An approach to classification of three different imaginary movements based on linear discriminant analysis transformations and applicable to brain-computer interface implementations is considered. First, search for discriminative frequencies individual for each subject and each movement is conducted. It is shown that this procedure leads to an increase in classification accuracy compared to conventional common spatial patterns algorithm followed by linear classifier considered as a baseline approach. In addition, an original approach to finding discriminative time segments for each movement is tested. This approach led to further increase in accuracy if Hjorth parameters and inter-channel correlation coefficients were used as features calculated for the found segments. Particularly, classification by the latter feature led to the best accuracy of 69,4% averaged over all subjects. Besides, scatter plots demonstrated that two out of three movements pairs were discriminated by the approach presented.