Convolutional Neural Network for Hand Gesture Recognition using 8 different EMG Signals

Abstract

The following paper presents the implementation of a versatile convolutional neural network architecture (CNN) for the recognition of 6 different commands by means of hand gestures using electromyographic signals. For this, a database consisting of 2880 multi-channel feature maps is built, that is, each dataset is composed of the processed signals of the 8 sensors of a Myo Armband, making use of power spectral density maps. The database is divided into 3 sets of equal size for training, validation and testing. With this, the architecture is trained, obtaining 98.4% accuracy in the validation and 99% in the tests, as well as the verification of the processing time that the network takes to obtain a result, this being 4 ms, demonstrating the ability of a shallow CNN to support multiple channels belonging to different sensors, achieving a high performance and having a reduced execution time that gives the possibility of being implemented in an application in real time.