ATD-GCN: A human activity recognition approach for human-robot collaboration based on adaptive skeleton tree-decomposition

Abstract

In contemporary intelligent manufacturing environments, scalability and flexibility have emerged as critical requirements for human-robot collaboration (HRC) systems. As the cornerstone of effective HRC implementations, human activity recognition (HAR) is indispensable for accurate understanding of human operational intent and improving HRC in dynamic industrial environments. Recently, graph convolutional networks (GCNs) based on skeleton data have become the most popular methods for HAR. Existing approaches frequently overlook the critical design of topology graphs and skeleton features, resulting in delayed responses and incorrect decisions in HRC systems. To address these challenges, we propose an adaptive human skeleton Tree-Decomposition approach (ATD-GCN). First, the Tree-Decomposition (TD) graph is defined by dividing skeleton nodes into levels with consistent semantic information, where adjacent levels are connected through three hierarchies. Second, we develop a Hierarchical Attention (HA) module to adaptively emphasize the dominant hierarchy within the TD graph. Furthermore, we apply 6DoF theory to the design of skeleton features and develop a multi-stream network integration strategy. Finally, we conduct comprehensive experiments on three datasets and validate the method in the HRC workshop. Experimental results show that our model achieves outstanding performance while remaining lightweight. The proposed ATD-GCN approach makes substantial contributions to HRC by enabling more responsive and adaptive robot behavior through enhanced HAR, thus advancing the development of human-centric smart manufacturing systems.