Detecting Driver Sleepiness Using Optimized Nonlinear Combinations of Sleepiness Indicators

This paper addresses the problem of detecting sleepiness in car drivers. First, a variety of sleepiness indicators (based on driving behavior) proposed in the literature were evaluated. These indicators were then subjected to parametric optimization using stochastic optimization methods. To improve performance, the functional form of some of the indicators was generalized before optimization. Next, using a neural network, the best performing sleepiness indicators were combined with a mathematical model of sleepiness, i.e., the sleep/wake predictor (SWP). The analyses were based on data obtained from a study that involved 12 test subjects at the moving-base driving simulator at the Swedish National Road and Transportation Research Institute (VTI), Linköping, Sweden. The data were derived from 12 1-h driving sessions for each test subject, with varying degrees of sleepiness. The performance measure (range [0,1]) for indicators was taken as the average of sensitivity and specificity. Starting with indicators proposed in the literature, the best such indicator, i.e., the standard deviation of the yaw angle, reached a performance score of 0.72 on previously unseen test data. It was found that indicators based on a given signal gave essentially equal performance after parametric optimization, but in no case was it better than 0.72. The best generalized indicator (the generic variability indicator) obtained a performance score of 0.74. SWP achieved a score of 0.78. However, by nonlinearly combining SWP with the generic variability indicator, a score of 0.83 was obtained. Thus, the results imply that a nonlinear combination of a measure based on driving behavior with a model of sleepiness significantly improves driver sleepiness detection.