Temporal windowing in CSP method for multi-class Motor Imagery Classification

Brain Computer Interface (BCI) is a system which straightly converts the acquired brain signals such as Electroencephalogram (EEG) to commands for controlling external devices. One of the most successful methods in Motor Imagery based BCI applications is Common Spatial method (CSP). In existing methods based on CSP, the spatial filters are extracted from the whole EEG signal as one time segment. In this study we use the fact that ERD/ERS events are not steady over time. This means that the importance of EEG channels vary for different time segments. Therefore we divide EEG signals into a number of time segments. Then we extract a feature vector from each time segment using CSP. We use OVR (One-Versus-the Rest) algorithm to break four classes problem into two classes problems. The considered four classes MI are left hand, right hand, foot and tongue. We used dataset 2a of BCI competition IV to evaluate our method. The result of experiment shows that this method outperforms both CSP and the best competitor of the BCI competition IV. In fact the effect of noise and outliers on extracted features is reduced by the proposed time windowing method.