DETECTION AND PREDICTION OF DRIVER’S MICROSLEEP EVENTS Martin Golz, David Sommer, Markus Holzbrecher, Thomas Schnupp University of Applied Sciences Schmalkalden D - 98 573 SCHMALKALDEN Germany Phone: +49 3683 688 4107 Fax: +49 3683 688 4499 E-mail:
[email protected]
ABSTRACT The detection of spontaneous behavioral events like short episodes of unintentional sleep onset during driving, which are usually called microsleep events, still poses a challenge. The analysis of only a small number of signals seems to be useful to detect such events on a second-by-second basis. Here we present an experimental investigation of 22 young drivers in our real car driving simulation lab. The experimental design was chosen to raise many microsleep events. A framework for adaptive signal processing and subsequent discriminant analysis was applied. In addition to the common estimation of Power Spectral Densities, the recently introduced method of Delay Vector Variance is utilized in order to get an estimate if the signal has undergone a modality change or not during the microsleep event under analysis. The fusion of the outcomes of both methods applied to three different types of signals, to the Electroencephalogram, the Electrooculogram and to Eyetracking signals, by modern methods of Computational Intelligence, namely the Support Vector Machine, leads to high classification accuracies with mean errors down to 9% for all subjects. It turned out that such low errors are only achievable in a relatively small temporal window around the onset of microsleep. Their prediction is feasible but with much higher errors. The signal processing framework has the potential to establish a reference standard for drowsiness and microsleep detection.
1 INTRODUCTION The detection of short-time brain states from ongoing biosignals is a challenging task not only in the area of clinical applications but also for e.g. future human-machine-interaction. As a special type of such an interface one can consider a system for detection of short intrusions of sleep into sustained wakefulness. In case of automobile drivers such events are believed to be a major factor in accident causation. During the recent years this topic has received broad attention from authorities, from the public and as well as from the research community. Most research projects in this area, e.g. the EU projects AWAKE (2001–2004) and SENSATION (2004–2007), are engaged in developing sensors to monitor driving impairments due to fatigue and drowsiness. These impairments arise on a time scale of some ten seconds and are typically developing as waxing and waning patterns. Some doubts still exist about the feasibility of detecting short sleep intrusions under demands of attentiveness in ongoing biosignals on a time scale of, say, one to five seconds [Sagberg et al. 2004]. Many biosignals which are more or less coupled to drowsiness do not fulfill these temporal requirements. For example, electrodermal activity and galvanic skin resistance are too slow in their dynamics to detect such suddenly occurring events. The EEG is a relatively fast and direct functional reflection of mainly cortical and to some low degree also of subcortical activities. Therefore, it should be the most promising signal for microsleep detection. The electrooculogram (EOG) is a measurement of mainly eye and eyelid movements. Their endogenous
components are coupled to the autonomic nervous system which is affected during drowsiness and wake-sleep transitions. Disadvantageously, the electrophysiological measurements of brain electric and of eye movement activity are non-contactless and are corrupted by large noise which is originated by other simultaneously ongoing processes. This leads to more or less extensive signal processing and pattern recognition. Another signal type is the contactless working eyetracking which is also featured by high temporal resolution. It comes out with time series of the pupil diameter and of the eyegaze location. We suppose that there should be characteristic short-time-stationary patterns in all three signal sources, perhaps reflecting brain microstates associated to microsleep. Using machine learning algorithms it should be possible to detect these patterns whereby it is a priori not clear how stable and how affected by disturbances they are. The main concern of this contribution is to find out which signal is most valuable for MSE detection and prediction. Furthermore, it will be investigated if combinations of these signals can lead to better results or not.
2 EXPERIMENTS 2.1 Participants Subjects were recruited by roundmail to all students of our university. A donation of 50 Euros for participation was announced. Students got free access to a registration web page where they got further information about the procedures and their aims. In case of interest, they had to give in initial personal data, e.g. information on general sleep-wake rhythm and health status. From all students fulfilling the requirements to participate in the study (ca. 86%) twentysix healthy subjects were selected randomly (21 male, 5 female; mean age 24.4 ± 3.1 years, range 19-28 years). In addition, the Pittsburgh Sleep Quality Index (PSQI) [Buysse et al. 1989] was administered. The mean PSQI score was 3.8 ± 1.6. No subject reported PSQI larger than 5.0 and no one reported chronic or current major medical illness or injury, medication or drug consumption, shift work or transmeridian travel within the last three months prior to the study. During the week preceding the study subjects had to keep a sleep diary to assess sleep habits. They were instructed not to take daytime naps during that time, i.e. to go to sleep only once a day and to refrain from excessive physical activity, caffeine, and alcohol consumption. Finally, they were told not to consume alcoholic or caffeine beverages during the day before the experiment.
