Estimation of Ankle Angle Based on Multi-Feature Fusion with Random Forest

The electromyography (EMG) has an important application in continuous movement estimation of the human lower limb during walking, which can be potentially implemented in to the fields of the intelligent prosthesis, medical rehabilitation, and man-machine control. Considering the evaluation requirements of the accuracy and speed, in this paper we present an algorithm combined with feature fusion and random forest (RF) to estimate the ankle joint angle. Firstly, surface electromyographic signals (sEMG) were experimentally collected from the related muscles in the lower extremity. Secondly, three domain features (mean value, variance and waveform length) were extracted from the sEMG which were denoised and preprocessed. Thirdly, the features were taken into the RF input layer and estimated the joint angles of the ankle. Finally, accuracy and speed indexes were verified and compared between the RF and neural network algorithms by using the mean square root error and the correlation coefficient. The results show that the training time of the RF algorithm is far less than the BP neural network while ensuring the accuracy of the estimation, which has a valuable reference for further research on gait analysis and the prosthesis control.