Skeleton-based multi-stream adaptive-attentional sub-graph convolution network for action recognition

Huan Liu, Jian Wu, Haokai Ma, Yuqi Yan, Rui He

Abstract: Recently, graph convolutional networks have achieved remarkable performance with skeleton-based action recognition methods. However, there is potential correlation between different parts of the human body. Many studies have ignored the fact that different actions are the result of the interaction of different human body parts, and that operating on the whole graph provides inadequate information to characterize the action category. In this study, to pay more attention to this problem and further improve the accuracy of action recognition models, sub-graphs based on the depth-first tree traversal order were used to represent the importance and correlation characteristics of joint and bone parts. In addition, beyond the physical structure of the body, joint and bone motion information was also introduced to represent changes in human body parts with movement. To improve the performance of this method, an adaptive-attentional mechanism was added to learn unique topology autonomously for each sample and channel domain. The multi-stream adaptive-attentional sub-graph convolution network was thus proposed for action recognition. The resulting model achieved competitive results on the NTU-RGB + D60 dataset based on 2D or 3D skeleton poses. The experimental results demonstrated the efficacy of our proposed method.

External IDs:dblp:journals/mta/LiuWMYH24