Monitoring of drowsiness and microsleep

Summary form only given. An overview is presented of different studies on monitoring and detection of drowsiness and microsleep (MS)during driving simulation. At first a framework is presented how to utilize methods of pattern recognition and of computational intelligence in order to detect and to predict MS. Secondly, different biosignals are compared due to their value for MS detection. Data fusion at the feature level of different electroencephalographic (EEG) and electrooculograhpic (EOG) signals leads to low error rates of classification between MS and Non-MS. Thirdly, results of different classification algorithms are compared and the advantages as well as disadvantages of Support-Vector Machines for this task are highlighted. Fourthly, investigations of automatic relevance determination are introduced. This results in weight variables of the power spectral densities. It is shown that the emphasis on the alpha and theta band of the EEG in detecting MS is a very limited approach. Fifthly, the problem of the inter- and intra-individual variability is addressed. Cross validation analysis revealed that the EEG characteristics during MS are very specific from individual to individual. Recently, other authors also reported this problem. Sixthly, one has to keep in mind that the MS detection algorithm has been established by learning from data. For this, only clear-cut examples of MS and Non-MS were selected. They cover approximately 15% of the whole time of driving. The remaining 85 % of time comprise Non-MS as well as dubious behavioral states. Therefore, the question arises how well the MS detection algorithms works when applied consecutively. It turned out that there is a high contrast in the detector output signal. It will be argued why validation analysis is difficult in principle. Seventhly, the MS density is presented as a new measure of drowsiness and strong central fatigue. High correlation coefficients to subjectively self-ratings of sleepiness as well as- to objectively assessed driving performance measures confirm this measure. Furthermore, strong relations of MS density to the occurrence of incidents and accident were observed. Eighthly, the application of our methodology to evaluate commercial fatigue monitoring technologies (FMT) is presented. Results suggest that under laboratory conditions current FMT devices are reliable when temporal resolution is not too fine and data averaged across several subjects were utilized. But FMT fail in giving a valid drowsiness measure at high temporal resolution and to predict subjective sleepiness as well as driving performance on an individual level.