Cox proportional hazard models were used to study relationships between the event that a driver is leaving the lane caused by sleepiness and different indicators of sleepiness. In order to elucidate different indicators performance, five different models developed by Cox proportional hazard on a data set from a simulator study were used. The models consisted of physiological indicators and indicators from driving data both as stand alone and in combination. The different models were compared on two different data sets by means of sensitivity and specificity and the models ability to predict lane departure was studied. In conclusion, a combination of blink indicators based on the ratio between blink amplitude and peak closing velocity of eyelid (A/PCV) (or blink amplitude and peak opening velocity of eyelid (A/POV)), standard deviation of lateral position and standard deviation of lateral acceleration relative road (ddy) was the most sensitive approach with sensitivity 0.80. This is also supported by the fact that driving data only shows the impairment of driving performance while blink data have a closer relation to sleepiness. Thus, an effective sleepiness warning system may be based on a combination of lane variability measures and variables related to eye movements (particularly slow eye closure) in order to have both high sensitivity (many correct warnings) and acceptable specificity (few false alarms). (A) Reprinted with permission from Elsevier.
Abstract