-the-arts methods.