Abstract
Vigilance or sustained attention is an important aspect for people who engaged in long time attention demanding tasks such as monotonous monitoring and driving. Vigilance detection has been an important topic in the field of brain-computer interface (BCI) research. However, the study is limited due to the low SNR (Signal-Noise Ratio) nature of EEG. Common spatial pattern (CSP) is a one of the most effective algorithms for feature extraction method in the BCI study area. The CSP seeks for an optimal projection direction (spatial filter) by maximizing the variance of one class and simultaneously minimizing the variance of the other class. There is one drawbacks exists in the traditional CSP, that is, the CSP is proposed relies on the assumption that data in each class follow the Gaussian distribution. However, this assumption is not always true for EEG data in practice, especially in the research of vigilance detection based EEG (e.g. during sleep). Thus, traditional CSP suffers performance degradation in case of non-Gaussian distributions. In this paper, we extend the traditional CSP to the general version and proposed nonparametric CSP (NCSP) algorithms which do not explicitly rely on the assumption of the underlying class Gaussian distribution and we then develop a new efficient algorithm based on matrix deflation to solve the proposed NCSP algorithm and its extensions-nonparametric multi-class CSP (NMCSP). Experimental results on EEG-based vigilance estimation and motor imagery recognition task demonstrate the effectiveness and efficiency of our proposed algorithms.