Analysis of EEG signals for classification of car driver's vigilance level

The aim of research presented here is a qualitative analysis of electroencephalographic signals (EEG signals), which reflect the electrical activity a car driver’s brain, in order to indicate his/her level of vigilance. This level is classified into three classes of mentation, wakefulness and somnolence (micro-sleep). The research is conducted in cooperation with the Joint Laboratory of System Reliability of the Czech Technical University in Prague. In the laboratory there is installed a fully equipped car surrounded by 270° projection screens. Actuators of the car are interconnected with a computer that controls various types of projections.

EEG data are collected during experimental sessions involving volunteer and professional car drivers. The drivers undergo different driving scenarios. Scanning cap with 19 electrodes is fixed on their head and EEG signals are scanned and recorded. Signals are classified according to objective (response time on an acoustic signal) and subjective (video analysis of facial grimacing) criterions. Collected data are then processed by mathematical methods to built up classification models to monitor driver’s alertness and helping him to drive safely.

In our team, three methods of EEG data analysis are investigated. All approaches work with the data transformed from time to frequency domain, i.e. with EEG spectrograms.

We have a strong interest in the application of classification trees and random forests methodologies. Promising results were obtained for mixed models where classifiers established for individual persons are combined with the classifiers established for the whole population of tested persons. The combination of classifiers is based on the application of dynamic confidence measures. The other approach employs explorative data-mining methods for indication of conditional dependencies between different levels of vigilance and characteristics of EEG data spectrograms. Found dependencies are used as the basis for building radial fuzzy inference models to process the inherent vagueness accommodated in EEG data.

[1] Coufal D. et al.: Classification Models for Somnolence Detection in EEG Spectra. EURISBIS 09 Book of Abstracts, pp.80-81, Cagliari, Italy

[2] Klaschka J.: EEG Classification Models in Driver's Microsleep Prevention, Proc. of  ISBIS 2010, pp.58., Portorose, Slovenia.

[3] Štefka D., Holeňa M.: Dynamic Classifier Systems and their Applications to Random Forest Ensembles, Springer LNCS 5495, pp. 458-468, ICANNGA 2009, Kuopio, Finnland