A Model for Real-Time Hand Gesture Recognition Using Electromyography (EMG), Covariances and Feed-Forward Artificial Neural Networks

Abstract

Hand gesture recognition has many applications that require models to work in real time and with high recognition accuracy. The problem of hand gesture recognition involves identifying the time, duration and the class of a given movement of the hand. In this paper, a user-specific hand gesture recognition model is proposed. This model is based on electromyographies (EMGs) measured with the Myo Armband, covariances together with a bag of functions for feature extraction, and a shallow feed-forward neural network for classification. This model recognizes 5 gestures: fist, finger spread, wave in, wave out and double tap. The model is trained per user with 25 repetitions for each gesture to recognize. The model was designed, trained and tested using the data of 120 users. The recognition accuracy of this model is 92.45%, with a standard deviation of 11.00%, and an average time of processing of (40.58 ± 1.62)ms, which is less than the permitted delay 300ms for real-time gesture recognition models. Finally, for reproducing the results, the code and the database used for this paper are publicly available online.