Towards Robot-Assisted Post-Stroke Hand Rehabilitation: Fugl-Meyer Gesture Recognition Using sEMG

Abstract

Robot-assisted rehabilitation training requires to identify the patient's motion intention effectively. These motions are usually originated from rehabilitation actions included in the Fugl-Meyer assessment scale. Surface electromyography (sEMG) is the most commonly used physiological signal for identifying the motion intention of patients. The use of sEMG to classify different gesture patterns is one key technology for the human-machine interaction. Therefore, this paper investigates a Fugl-Meyer hand gesture recognition method towards robot-assisted hand rehabilitation. The experiment data set including eight hand gesture information is collected from six volunteers. Six single features (Difference Absolute Mean Value (DAMV), Integral of Absolute Value (IAV), Variance (VAR), Autoregressive Coefficients (AR), maximum value of Discrete Wave Transformation (DWTmax) and standard deviation of Discrete Wavelet Transform (DWTstd)) are used to recognize the gesture. The experimental results demonstrate that: (1) a segment length of 250 ms contains enough information to estimate the hand gestures and leaves sufficient time to do feature extraction and gesture recognition; (2) by comparing the performance of different single features, DWTstd wins the highest accuracy (i.e., 96%); (3) the combination of single features into a multi-feature can effectively improve the recognition accuracy, where the best performance is achieved by multi-feature combining DAMV, IAV and AR under BP neural network classifier (the average accuracy is 97.71%); (4) as to different classifiers, BP neural network has a better performance than Support Vector Machine (SVM) and Extreme Learning Machine (ELM).