Multi-scale Dilated Attention Graph Convolutional Network for Skeleton-Based Action Recognition

Due to the small size, anti-interference and strong robustness of skeletal data, research on human skeleton-based action recognition has become a mainstream. However, due to the incomplete utilization of semantic information and insufficient time modeling, most methods may not be able to fully explore the connections between non-adjacent joints in the spatial or temporal dimensions. Therefore, we propose a Multi-scale Dilated Attention Graph Convolutional Network for Skeleton-Based Action Recognition (MDKA-GCN) to solve the above problems. In the spatial configuration, we explicitly introduce the channel graph composed of high-level semantics (joint type and frame index) of joints into the network to enhance the representation ability of spatiotemporal features. MDKA-GCN uses joint-level, velocity-level and bone-level graphs to more deeply mine the hidden features of human skeletons. In the time configuration, two lightweight multi-scale strategies are proposed, which can be more robust to time changes. Extensive experiments on NTU-RGB+D 60 datasets and NTU-RGB+D 120 datasets show that MDKA-GCN has reached an advanced level, and surpasses the performance of most lightweight SOTA methods.