A Topology-Aware Graph Learning Framework for Occlusion-Robust Skeleton-Based Human Action Recognition

Skeleton-based human action recognition is attractive for practical vision systems because it represents motion compactly, but occlusion and pose-estimation failures can remove spatially adjacent joints for extended periods and degrade graph-based recognition. We propose a Topology-Aware Graph Learning (TAGL) framework that couples three mechanisms: Neighborhood Expansion Masking generates spatially correlated, sequence-level joint loss during training; Adaptive Multi-Hop Graph Learning reweights information from different graph distances when local neighborhoods become unreliable; and Holistic Topological Mapping provides complementary global structural cues through persistent topology. TAGL achieves 93.3%/97.9% top-1 accuracy on NTU RGB+D 60, 90.4%/91.6% on NTU RGB+D 120, 97.0% on Northwestern-UCLA, and 47.2% on UAV-Human. Across six heterogeneous skeleton corruptions, its average accuracy reaches 85.8%, compared with 79.3% without NEM and 84.5% with random masking, while requiring 1.78 M parameters and 2.58 GFLOPs. These results show that TAGL improves tolerance to incomplete and structurally degraded skeleton observations without excessive computational overhead. Its compact and corruption-aware design is therefore promising for skeleton-based monitoring systems operating under imperfect pose observations, including safety surveillance, human–robot interaction, and rehabilitation-oriented motion analysis.