Deep Residual Split Directed Graph Convolutional Neural Networks for Action Recognition

Abstract

Human action recognition is the basis technology of human behavior understanding, and it is a research hotspot in the field of computer vision. Recently, some studies show skeleton data (i.e., joint points and edges) is naturally more conducive to mining the connotation of human action, so exploring the relationship between joints and bones is helpful to improve action recognition. In this article, regarding skeleton data as a directed graph, we design directed graph convolutional neural networks with a novel residual split structure. First, we construct a directed graph represent model to extract human behavior by two kinds of graph models. Second, we use a novel residual split block to construct graph convolution neural network. Different from the traditional residual networks, we split high-dimensional features into several shallow features with the same dimension. It can not only ensure the diversity of mining features, but also avoid the gradient disappearing. Finally, during training, we use random sampling of data to reduce the burden of network training. Experiment results show that the proposed method achieves higher recognition rate than the comparative methods on the NTU-RGBD dataset.