Hand Gesture Recognition Using Instant High-density EMG Graph via Deep Learning Method
Dezhen Xiong, Daohui Zhang, Xingang Zhao, Yiwen Zhao
Abstract
Electromyography (EMG) shows excellent potential for human-machine interaction (HMI) tasks. It reflects the physiological intention of human beings, which contributes to a more intuitive human-machine interface. The sequence of EMG signals acquiring during a period is most commonly used for feature extraction and gesture recognition. The instant EMG graph, which is always thought to be useless due to too much noise inside it, can also be used to recognize movement intention. This work explores a new path to recognize EMG patterns without feature engineering, using both deep learning and machine learning methods. This paper proposes a novel scheme to classify hand gestures through the instant graph of high-density electromyography (HD-EMG) using the deep learning method. Four types of recurrent neural networks (RNNs) with units including long short-term memory (LSTM), gated recurrent unit (GRU), and bidirectional version of them are chosen to extract temporal information automatically from EMG data. By a simple 100 frame majority voting, which corresponds to a 100ms window, the best performance of 98.57% is achieved by Bi-LSTM. Besides, the machine learning-based method also achieves an accuracy of 95.66%, which shows the instant EMG graph method's great potential.