Class-aware topology-guided graph attention for large-scale skeleton-based firefighter action recognition

초록

State-of-the-art graph-based action recognition frameworks, such as Spatio-Temporal Graph Convolutional Networks (ST-GCN) and Two-Stream Adaptive Graph Convolutional Networks (2s-AGCN), have advanced skeleton-based spatio-temporal modeling by learning dynamic joint-wise dependencies across frames. Nevertheless, when applied to the high-variability firefighter behavior analysis—where complex tool-body interactions and subtle intent-driven motions coexist—these models often struggle with fine-grained separability across large action classes. To address this, we propose a Class-Aware Topology-Guided Graph Convolutional Network (CATS-GCN). The novelty of our approach lies in a principled framework that integrates (i) anatomical topology constraints, (ii) class-prototype guided attention, and (iii) multi-head sparsity-diversity regularization into a unified learning objective. We evaluated the proposed framework on a large-scale firefighter action dataset comprising 341 action classes and 13,981 samples. Across five runs with random initializations, CATS-GCN achieves a mean Top-1 accuracy of 93.49%, consistently outperforming all baselines. Owing to its modular structure and interpretable attention patterns, the framework offers practical utility for automated training systems by facilitating correct posture maintenance, reducing injury risk, and enhancing operational readiness. © 2026 Elsevier B.V.

키워드