A Measure of Strong Driver Fatigue

Strong fatigue during sustained operations is difficult to quantify because of its complex nature and large inter-individual differences. The most evident and unambiguous sign is the occurrence of microsleep (MS) events. We aimed at detecting MS utilizing computational intelligence methods. Our analysis was based on biosignal and video recordings of 10 healthy young adults who completed 14 sessions over two nights in our real-car driving simulation lab. Visual scoring by trained raters led to 2,290 examples of MS. Only evident events accompanied by prolonged eyelid closures, roving eye movements, head noddings, major driving incidents, and drift-out-of-lane accidents were regarded as MS. All other cases with signs of fatigue were regarded as dubious. The same amount of counterexamples (Non-MS) where continued driving was still possible were picked out from the recordings. Non-MS and MS examples covered only 15% of the whole time. Support-Vector Machines were utilized as classifiers and were adapted to these two classes of examples. If such classifiers were applied consecutively, then 100% of time is covered. Validation analysis demonstrated that the classifier gained high selectivity and high specificity. Based on this complete coverage, the percentage of MS in a predefined time span can be calculated. This measure was highly correlated to deteriorations in driving performance and to subjective self-ratings of sleepiness. We conclude that reliable detection of MS is possible despite large intra- and inter-individual differences in behaviour and in biosignal characteristics. Therefore, the percentage of detected MS gives an objective measure of strong driver fatigue.