Skeleton-Based Action Recognition with Adaptive Connections and Motion Modeling

In the field of skeleton-based action recognition, ST-GCN has gained popularity because of its strong capability to model graph data and its effective use of prior knowledge regarding human joint connections. However, it faces certain challenges and limitations, such as difficulties in transferring information between non-adjacent joints and dealing with subtle variations in action speed and acceleration. To address these issues, we present a Graph Convolution Network based Adaptive connections and Motion modeling (AM-GCN). Firstly, we remove the constraints of skeleton graphs by using learnable adjacency matrices to represent dynamic connections between joints, resulting in information exchange between non-adjacent joints. Additionally, we propose a joint motion modeling technique that combines position, velocity, and acceleration information to provide richer prior knowledge about human movements, further enhancing the model's performance. Extensive experiments conducted on the NTU-RGB+D dataset show that our method performs well in skeleton-based human behavior analysis tasks, demonstrating its potential and value in practical applications.