PART AWARE GRAPH CONVOLUTION NETWORK WITH TEMPORAL ENHANCEMENT FOR SKELETON-BASED ACTION RECOGNITION

In recent years, skeleton-based human action recognition has attracted broad research interests, and methods based on graph convolution networks have demonstrated excellent performance. However, how to extract the distinguishing spatio-temporal information effectively remains an essential problem. To address the problem, we propose a novel part aware graph convolution network with temporal enhancement, which can adaptively evaluate the activity level of each part of the body in the action sequence and enhance the extraction of temporal information. Considering that the range of motion of body parts in the action sequence is greater than that of joints, we manually divide the five major parts of the body and generate a skeleton sequence with different attention weights by using part-based attention module. Then, a temporal enhanced module is used to model actions with different duration. Experiments show that our method achieves the state-of-the-art performance.