An Attention-Aware Model for Human Action Recognition on Tree-Based Skeleton Sequences

Skeleton-based human action recognition (HAR) has attracted a lot of research attentions because of robustness to variations of locations and appearances. However, most existing methods treat the whole skeleton as a fixed pattern, in which the importance of different skeleton joints for action recognition is not considered. In this paper, a novel CNN-based attention-ware network is proposed. First, to describe the semantic meaning of skeletons and learn the discriminative joints over time, an attention generate network named Global Attention Network (GAN) is proposed to generate attention masks. Then, to encode the spatial structure of skeleton sequences, we design a tree-based traversal (TTTM) rule, which can represent the skeleton structure, as a convolution unit of main network. Finally, the GAN and main network are cascaded as a whole network which is trained in an end-to-end manner. Experiments show that the TTTM and GAN are supplemented each other, and the whole network achieves an efficient improvement over the state-of-the-arts, e.g., the classification accuracy of this network was 83.6% and 89.5% on NTU-RGBD CV and CS dataset, which outperforms any other methods.