BSCAL: Gesture Recognition Network Based on Dual-Stream Information Fusion of EMG and IMU Signals

Surface electromyography (sEMG) enables gesture recognition for rehabilitation and human-computer interaction but faces challenges from noise, inter-subject variability, and limited motion dynamics characterisation.

We propose BSCAL, a dual-stream network integrating sEMG and inertial measurement unit (IMU) signals. The architecture combines convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and spatiotemporal attention to process eight-channel sEMG and six-channel IMU data from 20 participants performing six gestures. Kalman filtering, Z-score normalisation, and gated fusion optimise multimodal feature integration.

BSCAL achieved 90.41% ± 1.54% recognition accuracy, surpassing single-modality models and state-of-the-art approaches. Ablation studies validated contributions from CNN (local features), LSTM (temporal dependencies), and attention (key feature enhancement).

By synergistically integrating neuromuscular patterns from EMG and kinematic data from IMU, BSCAL delivers a precise and scalable solution for gesture recognition, thereby advancing the development of wearable hand rehabilitation robots and assistive technologies for motor function recovery.