Discriminating sensors for driver's impairment detection

In this work, the authors present a statistical analysis applied to different sensors on the driver's impairment detection problem. Their goal is to get a minimal number of discriminating sensors to match the requirements of an industrial prototype. The signals coming from these group of sensors are used to create artificial variables based on several mathematical transformations (wavelets, standard deviations) in order to fuse this information at a first level. The authors' statistical study is based on several steps: 1) the observation of the variation of the mean and variance of the different variables, 2) the test of Hypothesis F concerning a population's variance and then, an hypotheses test based on Student's t distribution for the means, 3) principal components analysis (PCA); and 4) general performance of the diagnosis system using or not the different variables. This study has been performed using experimental data coming from 10 drivers in real experiments involving fatigued drivers at a closed circuit and over a motorway. These experiments have been realised using the CopiTech demonstrator which is equipped with a group of sensors measuring physiologic, mechanical and environmental status in real time. The analysis is validated by the physical state of the driver based on EEG. Results stress that the better discriminating sensors are: a) lateral position, b) steering wheel angle, and c) vehicle speed.