Multi-phase and Multi-mode Gait Recognition with Few Channels and Few Features Based on the Fusion of sEMG and...

To address sensor redundancy and single-dimension (either phase or pattern) in gait recognition, this study proposes a few-channel multi-information fusion framework to achieve concurrent gait pattern-phase recognition, balancing recognition accuracy, system simplicity, and user comfort.

Ten healthy males performed five gait patterns (level walking, stair ascent/descent, ramp ascent/descent). Lower limb surface electromyogram (sEMG), acceleration, and plantar pressure signals were collected. Gait cycles were split into five phases using plantar pressure signals, features were extracted from intra-phase sEMG and acceleration signals, selected via Maximum Relevance Minimum Redundancy (mRMR), and classified by Extreme Gradient Boosting (XGBoost) for gait recognition.

Firstly, a study of five-class classification of gait phases for one single terrain using sensor data from one single muscle location showed that signals at the medial calf (gastrocnemius) or thigh (rectus femoris) had better classification performance with an average accuracy of 91.58%. Secondly, when the five gait phases from five different terrains were combined into a 25-class classification task, the results of fusing the two-channel signals from the calf and the two-channel signals at the thigh were analyzed separately. With fewer channels and features, the classification accuracy reached 88.5%, covering 25 classes.

Considering that lower leg amputations constitute a significant proportion of all major lower limb amputations, we suggest that the fusion of signals from the rectus femoris and semitendinosus positions as the signal sources should be selected for intelligent prosthesis control.