Desarrollo de un modelo de reconocimiento de gestos de la mano utilizando señales EMG y Deep Learning.

Abstract

Hand gesture recognition (HGR) is one of the fields of research that has successfully developed humanmachine interaction applications in the last years. In this work, we propose the creation of a HGR model based on a Convolutional Neural Network (CNN). Then, we adapt a Long Short-Term Memory (LSTM) layer to the architecture of the model to observe its effect on the classification accuracy, recognition accuracy and processing time. The input of the model are the spectrograms created using Surface electromyography (sEMG) on the forearm through the commercial sensor Myo Armband. For testing, we performed experiments using a public EMGs dataset of 612 users, and we measured and compared the classification and recognition accuracy between 5 different gestures and the no gesture. The results were evaluated for the proposed model (CNN-based model) and its adaptation to use the LSTM layer (CNNLSTM- based model). The results showed that the classification accuracy reaches up to 90.49% ± 9.70% and the recognition up to 86.83% ± 11.30% for the CNN-based model, and classification accuracy up to 92.93% ± 8.23% and recognition up to 91.60% ± 8.81 % for the CNN-LSTM-based model.