Multiple tangent space projection for motor imagery EEG classification

Due to its non-invasiveness and easiness to implement, EEG signals decoding are in base of most based brain computer interfaces (BCI) studies. Given the non-stationary nature of these signals, a preprocessing phase is needed. An interesting idea to perform the preprocessing is the use of spatial covariance matrices. In the last years, spatial covariance matrices based preprocessing was extensively used in electroencephalography (EEG) signal processing and spatial filtering for Motor imagery (MI) BCI. Spatial covariance matrices lie in the Riemannian manifold of Symmetric Positive-Definite (SPD) matrices, therefore, the use of Riemannian geometry is attracting a lot of attention and showing to be simple, robust, and providing good performance. This paper explores the idea of enhancing the information provided to the classifier by the combination of different covariance matrices projections from their native Riemannian space to multiple class-depending tangent spaces. We demonstrate that this new approach provides a significant improvement in model accuracy.