SleepEYE is a collaborative project between Smart Eye, Volvo Cars and VTI (the Swedish National Road and Transport Research Institute) within the competence centre Virtual Prototyping and Assessment by Simulation (ViP). The main objectives of the project were: * Development and evaluation of a low cost eye tracker unit, and * Determination of indicator thresholds for sleepiness detection. Two literature reviews were conducted in order to identify indicators of driver sleepiness and distraction. Three sleepiness indicators — blink duration, blink frequency and Perclos — were implemented in the camera system. The aims of the study were firstly to develop and evaluate a low cost 1-camera unit for detection of driver impairment, and secondly to identify indicators of driver sleepiness and to create a sleepiness classifier for driving simulators. The project included two experiments. The first was a field test where 18 participants conducted one alert and one sleepy driving session on a motorway. 16 of the 18 participants also participated in the second experiment which was a simulator study similar to the field test. The field test data was used for evaluation of the 1-camera system, with respect to the sleepiness indicators. Blink parameters from the 1-camera system was compared to blink parameters obtained from a reference 3-camera system and from the EOG. It was found that the 1-camera system missed many blinks and that the blink duration was not in agreement with the blink duration obtained from the EOG and from the reference 3-camera system. However, the results also indicated that it should be possible to improve the blink detection algorithm since the raw data looked well in many cases where the algorithm failed to identify blinks. The sleepiness classifier was created using data from the simulator experiment. In the first step, the indicators identified in the literature review were implemented and evaluated. The indicators also included driving and context related parameters in addition to the blink related ones. The most promising indicators were then used as inputs to the classifier. The final set of indicators were an estimated KSS value that was based on the value the driver reported before the driving session (KSSestSR), standard deviation of lateral position (SDLP) and fraction of blinks > 0.15 s (fracBlinks, for EOG based and 1-camera-based). An optimal threshold for discriminating between KSS above and below 8 was determined for each indicator. The performances were in the range of 0.68—0.76. Two decision trees based on the selected indicators were created: one using the fracBlinksEOG and one using fracBlinks1CAM. The performances of the two trees were 0.82 and 0.83 respectively (on the training dataset), i.e., the overall performance of the EOG based and the 1-camera-based classifier were similar, although individual differences could be seen. The performance decreased to 0.66 when using a validation dataset from another study, which illustrates the difficulties in creating a generalized sleepiness classifier. (Author/publisher) This report may be accessed by Internet users at http://www.vti.se/en/publications/camera-based-sleepiness-detection/
Samenvatting