2.2 Subject preparations Three days before the experimental night subjects were familiarized with the lab equipment and had to drive on a 20 min training course in the driving simulator. Two female subjects complained about simulator sickness and were excluded from further investigations. During the experimental nights one further subject has quitted because of simulator sickness and one because of back pain. Therefore, twenty-two subjects finished experiments completely. All subjects gave written informed consent and gave a written declaration on their transfer home after experiments. Only driving as passenger or, in case of campus residents, walking was allowed. In addition, subjects had to carry a wrist actometer during the three days and nights preceding the experiments. Actograms were checked immediately after arrival of the subject to the experimental night, normally at 11 pm. Primarily, we checked total sleep length (6 … 10 hrs), time-since-sleep (14 … 16 hrs) and if the subject accomplished the demand of no nap.
2.3 Equipment Experiments were conducted in our driving simulation lab consisting of an operator room and a fully dark, temperature controlled simulator room (Fig. 1). Subjects had to drive a real small 2
city car (GM Opel “Corsa”) on a slightly winding main road under conditions of night vision. No oncoming traffic is simulated in order to maintain a high level of monotony. The driving scene is projected on a projection plane 2.6 m in front of the subject; maximal visual angle is 56 deg. In case of complete road departures a force feedback to the steering wheel is generated which is in nearly all cases effective enough to awaken drowsy subjects.
Figure 1: Real car driving simulation lab which is specialized for recording of overnightdriving simulations. A real small city car in conjunction with an interactive 3D driving simulation software is utilized to present a monotonic lane tracking task to the subjects. Subject behavior is recorded by infrared video cameras. In addition, driving performance and biosignals of the subject are recorded. Measured variables of the driving simulator are lane deviation, velocity, steering angle, and pedal movements; sampling rate is 10 sec-1. Furthermore, electropolygraphy is derived. Seven signals of EEG (C3, Cz, C4, O1, O2, A1, A2, common average reference), two of EOG (vertical, horizontal), one of ECG, and one of EMG (m. submentalis) were sampled at a rate of 128 sec-1. Further six signals were recorded by an eye tracking system (ETS, binocular) at a rate of 250 sec-1. For each eye the pupil size and the two coordinates of eye gaze on the plane of projection are measured. In addition, three video camera streams are recorded: (i) of subjects left eye region, (ii) of her/his head and of upper part of the body, and (iii) of driving scene. Video recordings are used for online and offline scoring as explained later.
2.4 Experimental Design Driving started at 1:00 am after a day of normal activity and a time since sleep of at least 16 hours. In all, subjects had to complete seven driving sessions lasting 35 min, each preceded and followed by vigilance tests and responding to sleepiness questionnaires. The vigilance tests consisted of two visuomotoric tasks and of one electrophysiological test; reports on their methodology and of their results will be given in another paper. Before the next driving session a 10 min long break was inserted for subjects needs and for motivational conversation. Experiments ended at 8:00 am (Figure 2). On the one hand, our design has the disadvantage of non-continuous driving due to questionnaires, vigilance tests and breaks. But on the other hand a large total time on duty is gained and time of day effect due to passing through the circadian trough can be observed. We experienced earlier that it is hard to motivate a subject for continuous driving in a 3
simulator for longer than two or three hours; most of them are willing to give up when the first microsleep episodes (MSE) arise. We believe that our design results in much more examples of MSE than in continuous driving. This way we have found 3,573 MSE (per subject: mean number 162 ± 91, range 11 - 399).
Figure 2: Chronology of one experimental night from 11 pm to 8 am. Each subject had to complete seven driving sessions, three vigilance tasks (VT 1 - 3), and had to respond to the Visual Analogue Scale (VAS) and Thayer’s Activation-Deactivation Adjective Checklist (ADACL).
