A Multi-Modal Approach to Gesture Recognition: Exploring Semg Statistical and Eegchaotic Features

Abstract

Gesture recognition makes possible a more intuitive and natural interaction between humans and machines. A key aspect is collecting data of representative human activity; many approaches work with sEMG and EEG as they provide information about muscle activity and brain activity. However, these signals present significant variability because of their complex nature and non-linear dynamics. The aim of this work is to recognize three human arm gestures from only sEMG and EEG collected by two commercial wearable devices: the Mindrove armband of eight channels and the Muse 2 of four electrodes. The three gestures were recorded from 10 participants performing three different activities, each involving different gestures. The sEMG were processed from a classical perspective of data cleaning and extraction of statistical features, but four chaotic descriptors were also extracted to compare which approach obtained the best classification results. The EEG were processed to extract five frequency bands (delta, theta, alpha, beta, and gamma), from which chaotic descriptors were also computed. We used three machine learning models: Bagged Trees, FineTree, and RUSBoosted Trees, where the Bagged Trees model had the highest accuracy among the three. We achieved classification accuracies of more than 97.49%. For EEG, the maximum recognition accuracy was 99.68% using chaotic descriptors extracted from the beta and gamma frequency bands. These results show the potential of the features here used to build a robust gesture recognition framework to later integrate a collaborative robot that responds to the user's gestures.

Keywords: EEG, sEMG, Chaotic Descriptors, Biomedical Signal Processing, Gesture Recognition.