A Dynamic Topology-Unshared and Occlusion-Robust Network for Skeleton-Based Action Recognition

This paper introduces an advanced hybrid network architecture that integrates Graph Convolutional Networks (GCN) with Transformer models to enhance human action recognition in complex scenarios, particularly in environments with frequent occlusions, such as power stations. We developed a dynamic topology modeling approach with channel-specific unshared weights that adaptively refine the feature matrices and improve recognition accuracy. Our results demonstrate that the proposed hybrid network outperforms traditional models, especially in handling occluded or incomplete skeletal data. We further validated the effectiveness of our model against a range of standard skeleton-based recognition methods, show-casing significant improvements in accuracy. The hybrid model proves particularly adept at maintaining high accuracy even in challenging conditions with substantial occlusions, aligning with the operational realities of power production facilities. This study not only advances the field of action recognition with its robust approach to occlusions but also sets a new benchmark for the practical application of deep learning technologies in industrial settings, ensuring enhanced operational safety and efficiency.