Driver drowsiness estimation by fusion of lane and eye features using a multilevel evidence theory

Driver drowsiness has been considered as a significant contributing factor to severe traffic accidents. Most of studies about monitoring driver drowsiness have investigated simple functions of performance, such as deviation of lane position and percentage of eyelid closure. However, any single function cannot be verified to work well in the complex road conditions. Thus, in this paper, a nonintrusive surveillance system is proposed to estimate driver drowsiness through fusion of visual information on lane and driver in a multilevel framework with evidence theory. Based on expert knowledge and data statistics, various visual features extracted from lane and eye tracking are analyzed for their correlation with the subjective Observer Rating of Drowsiness (ORD) scale. Various feature sets are then combined individually using a non-distinct combination rule at a low level. Fusion of the distinct results from the low level is processed at a higher level, where it could determine the driver´s state. The system has been validated in real world and the experiment results show its efficiency in real-time surveillance.