ResGAT-SAGPool Hybrid Architecture: A Spatio-Temporal Graph Feature Learning Framework for Dynamic Gesture Recognition Based on Flexible Sensors

Dynamic gesture recognition serves as a critical foundation for next-generation immersive virtual/augmented reality (VR/AR) systems and human-computer interaction (HCI). To address the challenges posed by high-dimensional heterogeneous spatiotemporal sensor data in dynamic gesture analysis, we propose ResGAT-SAGPool: a hybrid neural architecture integrating Residual Graph Attention Networks with Sparse Adaptive Graph Pooling. This framework establishes an efficient spatiotemporal graph learning paradigm that optimally balances dynamic topology adaptation with computational efficiency through two innovative components: (1) Spatial dimension: A multi-head residual graph attention mechanism enhances spatiotemporal dependency modeling while preserving local gesture sensitivity through skip-connection empowered feature refinement; (2) Temporal dimension: An adaptive hierarchical pooling strategy performs node importance-aware down sampling, achieving 53.3% parameter reduction while maintaining critical temporal patterns. Evaluated on multimodal smart glove datasets (50 word-level gestures and 20 sentence-level patterns), our framework demonstrates superior performance with 92.8% word recognition accuracy (+1.5% improvement) and 91% sentence recognition accuracy (+6.4% enhancement) compared to 1D-CNN baselines. The proposed solution offers twofold advantages for intelligent wearable systems: 58% model compression rate enabling edge deployment, and 92% cross-user generalization accuracy, making it particularly suitable for industrial control systems and IoT-enabled medical rehabilitation applications.