Experts state that driver drowsiness is responsible for about 30% of severe traffic accidents. Driver monitoring systems, such as the Mercedes-Benz Attention Assist aim to reduce these road-crashes caused by fatigued drivers using standard equipment sensors. In this article, new measures (features) for detecting drowsiness are proposed in addition to promising features in literature. Most studies in literature are based on driving simulator data, whereas this article focuses on real world driving. External influences such as road condition, road bumps and cross-wind are furthermore taken into account. The presented results are based on a large selection of the Mercedes-Benz drowsiness database which covers over 1.2 million kilometers of measurements. Features are analyzed for their correlation with the subjective Karolinska Sleepiness Scale (KSS). The performance of a combination of features is assessed by sophisticated classifiers and dimension reduction techniques. Even after these improvements, the classification results do not reach the results obtained in a driving simulator. (Author/publisher)
Abstract