Wrist Motion Regression Using EMG Attention Feature Fusion Algorithm

The visual hand tracking technology in Mixed Reality (MR) headsets is susceptible to environmental occlusion and lighting interference, which significantly degrades interaction accuracy. To address this issue, this study proposes an electromyography (EMG)-based wrist motion regression framework that provides bioelectric compensation in scenarios where visual tracking fails, enhancing the robustness of MR systems. A hybrid TCN-LSTM model is developed, integrated with channel attention and handcrafted feature fusion to enhance wrist angle estimation. To support training and real-time evaluation, we collected a synchronized EMG-MR dataset from 13 subjects at a sampling rate of 500 Hz, segmented using a 0.1 s sliding window with 50% overlap. Experimental results demonstrate that the proposed method achieves lower mean absolute error (MAE) and higher coefficient of determination (R²) than baseline models, reaching an average R² of 0.828 ± 0.029. Furthermore, attention weight analysis reveals consistency with physiological muscle activation patterns, enhancing the interpretability of the model’s decision-making process.