Towards a Reliable Ground Truth for Drowsiness: A Complexity Analysis on the Example of Driver Fatigue

Abstract

The increasing number and complexity of advanced driver assistance systems (ADAS) pave the way for fully automated driving. Automated vehicles are said to increase road safety and prevent human-made (fatal) accidents, amongst others. In the lower levels of automation, however, the driver is still responsible as a fallback authority. As a consequence, systems that reliably monitor the driver's state, especially the risk factor drowsiness, become increasingly essential to ensure the driver's ability to take over control from the vehicle on time. In research, the use of supervised machine learning for drowsiness detection is the prevalent method. As the ground truth for drowsiness is both application- and user-dependent, and no golden standard exists for its definition, measures are usually applied in the form of observer ratings. Also, in this work, observer ratings were investigated with regard to the required level of detail/complexity. To this end, video data, recorded within a simulator study (N = 30) comprised of each 45-minute manual and automated driving sessions, were evaluated by trained raters. Correlation analysis results show that - depending on the number of drowsiness levels - a comparable ground truth can be generated by reducing the rating frequency and thus the rating complexity by a factor of five. The knowledge gained can be used in future studies in this research area, the collection of a reliable and valid ground truth of drowsiness, as well as for improving the process in developing interactive drowsiness detection systems.