Motor Imagery Classification Based on Subject to Subject Transfer in Riemannian Manifold

Motor imagery based brain computer interface requires large number of labeled subject specific training trials to calibrate system for new subjects. This is due to huge variations in individual characteristics. Major challenge in development of brain computer interface is to reduce calibration time or completely eliminate. Existing approaches rise up to this challenge by incorporating Euclidean representation of the individual variations from other subjects’ trials. They use covariance matrices from other subjects but do not consider the geometry of the covariance matrices, which lies in space of Symmetric Positive Definite (SPD) matrices. This inevitably limits their performance. We focus on reducing calibration time by introducing Riemannian approach by incorporating geometrical properties of covariance matrices in the subject to subject transfer. Our method outperforms the state of the art methods on the BCI competition dataset IVa. Our proposed method yielded accuracy of 77.67%, 100%, 75%, 87.05% and 91.67% for five subjects (aa, al, av, aw and ay respectively) in the dataset resulting in an average accuracy of 86.27%.