Finger Joint Angle Estimation based on sEMG signals and deep learning method

Conventional movement recognition methods are normally based on classification algorithms, which could only provide discrete movement classification rather than natural human body continuous movements. In this paper, we utilized the deep learning methods to estimate eight complicated movements of fingers by extracting the kinematic information based on surface electromyographic (sEMG) signals. Aiming at realizing continuous estimation, we adopted four representative models, AlexNet, Residual neural network (ResNet), Long Short-term Memory network (LSTM) and Gate Recurrent Unit (GRU) in this study. Convolutional kind models (AlexNet and ResNet) are chosen because of their irreplaceable feature extraction ability. And recurrent kind models (LSTM and GRU) are chosen because they are suitable for time-series signal processing. We took 10 degrees of freedom (DoFs) of joint angles from one hand as the target, 12 channels of sEMG as input and trained the models with the stochastic gradient descent and backpropagation. The models were tested on 8 abled subjects. The results indicated that the employed AlexNet turned out to show the best estimation performance and stability than other models. We realized the AlexNet is more suitable for sEMG based continuous movement estimation.