DAA-GCN: A Dynamic Adaptive Attention Graph Convolutional Network for Robust Skeleton-Based Action Recognition

Skeleton-based action recognition has attracted increasing attention due to its compact data representation and robustness against appearance variations. Although Graph Convolutional Networks (GCNs) have demonstrated strong performance by modeling spatial dependencies among joints, they still face challenges in capturing long-range interactions and multi-scale temporal dynamics. To address these limitations, we propose a novel framework, DAA-GCN (Dynamic Adaptive Attention Graph Convolutional Network), which comprises two core modules: the Spatio-Temporal Adaptive Feature Extractor (STAFE) and the Multi-Perspective Fusion Graph Attention (MPFGA). STAFE integrates long-range spatio-temporal graph convolutions with multi-branch temporal convolutions to effectively capture both short-term and long-term motion patterns, while MPFGA enhances feature representations by combining global self-attention with local additive attention, thereby balancing global context with local structural information. We evaluate the proposed DAA-GCN on two benchmark datasets, NTU RGB+D 60 and NTU RGB+D 120, under both cross-subject and cross-view/setup protocols. Experimental results show that DAA-GCN consistently outperforms state-of-the-art methods, and ablation studies further confirm the effectiveness of each module in the overall architecture. In summary, DAA-GCN presents a robust and scalable solution for skeleton-based action recognition, with promising applications in human-computer interaction, video surveillance, and healthcare monitoring.