Intra-subject variability of fatigue estimation based on driving variables

Abstract

Fatigue Monitoring Technologies (FMT) aim at detecting dangerous states of subjects during driving. In a recently published technology report of Caterpillar Machine Research 35 FMT solutions were listed, whereby 22 of them were commercially available in 2006. Most of them utilize measurements of driver behavior such as video recordings of eyelid movements. Alternatively, 5 FMT solutions were based solely on the analysis of driving performance variables such as standard deviation of lateral position in lane, of steering angle and of longitudinal speed. Based on these variables we investigated how accurate fatigue states can be estimated under ideal conditions, i. e. during simulated driving in the lab. Non-linear discriminant analysis utilizing Support-Vector Machines resulted in mean test set errors of 13 % and standard deviations of 2 %. Preliminary results on intra-subject variability were obtained by cross validation analysis of independent data sets of one subject who finished 10 experimental nights in our real-car driving simulation lab. It turned out that intra-subject variability was relatively high; mean test set errors ranged between 9 % and 24 % when the discriminant function was validated on data of one data set against all other data sets. In conclusion, results imply that driving performance variables alone may not have the potential to feed FMT devices, even under ideal conditions. We recommend aiming at data fusion of heterogeneous sources.