Human Skeleton Graph Attention Convolutional for Video Action Recognition

Abstract

Action recognition based on human skeleton information is a hot topic in the field of computer vision, how to represent the human skeleton graph structure is the key of the method. Graph convolutional network is widely used to extract spatial features of human skeleton. However, the graph convolutional network shares the same weight for neighborhood of each node. In this paper, we propose Human Skeleton Graph Attention Convolutional Neural Network, which introduces graph attention convolution mechanism to extract the spatial features of human skeleton. The model improves the spatial feature extraction of skeleton graph based on the feature relationship of node neighborhood. The experimental results on Kinetics and NTU-RGB+D datasets show that the model can obtain better representation of spatial features, and can achieve better accuracy.