Dynamic-static partitioning mask and multi-dimensional attention mechanism for skeleton-based action recognition

For skeleton-based action recognition, existing GCN-based methods utilize the adjacency matrix to model the complex dynamic relationships in human joint topology. However, most of these methods ignore the heterogeneity between dynamic and static regions. In this paper, we propose a gradient-based dynamic-static partitioning mask strategy, which dynamically generates masks to distinguish between dynamic and static regions. This strategy enables a more precise capture of key action components while retaining the advantages of GCNs in handling the non-Euclidean topology of skeleton data. Additionally, we introduce the Spatio-Temporal Channel Attention Mechanism (STCA), which plays a crucial role in enhancing the effectiveness of the partitioning mask strategy by adaptively adjusting the focus on temporal, spatial, and channel dimensions of the input data. Experimental results on two large-scale datasets, NTU RGB+D and NTU RGB+D 120, demonstrate that the proposed method is highly competitive compared to the state-of-the-art methods.