Global-Local Collaborative Graph Convolutional Network for Skeleton-Based Action Recognition

Skeleton-based action recognition has emerged as a critical challenge in computer vision. In this task, local and global features describe actions from the perspectives of fine-grained details and overall patterns. However, existing methods lack sufficient exploration of the two types of features. How to adaptively coordinate global dynamics with local detailed features and fully explore their complementarity remains a critical issue to be addressed. To solve this problem, we propose a Global-Local Collaborative Graph Convolutional Network(GLC-GCN), which achieves deep interaction and optimized collaboration between global and local features through their flexible representation, thereby effectively enhancing the recognition performance for diverse actions. In the spatial dimension, we design a flexible dual-graph mechanism: one graph adaptively captures global spatial semantics by learning dynamic joint relationships, while the other enhances the representation of local motion changes by emphasizing salient joints. These two graphs are integrated through a learnable fusion strategy that balances their contributions based on feature consistency, ensuring coordinated coarse and fine-grained information modeling. In the temporal dimension, we adopt a CNN-Transformer dual-branch framework to capture short-term transitions and long-term temporal dependencies jointly. Furthermore, a Salient Location Selection is introduced to highlight informative temporal segments and suppress ambiguous frames. GLC-GCN has been extensively experimented on three action recognition datasets: NTU RGB+D, NTU RGB+D 120, and NW-UCLA, with results demonstrating that it is competitive with state-of-the-art methods.