Regressing Variations of Joint Angles for Continuously Estimating Hand Finger Movements From sEMG Signals Using Recurrent Neural Networks

Since surface electromyography (sEMG) signals reflect the behavior of a hand prior to the actual activities, they have been used as various interfaces for supporting the functions of conveying the hand behavior. Recently, many systems have been developed to continuously estimate hand finger movements from sEMG signals using neural network models for time-series learning. Most systems have utilized the joint positions and joint angles of fingers as targets in the regression processing. However, using the joint positions and joint angles themselves is not considered reasonable to learn the relationship between sEMG signals and hand finger movements with the models because the muscle activities generate the hand finger movements. Therefore, we propose to use an amount of variation of joint angles as a target in the regression processing for continuous estimation of hand finger movements. Then, we perform an experiment to compare the estimation performance between variations of joint angles and joint angle themselves in the finger movements using the time-series learning models, RNN (Recurrent Neural Network), LSTM (Long-Short Term Memory), and GRU (Gated Recurrent Unit). The experimental results show that all the models can better reproduce the time series of finger movements based on sEMG signals under the condition of variations of joint angles set to the targets in the regression.