Several studies show that up to one in four severe traffic accidents can be attributed to drowsiness. Drivers often over-estimate their fitness level or are not aware of the danger that always accompanies drowsy driving. Since associations like the NHTSA pointed to the relevance of this topic, more and more research has been conducted and in the meantime there is also a variety of commercial systems on the market to address this risk. In this paper, we do not aim to find new methods of detecting drowsiness of a driver. Our approach is rather to choose an established method and enhance it in a way that it not only performs well in a driving simulator but also in real world drives.
The chosen drowsiness detection method is the observation of the steering wheel angle signal. It has been shown that the frequency of occurrence of a typical steering pattern, which can roughly be described as a deadband followed by a rather fast correction, is an indicator for the state of drowsiness of a driver. The advantage over other techniques like camera-based detection is that it can run in standard equipped cars. Thus it is available for the largest number of drivers and can thereby achieve the greatest effect on accident avoidance.
We investigate the chosen detection method in real world drives and discuss which other effects not related to drowsiness can evoke the described steering pattern. We focus on environmental effects like crosswind and can show that those events may lead to an increase of the amount of steering patterns. Finally, we quantify the influence on drowsiness measures. The underlying database comprises more than two million kilometers of more than one thousand drivers, all real-world drives.
Our evaluation shows that particularly on routes or in situations where those environmental influences accumulate, the drowsiness measure can be affected to an extent that leads to false triggering of the system. Therefore, we suggest measures that can be taken to reduce the influence of steering patterns that are not related to the driver’s drowsiness state.
The aim of most drowsiness detection systems is to inform a driver when his state has reached a critical level and to motivate him to take appropriate measures. This presupposes confidence in the system. False warnings will negatively affect the credibility of the system.
Our purpose is to show the importance of enabling this kind of system to recognize external influences, thus making detection more robust. We consider it very important to make such systems as reliable and credible as possible, as otherwise the driver will not take the advice the system will give him. Limiting the influence of external factors is a key to achieving this goal.