Raw sEMG-based Real-time Gesture Recognition with Recurrent Neural Networks

Abstract

Surface Electromyography (sEMG) is widely applied in controlling assistant devices such as prostheses and wheelchairs due to it is convenient to access and use. However, real-time sEMG gesture recognition must deal with low recognition accuracy and time delay, limited by selected features and recognition algorithm. In this study we utilize the recurrent neural networks (RNNs) to recognize raw sEMG signals real-time without the feature extraction methods to improve the performance of gestures recognition. The proposed methodology was evaluated on three subjects with a set of six gestures. The performance of Long Short-Term Memory network (LSTM) and Gated Recurrent Unit (GRU) are compared and discussed. The results show that by using GRU, the average accuracy is 97.32% with the time delay of 80 ms, and the results for LSTM are 96.17% and 160 ms. This indicate that that GRU achieves higher accuracy than LSTM in our dataset and has shorter response time. In the future, we will apply the proposed methodology to actual assistant devices.