Viusal Cues Extraction for Monitoring Driver's Vigilance

Results based on traac accident analysis indicate that a leading cause for automobile accidents is due to a diminished vigilance on the part of the driver. Two major factors that may result in a reduction of a driver's alertness are drowsiness and intoxication. Developing eeective technologies for measuring drowsiness/intoxication is therefore imperative for reducing automobile accidents. The development of technologies for monitoring driver's vigilance is a major challenge in the eld of accident avoidance systems. In order to accurately and robustly characterize a driver's vigilance level, information from diierent sources (e.g., driver's physiological conditions and physical reactions, sensing of vehicle behavior, response of driver, weather conditions, etc.) needs to be collected and processed. This task poses two major challenges: (i) how to extract appropriate information about the driver's condition in a non-intrusive way, and (ii) how to systematically integrate diverse sources of evidence in a uniform manner so that a consistent overall evaluation of the driver' vigilance can be achieved. Towards the rst challenge, we propose to use non-intrusive techniques based on an analysis of driver's facial images. Towards the second challenge , we propose using Bayesian networks which provide a mathematically coherent and sound basis for systematically aggregating visual evidences from diierent sources, augmented with relevant contextual information that may lead to a decline in a driver's vigilance level. This report describes our latest efforts in developing techniques for extracting various visual cues typically reeecting the state of vigilance of the driver.