MEL-FREQUENCY BAND STRUCTURE BASED FEATURES FOR MOTOR IMAGERY TASK CLASSIFICATION

Differentially enabled communities face much difficulties and challenges in their life time while commuting from one place to another. Power wheelchairs were designed to aid the movement of these differentially enabled subjects and a Brain Computer Interface can also be applied to replace the existing conventional joystick method of controlling the movement of a wheelchair without using hands. In this research work, a simple protocol is proposed to record the EEG signals emanated from a subject while the subject performed four different kinesthetic motor imagery tasks. The noise present in the EEG signals are removed and three different feature sets, namely, power spectral density, Mel-frequency cepstral coefficients and Mel-frequency band structure based energy features are extracted. The extracted features are then associated to the type of motor imagery tasks and three multi-layer Perceptrons trained with Levenberg-Marquardt method are developed. The performance of the three Perceptron models are evaluated in term of classification rate and compared. From the results, it is observed that the Perceptron model trained with Mel-frequency band structure based features has yielded a higher classification accuracy for all 5 subjects, which is between 92.64-97.72%. The obtained result clearly indicates that the Mel-frequency band structure based features has potential to classify the four different motor imagery tasks.