A Survey of Recent Graph-Based Methods for Skeleton-Based Action Recognition

Skeleton-based action recognition (SBAR) leverages 3D joint trajectories to recognize human activities while offering privacy, robustness to illumination/background changes, and computational efficiency. Recent progress is dominated by spatio-temporal graph neural networks (ST-GNNs) that model the human body as a graph and learn data-adaptive connectivity, hierarchical structure, compact representations, multimodal supervision, and efficient temporal fusion. This survey focuses on five representative methods CTR-GCN, HD-GCN, InfoGCN, Language Supervised Training (LST), and Temporal Channel Aggregation (TCA-GCN) and positions them within the broader SBAR literature. We analyze modeling assumptions, architectural choices, training objectives, and empirical results on NTU RGB+D 60/120 and Northwestern-UCLA. We additionally contextualize the trajectory from dynamic topology learning to emerging foundation and sequence models reported in 2024-2025. Finally, we summarize open challenges and provide research directions for scalable, robust, and semantically grounded SBAR.