A Lightweight Dynamic Gesture Recognition Approach with Graph Convolution and Spatio-Temporal Decoupling
Year: 2026
Type: conference-paper
Abstract: In recent years, Graph Convolutional Networks (GCN) have become mainstream methods in dynamic hand gesture recognition due to their strengths in modeling the spatial structure of skeletal joints. However, existing models still suffer from structural redundancy and insufficient decoupling of spatio-temporal information, which limits their discriminative capability and deployment efficiency. In this paper, we propose a lightweight dynamic graph modeling approach that integrates GCNs with an improved temporal processing module, constructing a hybrid neural network architecture featuring temporal decoupling and feature re-aggregation mechanisms. Specifically, we first introduce a spatial graph convolution module that combines semantic difference modeling and structure-aware mechanisms to enhance the topological semantic relationships among joints. Then, we develop an improved temporal processing module to independently model the temporal dynamics of joints. Finally, a global average pooling and channel mapping mechanism is designed to effectively compress feature dimensions while improving the model's discriminative power and computational efficiency. Experiments conducted on two public hand gesture datasets (Shrec17 and DHG14/28) demonstrate that the proposed method achieves leading accuracy while maintaining a parameter size of only 0.14M, which is significantly lower than that of mainstream methods, striking a good balance between lightweight design and performance. This study provides a new paradigm for spatio-temporal structure decoupling in dynamic graph modeling and shows strong potential for applications on resource-constrained devices.
Institution: Jilin Animation Institute
Language: English
Subfield: Human-Computer Interaction
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204806429