Multi-modality dense graph convolution network for skeleton-based action recognition

Abstract

Skeleton-based human action recognition has emerged as a vibrant area of research in recent years.. In action recognition based on skeleton, the method based on graph convolution network is through modeling space to explore physical dependence between body joints, has obtained the remarkable performance, however, most GCN method uses the stratification to aggregate a wide range of field information, which makes the joint characteristics in long distance transmission is weakened. In this paper, dense graph convolutional networks are proposed to overcome these shortcomings. To enhance the informative skeleton features and create a compact representation at an early fusion stage, an MMF is developed. Then, Our proposed method utilizes a dense graph convolution operation to improve the local context information of each joint. Finally, feature channels and action classes are specific, In our model, it is crucial to assign weights to important features and focus on the relevant information. Therefore, To improve the identification of features in our model, we introduce a channel joint attention module. A large number of experiments on three large datasets (NTU-RGB+D, NTU-RGB+D120 and Kinetics) have demonstrated the performance of the proposed MMD-GCN for skeleton-based action recognition.