GLST-Net: Global and local spatio-temporal feature fusion network for skeleton-based action recognition

Abstract

The Graph Convolutional Network (GCN) has been widely applied in skeleton-based action recognition. However, GCNs typically operate based on local graph structures, which limits their ability to recognize and process complex long-range relationships between joints. The proposed GLST-Net consists of three main modules: the Global–Local Dual-Stream Feature Extraction (GLDE) module, the Multi-Scale Temporal Difference Modeling (MTDM) module, and the Temporal Feature Extraction (TFE) module. The GLDE is designed to capture both global and local feature information throughout the motion process and dynamically combine these two types of features. Additionally, since motion is defined by the changes observed between consecutive frames, MTDM extracts inter-frame difference information by calculating differences across multiple time scales, thereby enhancing the model’s temporal modeling capability. Finally, TFE effectively strengthens the model’s ability to extract temporal features. Extensive experiments conducted on the challenging NTU-RGB+D and UAV-Human datasets demonstrate the effectiveness and superiority of the proposed method.