Implementação de uma metodologia para classificação de imagética motora em módulo Orange Pi Zero

Abstract

Various people around the world live with some form of physical disability. For this part of the population that has some motor deficiency, can communicate or perform tasks that are impossible to perform is of utmost importance the development of technologies such as Brain Computer Interface (BCI). Because a BCI application works by converting the brain signals into commands to control external resources, such as orthotics, wheelchair, among other controllable devices. Being able to contribute significantly to the improvement of the living condition of these people. The main goal of the current study is to implement in a portable platform the techniques already tested in a computer, referring to a system of detection of movement intention based on EEG signals, implemented by (SILVA, 2017). This current study presents initial experiments for the identification of motor imagery, referring to the detection of intention of movement of the hands or detection of intention of movement referring to hand or foot. Aiming to contribute to the increase of reliability in the detection of intentions to the point of allowing its integration in a BCI system with practical application. The results are presented referring to the training stage of the algorithm and simulation in real time, where the trained algorithm is executed in the portable platform. In the methodology proposed by (SILVA, 2017), the signal is filtered in cerebral rhythms related to the tasks of motor imagery. After filtering, statistical attributes of these brain rhythms are extracted and fed to a classifier (used Random Forest and Support Vector Machine) that performs the prediction on the set of tests. In this current study, instead of the classifier being applied on the test set in a single time, the prediction of each attempt is performed one by one, in the real-time simulation step. The tests were performed with unbalanced and balanced classes. Were also performed tests with right hand and left hand detection instead of using the of hand and foot classes, as suggested by reviewers of the paper presented by (SILVA, 2017), in order to verify the performance of the proposed methodology.