2.5 Scoring of microsleep events Driving tasks were chosen intentionally monotonous and with time-since-sleep up to 24 hours to support drowsiness and occurrence of MSE. The latter are defined as short intrusions of sleep into wakefulness under demands of attention. They were detected online by two operators who observed the subject utilizing three video camera streams (section 2.3). Typical signs of MSE are e.g. prolonged eyelid closures, roving eye movements, head noddings, major driving incidents and drift-out-of-lane accidents. This step of online scoring is critical, because there are no unique signs of MSE, and their exact beginning is sometimes hardly to define. Therefore, all events were checked offline by an independent expert and were corrected if necessary. Unclear MSE characterized by e.g. short phases with extremely small eyelid gap, inertia of eyelid opening or slow head down movements were excluded from further analysis. Non-MSEs were selected at all times outside of clear and of unclear MSE. We have picked out the same amount of non-MSE as of MSE in order to have a balanced data set. Our intention was to design a detection system for clear MSE versus clear Non-MSE classification, assuming that such a system can not only detect the MSE recognized by human experts, but would also offer a possibility to detect unclear MSE cases which are not easily recognizable by experts. Most of the recorded signals are useful for fatigue evaluation by a temporal resolution of say 3 or 5 minutes. We believe that for MSE detection the EEG, eye movements (EOG, ETS) and the pupil diameter may be useful, but other variables, like e.g. EMG, ECG, lane deviation, or steering angle are not sensitive enough for detections on a second-by-second basis needed for MSE detection. Results of fatigue evaluation will be published elsewhere. In the following, data analysis of EEG, EOG, and ETS is presented.
3 DATA ANALYSIS Processing of the above mentioned biosignals in order to detect spontaneous events like MSE is primarily a task of discriminant analysis which typically comprises of pre-processing, feature extraction, classification, and validation. In pre-processing mainly three steps have to be performed: signal segmentation, artefact removal and missing data substitution. Segmentation of all signals was done with respect to the observed temporal starting points of MSE / Non-MSE using two free parameters, the
4
segment length and the temporal offset between first sample point of the segment and starting point of the event. The first parameter adjusts the trade-off between temporal and spectral resolution whereas the second parameter controls the location of the region-of-interest on the time axis. Therefore, both parameters are of high importance and have to be optimized (section 4). Artifacts in the EEG are signal components which are presumably originated extracerebrally and often exhibit as transient, high-amplitude voltages. For their detection a sliding double data window was applied, in order to compare the power spectral densities in both windows. When the mean squared difference of them is higher than a thoroughly defined threshold value, then the condition of stationarity should be evidently violated and as a consequence this example of MSE or NMSE is excluded from further analysis. In all, 14 MSE and 223 NMSE were excluded; the latter were all exchanged by new examples drawn from the original data set. Missing data problem occurred in all six eyetracking signals during every eyelid closures. This is caused by the measuring principle. They are substituted by data generated by an autoregressive model which is fitted to the signal immediately before the eyelid closure. This way, artificial data replace missing data under the assumption of stationarity. Nevertheless, this problem should be not important enough to give more insight. For instance, periods of missing data are in the size of 150 msec which is small compared to the segment length of 8 sec (section 4). In the stage of feature extraction two completely different methods were applied. First, we utilized the common periodogram as a direct method to estimate logarithmic power spectral densities (PSD) [cf. Percival & Walden 1994]. This method assumes that the signal is stationnary and their generating system is linear. PSD values were afterwards summed in spectral bands. As it was shown recently, this step of feature reduction has potential to optimize detection performance [Sommer & Golz 2007]. It was found empirically that for quantitative EEG analysis the summation of PSD values in equidistant band is much more optimal than the common summation in the delta, theta, alpha and beta band. The lower and upper cut-off frequencies have been found to be 0.5 Hz and 23.0 Hz, respectively, and the width of the spectral bands has been found to be 1.0 Hz [Sommer & Golz 2007]. The second applied method was the recently introduced method of Delay Vector Variance (DVV) [Gautama et al. 2004a]. DVV transforms the signal to the state space using time delay embedding [cf. Kantz & Schreiber 2004]. This has the advantage that signals which show a high degree of irregularity in the time domain are mapped on relatively simple trajectories in the state space if their generating system can be described by coupled, ordinary differential equations. Simple statistical tests, like the unpaired t-test, in the state space can then be utilized to estimate to which degree the signal may be generated by a nonlinear system and to estimate how large may be the amount of stochasticity in the signal. Both features are important and are dependent on one free parameter which controls the degree of similarity in the sate space. Therefore, two feature sets are generated by DVV. They may vary over time if the signal generating process alters as it might by when a MSE is oncoming. In this line, DVV may be better suited to detect signal alterations than PSD estimation. Two modern Soft Computing methods were utilized for the stage of classification, namely Optimized Learning Vector Quantization (OLVQ1) [Kohonen 2001] and Support Vector Machines (SVM) [Cortes & Vapnik 1995]. Both are stochastic learning methods and have the ability to adapt a discriminant function without any presumptions on the data distribution. In order to gain good adaptivity and also high generalizability several internal parameters have to be optimized which is much more computational time consuming than in basic statistical testing procedures. Further inside into this topic can be found elsewhere [Golz et al. 2001, Sommer et al. 2005, Golz et al. 2007].
