Accurate and Robust Locomotion Mode Recognition Using High-Density EMG Recordings from a Single Muscle Group.

Existing methods for human locomotion mode recognition often rely on using multiple bipolar electrode sensors on multiple muscle groups to accurately identify underlying motor activities. To avoid this complex setup and facilitate the translation of this technology, we introduce a single grid of high-density surface electromyography (HDsEMG) electrodes mounted on a single location (above the rectus femoris) to classify six locomotion modes in human walking. By employing a neural network, the trained model achieved average recognition accuracy of 97.7% with 160ms latency, significantly better than the model trained with one bipolar electrode pair placed on the same muscle (71.4% accuracy). To further exploit the spatial and temporal information of HDsEMG, we applied data augmentation to generate artificial data from simulated displaced electrodes, aiming to counteract the influence of electrode shifts. By employing a convolutional neural network with the enhanced dataset, the updated model was not strongly affected by electrode misplacement (93.9% accuracy) while models trained by bipolar electrode data were significantly disrupted by electrode shifts (29.4% accuracy). Findings suggest HDsEMG could be a valuable resource for mapping gait with fewer sensor locations and greater robustness. Results offer future promise for real-time control of assistive technology such as exoskeletons.