Brain-computer interface (BCI) is getting increasing attention where classification of motor imagery (MI) using electroencephalography (EEG) signal plays a vital role. In traditional EEG-based BCI setup, after applying pre-processing like band-pass filtering and spatial filtering, features are extracted and are fed to the classifier. However, most of the traditional features are extracted from a single time window, which is usually the full-time frame of a cue-based MI signal. Such features are usually statistical characteristics like log-variance of the whole-time signal. Thus, the information, which is localised in time and crucial for subject-specific MI classification, is not best captured. In this work, a new time-localised approach is proposed where multiple time windows are used for feature extraction. We have developed a number of feature representations using those time windows. Our experimental results corroborate that the proposed approach can achieve higher accuracy of classification when compared to methods using conventional features using the same platform.
