Deep learning approach for sign language's handshapes recognition from EMG signals

Abstract

Electromyographic (EMG) signals are beginning to be increasingly used as alternatives for gesture recognition. Indeed, EMG signals are used to control machines (drones, robots, VR, PowerPoint presentation…). In the field of assistive technologies, EMG signals offer new solutions for people with disabilities. Namely, the control of prostheses for disabled people or the control of wheelchairs or computers for paralyzed people. Through this work we aim to use electromyographic signals for the recognition of sign languages alphabets. The signals are received by 8 sensors mounted on an armband that the deaf person can wear on the forearm. We propose a new approach for the classification of EMG signals using an approach based on new deep learning solutions. We used a sequence of CNN layers followed by two LSTM layers to extract features from Electromyographic data. We implemented and evaluated the proposed approach using 28 handshapes representing Arabic alphabets. We obtained an average accuracy of 98,49%.