Objectives The research on action recognition based on human skeleton data has made good progress. However, most of the existing methods only consider the spatial position information of joint points and ignore the regional change characteristics of joint points. In order to solve this problem, an action recognition method considering the regional characteristics of skeleton is proposed.
Methods First, based on the human body structure, the joints and bones are constructed into a spatiotemporal skeleton map which represents the human action characteristics. The skeleton map is divided into regions according to the law of human movement. Then, the coordinates of nodes in the skeleton graph and the angle change data of the nodes in the region are used as the inputs of the spatiotemporal graph convolution network. Finally, the prediction results of the two data streams are fused to realize human action recognition.
Results In order to prove the effectiveness of the proposed method, it is verified on Florence 3D dataset and dynamics action dataset, respectively. The results show that the accuracy of the proposed method reaches 91.1% on Florence 3D dataset, which is 1.9% higher than that of a single data stream. The accuracies of Top-1 and Top-5 on dynamics action dataset reach 32.4% and 54.2%, respectively.
Conclusions Compared with the existing methods, the proposed method is proved to have better recognition accuracy and higher effectiveness through multiple sets of experiments.
DownLoad: