Summary: This paper examined a steering behavior based fatigue monitoring system. The advantages of using steering behavior for detecting fatigue are that these systems measure continuously, cheaply, non-intrusively, and robustly even under extremely demanding environmental conditions. The expected fatigue induced changes in steering behavior are a pattern of slow drifting and fast corrective counter steering. Using advanced signal processing procedures for feature extraction, we computed 3 feature set in the time, frequency and state space domain (a total number of 1251 features) to capture fatigue impaired steering patterns. Each feature set was separately fed into 5 machine learning methods (e.g. Support Vector Machine, K-Nearest Neighbor). The outputs of each single classifier were combined to an ensemble classification value. Finally we combined the ensemble values of 3 feature subsets to a of meta-ensemble classification value. To validate the steering behavior analysis, driving samples are taken from a driving simulator during a sleep deprivation study (N=12). We yielded a recognition rate of 86.1% in classifying slight from strong fatigue.

In an overnight driving simulation study three commercially available devices of fatigue monitoring technologies (FMT) were applied to test their accuracy. 16 volunteers performed driving tasks during eight sessions (40 min each) separated by 15 minutes breaks. The main output variable of FMT devices, which is the percentage of eye closure (PERCLOS), the driving performance (standard deviation of lateral position in lane, SDL), the electroencephalogram (EEG) and electrooculogram (EOG) were recorded during driving. In addition, the subjective self-rated Karolinska sleepiness scale (KSS) was assessed every 2 min. As expected, Pearson product-moment correlation coefficient (PMCC) yielded significant linear dependence between KSS and PERCLOS as well as between SDL and PERCLOS. However, if PMCC was estimated within smaller data segments (3 min) as well as without averaging across subjects then strongly decreased correlation coefficients resulted. To further validate PERCLOS at higher temporal resolution its ability to discriminate between mild and strong fatigue was investigated and compared to the results of the same analysis for EEG/EOG. Spectral-domain features of both types of signals were classified using Support-Vector Machines (SVM). Results suggest that EEG/EOG indicate driver fatigue much better than PERCLOS. Therefore, current FMT devices perform acceptably if temporal resolution is low (> 20 min). But, even under laboratory conditions large errors have to be expected if fatigue is estimated on an individual level and with high temporal resolution.

A combination of linear and nonlinear methods for feature fusion is introduced and the performance of this methodology is illustrated on a real-world problem: the detection of sudden and non-anticipated lapses of attention in car drivers due to drowsiness. To achieve this, signals coming from heterogeneous sources are processed, namely the brain electric activity, variation in the pupil size, and eye and eyelid movements. For all the signals considered, the features are extracted both in the spectral domain and in state space. Linear features are obtained by the modified periodogram, whereas the nonlinear features are based on the recently introduced method of delay vector variance (DVV). The decision process based on such fused features is achieved by support vector machines (SVM) and learning vector quantization (LVQ) neural networks. For the latter also methods of metrics adaptation in the input space are applied. The parameters of all utilized algorithms are optimized empirically in order to gain maximal classification accuracy. It is also shown that metrics adaptation by weighting the input features can improve the classification accuracy, but only to a limited extent. Limited improvements are also obtained when fusing features of selected signals, but highest improvements are gained by fusion of features of all available signals. In this case test errors are reduced down to 9% in the mean, which clearly illustrates the potential of our methodology to establish a reference standard of drowsiness and microsleep detection devices for future online driver monitoring.

Comparing different novel feature sets and classifiers for speech processing based fatigue detectionis is the primary aim of this study. Thus, we conducted a within-subject partial sleep deprivation design (20.00-04.00 h, N = 77 participants) and recorded 372 speech samples of sustained vowel phonation. The self-report on the Karolinska Sleepiness Scale (KSS), and an observer report on the KSS, the KSS Observer Scale were applied to determine sleepiness reference values. Feature extraction methods of non-linear dynamics (NLD) provide additional information regarding the dynamics and structure of sleepiness speech. In all, 395 NLD features and the 170 phonetic features which have been computed partially represent so far unknown auditive-perceptual concepts. Several NLD and phonetic features show significant correlations to KSS ratings, e.g., from the NLD features for male speakers the skewness of vector length within reconstructed phase space (r = .56), and for female speaker the mean of Cao"s minimum embedding dimensions (r =-.39). After a correlation-filter feature subset selection different classification models and ensemble classifiers (by AdaBoost, Bagging) were trained. Bagging procedures turned out to achieve best performance for male and female speakers on the phonetic and the NLD feature set. The best models for the phonetic feature set achieved 78.3% (NaïveBayes) for male and 68.5% (Bagging Bayes Net) for female speaker classification accuracy in detecting sleepiness. The best model for the NLD feature set achieved 77.2% (Bagging Bayes Net) for male and 76.8% (Bagging Bayes Net) for female speakers. Nevertheless, employing the combined phonetic and NLD feature sets provided additional information and thus resulted in an improved highest UA of 79.6% for male (Bayes Net) and 77.1% for female (AdaBoost Nearest Neighbor) speaker.

No abstract