A Real-Time High-Density sEMG Gesture Recognition System Distilled from a Deep Model

Surface electromyography (sEMG) has been widely used in human-machine interaction systems for gesture recognition. While high-density EMG systems can enhance recognition accuracy and robustness, their practical deployment remains challenging due to increased hardware complexity and computational demands. In this study, we propose a real-time, high-density EMG-based gesture recognition system that utilizes a 64-channel electrode array and a knowledge distillation-based deep learning approach to enable efficient, wearable deployment. Our system processes sEMG signals in real-time using a lightweight student model, distilled from a VGG-16 teacher model, to balance recognition performance and computational efficiency. The model is further optimized for online streaming scenarios, employing a sliding window mechanism, cold time filtering, and robust calibration strategies to enhance real-world usability. We evaluated the system with 12 participants, achieving an average recognition accuracy of 88.12% across 11 hand gestures while maintaining a processing time of 9.7 ms per segment, meeting real-time requirements. These results demonstrate the feasibility of deploying high-density sEMG-based gesture recognition systems in practical, real-time applications.