Channel Influence in Armband Approach for Gesture Recognition by sEMG Signals

Abstract

This work presents an evaluation of the number of channel using an armband device for classification of hand gestures, aiming to reduce the feature vector dimensionality. Four classifiers and their optimized feature sets were used to recognize 6 gestures. Two search methods were applied to find the best channels: wrapper by Sequential Forward Selection and Binary Particle Swarm Optimization. Moreover, the data were organized in two approaches, analyzing the contribution of each channel and the contribution of each individual feature-channel. Each method resulted in different occurrences for the repeated channels, being the channels placed on flexor carpi ulnaris, palmaris longus, braquioradialis, and extensor digitorum muscles the most repeated. This analysis obtained an average of gain of 2 and a 60% of dimensionality reduction in classification, specially for the Support Vector Machine classifier, reaching 92.3% of hit rate with 9 inputs in Feature-Channel and wrapper approach. The applied methods indicate that there are some dependencies of features and channels in classification process. These dependencies could determine an ideal quantity of channels for a set of gestures and features not only depending on the classifier, but depending of the acquisition channels.