Multiclass motor imagery classification based on the correlation of pattern images generated by spatial filters

This paper proposes a new motor imagery classification method. This method applies the Common Spatial Pattern (CSP) technique that projects Electroencephalography (EEG) filtered signals into a different time-space, where an optimal variance for discrimination of different motor imagery tasks is obtained. By using the training data set, RGB pattern images of each mental task are created through the features obtained by the CSP technique. Then the normalized cross-correlation coefficients of the RGB pattern images with the images representing the mental state of the user in a segment of two seconds are computed. The mean and variance values of the correlation coefficients and the obtained CSP features are used to train six binary Support Vector Machine classifiers, which discriminate between imagined left hand, right hand and foot movements and an idle state (a state without any imagery movement). The results show an average accuracy of 83.18% with the training data set and an 81.58% with the testing data set. These results demonstrate that the proposed method is competitive compared with existing methods and may represents a successful alternative for a multi-motor imagery classification.