Recently, Brain-computer interface (BCI) oriented electroencephalographic (EEG) studies have received due attention for decoding human brain signals corresponding to a specific mental state and providing an alternate solution to the disabled or paralyzed persons for communicating with the computer, robotic arm, or various neural prostheses. In this paper, we propose a two-phase approach to distinguish EEG signals of different mental tasks. The first phase combines the cross-correlation features and slow cortical potentials mean extracted from the most significant channels to form feature vectors. The second phase performs a classification of these feature vectors using SVM and KNN classifiers. It further boosts the classification performance by creating an ensemble of SVM classifiers trained with complementary feature sets extracted during the first phase. EEG signals generated for the same mental task are similar in shape to each other and dissimilar to other activities. The basic principle of cross-correlation is to measure the similarity in shape between two signals which makes it suitable for the EEG analysis. We test the performance of the proposed approach on the BCI competition II dataset Ia representing the cursor movement EEG data for a healthy subject. Experimental results on this dataset demonstrate a significant improvement in the classification accuracy compared to other reported results. Moreover, the proposed work requires fewer channels and features compared to the recent study, which uses all six channels and 42 features, manifesting the efficacy of the proposed work.INDEX TERMS Brain-computer interface (BCI), cross-correlation, EEG classification, ensemble.