sEMG-based Gesture Recognition by Rotation Forest-Based Extreme Learning Machine

The motion information contained in surface electromyography (sEMG) signals contributes significantly to the prosthetic hand control. However, the accuracy and speed of gesture recognition from sEMG signals are still insufficient for natural control. In order to alleviate this problem, this paper propose a rotation forest-based extreme learning machine method (RoF-ELM) to improve the recognition performance based on sEMG signals. Firstly, the active motion segments were picked out and pre-processed by sliding window. Then, 104 features were extracted from each sample, and SVM-RFE method was used to reduce the feature space dimension. Finally, the RoF-ELM classification model was constructed and tested based on the EMG data for gestures Data Set containing a total of 72 recordings, each of which consists of six basic gestures. 100 trials were implemented to validate the performance of the proposed method, the results show that the RoF-ELM method have the highest accuracy (91.11%) across different subjects with relatively short runtime compared with decision tree (DT), ELM, random forest (RF), and RoF methods.