Mix-hops Graph Convolutional Networks for Skeleton-Based Action Recognition

Abstract

Skeleton-based human action recognition has drawn considerable research interest since it can robustly accommodate dynamic circumstances and complex backgrounds. By modeling the human body skeletons as graph structure, graph convolution network (GCN) has achieved great success in this field. However, these methods based on GCN are difficult to capture global features and relations only through a single-layer network. To obtain the relations between distant skeleton joints of human body, it is necessary to stack multiple graph convolution layers. In addition, the topology of the graph needs to be set manually in graph convolution operation and it is shared in all layers and time dimensions. In this paper, a novel mix-hops graph convolutional network (MHGCN) is proposed to recognize human action from skeleton data. The proposed module can fuse local features with global features through a layer of graph convolutional network. Besides, the topological structure of graph in our model changes with the time dimension and it can be individually learned in an end-to-end way through the BP algorithm. The experiments on several benchmark datasets show remarkable performance for human action recognition, demonstrating the effectiveness of our method.