DescriptionThe literature assigns human behavioral factors as the main cause in traffic accidents. In up to 95%, a contributory cause is indicated. Here, driver fatigue has a non-negligible influence. In Austria, the share of fatigue as the main causative factor amounts to 3.4%, but this is difficult to prove in accident analysis and may be much higher. Worldwide research estimates that fatigue is involved in about one third of fatal accidents.
State-of-the-art drowsiness detection systems usually rely on the evaluation of the steering signal, with a classification reliability of about 60%. This can be increased to about 80% by fusing other data such as time of day and lane keeping. Alternatively, the eyelid movement can be detected and evaluated with interior surveillance cameras, but here the robustness of the signal is often limited by changing light conditions and driver position. However, reliable and robust prediction of driver fatigue a few minutes before micro sleep is currently not available. In the near future, SAE Level 3 systems will be available on the market that will allow hands-off from the steering wheel, but this will render the in-vehicle signals worthless. Reliable recognition of the driver's state will be necessary in order to include the driver's influence in take-over situations switching between automated and manual driving in critical situations. This will have a significant impact on the market introduction of automated driving functions.
The WACHsens research project featured a driving simulator study with 92 test subjects, balanced according to age and gender. It included driving on a three-lane highway at night in a well-rested and tired state, each in manual and automated driving mode. Through elaborate measurement of physical state (EEG, ECG, skin resistance, eyelid movement, eye gaze, head position, respiration rate), data is generated that is used to robustly, accurately, and predictively classify driver fatigue using methods from machine and deep learning. The ground truth required for this purpose to train the artificial neural networks is determined by classifying vigilance into 4 classes through video observation.
Results of the study presented here show that using traditional machine learning methods improves the accuracy of drowsiness detection in manual driving to about 83% and in automated driving to about 76% by fusing vehicle, facial and heart rate signals, while other bio-signals such as EEG, skin resistance and respiration rate do not provide any improvement or are either too complex in data analysis or in the subsequent vehicle application. Using deep learning methods, this classification can be increase up to 86% in manual and 84% in automated mode, while being even more robust when one of the signals is lost. The classification in three different levels of drowsiness allows to detect oncoming drowsiness early enough to initiate countermeasures. In future we are seeking for cooperation for in-vehicle application
|Period||18 Jan 2022|
|Event title||Automotive Innovation Workshop: AUIN Workshop 2022|
|Degree of Recognition||International|
ASJC Scopus subject areas
- Automotive Engineering
Fields of Expertise
- Mobility & Production
Documents & Links
Project: Research area