Abstract
Currently, research on gesture recognition systems has been on the rise due to the capabilities these systems provide to the field of human-machine interaction, however, gesture recognition in prosthesis and orthesis has been carried out through the use of an extensive amount of channels and electrodes to acquire the EMG (Electromyography) signals, increasing the cost and complexity of these systems. The scientific literature shows different approaches related to gesture recognition based on the analysis of EMG signals using deep learning models, highlighting the recurrent neural networks with deep learning structures. This paper presents the implementation of a Recurrent Neural Network (RNN) model using Long-short Term Memory (LSTM) units and dense layers to develop a gesture classifier for hand prosthesis control, aiming to decrease the number of EMG channels and the overall model complexity, in order to increase its scalability for embedded systems. The proposed model requires the use of only four EMG channels to recognize five hand gestures, greatly reducing the number of electrodes compared to other approaches found in the literature. The proposed model was trained using a dataset for each gesture EMG signals, which were recorded for 20 s using a custom EMG armband. The model reached an accuracy of to 99% for the training and validation stages, and an accuracy of 87 +/- 7% during real-time testing. The results obtained by the proposed model establish a general methodology for the reduction of complexity in the recognition of gestures intended for human.machine interaction for different computational devices.