sEMG-based Hand Posture Recognition and Visual-Feedback Training

Abstract

Purpose sEMG-based recognition system provides the intuitive and accurate classification of various gestures and useful for human-computer interaction. However, the variability of the sEMG signal was caused by the individual differences and the change of the force or proficiency. In this study, the hand posture recognition algorithm was developed using sEMG sensors, and the visual-feedback training was performed to improve the classification performance.

Method Four healthy adults and a forearm amputee participated in this study to perform twelve hand postures (Rest, Hook, Spread, Scissor-sign, Finger pointing, V-sign, O.K.-sign, Cylindrical grasp, Spherical grasp, Lateral pinch, Palmar pinch, and Tip pinch). Commercial sEMG sensors (Delsys Trigno; 1926 Hz) were placed on the nine forearm muscles of the healthy adults (FDS, ED, EDM, EP, FCR, FCU, ECR, ECU, BRD) and the eight forearm muscles of an amputee (based on 3D reconstruction data from MRI; BRD, FCR, ECR, ED, ECU, FDP, FDS, FCU) (Figure 1). A LabVIEW-based radar plot was used for hand posture training (1 day for the healthy adults and 5 days for the amputee) to minimize the variability of the sEMG signal. In the session for recording the sEMG signal, all participants repeated the hand postures 10 times with 20% MVC. The 10-fold cross validation was performed on the ANN classifiers with the feature vector of MAV or the Hudgins’ set (MAV, WL, ZC, and SSC) on MATLAB.

Results and Discussion The classification accuracy was improved by increasing the number of training data in both trained and untrained sessions (Table 1). ANN classifiers of the healthy adults with MAV showed the classification accuracies of 84.3% and 85.8% in untrained session and trained session respectively. The classification performance was improved as 88.4%(untrained) and 91.6%(trained) by the Hudgins’ set. In untrained session of the forearm amputee, poor classification accuracies appeared in both MAV(31.7%) and the Hudgins’ set(33.0%). After the 5 days training, MAV and the Hudgins’ set showed the classification accuracy of 76.3% and 84.7% respectively. In this study, the Hudgins’ set and the posture training were useful to improve the classification accuracy. Furthermore, the great improvement of the classification performance appeared in the forearm amputee with an appropriate posture training (day 5). These results supported that the posture training with visual feedback had an effect to minimize the variability of sEMG signals.