Self-Supervised Topology-Aware Skeleton Masking for Robust Human Action Recognition

Self-supervised representation learning has become critical for addressing the annotation bottleneck in skeleton-based human action recognition on consumer electronics devices. However, most existing methods adopt a spatio-temporal decoupled training paradigm, which neglects the inherent synergy between a skeleton’s spatial topology and its temporal dynamics. Even recent hierarchical approaches fail to capture cross-level associations fully, limiting overall performance. To address these challenges, we propose a Topology-aware Spatio-Temporal Interaction Contrastive Network (TSI-CNet) within a self-supervised framework. Topology-driven spatial masks quantify joint centrality in the skeleton graph to prioritize highly connected nodes. Motion-driven temporal masks dynamically select high-change frames based on inter-frame motion magnitude. A dual-branch decoupling–interaction architecture further disentangles spatial and temporal features and achieves semantic fusion through a cross-domain alignment loss. Extensive experiments on five public datasets show that TSI-CNet achieves competitive performance with a lightweight design. The results confirm the effectiveness of the proposed topology- and motion-driven masking strategies, and ablation studies further validate the contributions of individual components. Robustness evaluations under seven simulated real-world perturbations also highlight the method’s resilience to spatial and temporal disturbances, underscoring its suitability for consumer applications such as smart cameras, AR/VR devices, and service robots.