5
The last stage of discriminant analysis comprises validation in order to estimate the true error of classification. The expectation value of the classification error based on the training data set has been shown to be biased [cf. Joachims 2002]. This error is called training set error and is a useful measure to check how good the adaptation of the discriminant function has been working. Several cross validation methods have been developed in order to get a second measure, the test set error. One cross validation method, the so-called “leave-one-out” scheme, is an almost unbiased estimator of the true classification error, but is computationally much more expensive than e.g. the “multiple-hold-out” scheme. For the latter case it has been shown numerically to perform also well in case of two practical biosignals applications [Sommer & Golz 2006]. Therefore, we will use the “multiple-hold-out” scheme when OLVQ1 is applied and will use the “leave-one-out” scheme when SVM is applied because in case of SVM a computational efficient implementation exists [Joachims 2002].
mse vs. non-mse5 mse vs. non-mse4 mse vs. non-mse3 mse vs. non-mse2 mse vs. non-mse1
45 40
Test Error [%]
35 30 25 20 15 10 5 0
-9
-8
-7
-6
-5
-4
-3
-2
-1
0
1
2
3
4
Offset [s] Figure 3: Mean test set errors versus temporal offset of the biosignals segments. For all five different data sets compared the optimal offset value is around -3 sec.
4 RESULTS 4.1 Ability of detection and prediction of microsleep Our data set consists of a total of 3,559 clear-cut MSE and of the same amount of NonMSE. As mentioned above, Non-MSE has been picked out at all times outside of clear and of unclear MSE. Five different types of Non-MSE were selected to show their influence on the detection accuracy: - Non-MSE1: only episodes of the first driving session (from 1:00 am to 1:35 am), - Non-MSE2: episodes of the first driving session and only during eyelid closures, - Non-MSE3: episodes in the first five minutes of each driving session, - Non-MSE4: only episodes between MSE where subject is drowsy,
6
- Non-MSE5: like Non-MSE4, but only during eyelid closures. Variation of the segment offset as a free parameter has led to a relatively steep error function (Fig. 3). An optimal offset value was found to be around -3 sec. In the same way an optimal segment length of 8 sec was found. Both optimal parameter settings mean that classification is working best when biosignals from 3 sec immediately before MSE to 5 sec after MSE onset are analyzed. Classification of MSE versus Non-MSE1 resulted best because it is easiest to discriminate between MSE, which are always ongoing under a high level of fatigue, and Non-MSE of the first driving session, which are at a relatively low level of fatigue. The biosignals of both classes must have characteristic differences and therefore a relatively good discrimination is possible with mean errors down to 5 %. Classification of MSE versus Non-MSE3 is always at higher errors because a lot of Non-MSE segments are like MSE segments of higher levels of fatigue. That’s why it is more complicate to find a good discrimination function. When segments of Non-MSE5 are processed then this task is much more difficult because segments of both classes, MSE and Non-MSE5, are of the same highest level of fatigue. Therefore the minimum of the error function is by 10% higher than for the easiest case when segments of Non-MSE1 are processed. One could argue that mostly MSE are starting at eyelid closures and, therefore, we did perhaps nothing else than a simple detection of eyelid closures. But this was clearly not the case, because eyelid closures of MSE versus eyelid closures of Non-MSE (type 4) were discriminated with nearly equal test errors than Non-MSE without eyelid-closures (type 5). Only the first mentioned case, MSE against Non-MSE of the first session, was slightly more difficult to discriminate if both comprise eyelid closures (type 2). In the following, all results presented were obtained from the most difficult types of Non-MSE (Non-MSE4 and NonMSE5), because this is of highest interest for sensor applications.
4.2 Best signals Next, we pursued the question if one type of measurement (EEG, EOG, ETS) contains enough discriminatory information and which single signal inside of one type is most successful. Our empirical results suggest that the vertical EOG signal is very important (Fig. 4) leading to the assumption that modifications in eye and eyelid movements have high importance, which is in accordance to results of other authors [Galley et al. 1999]. In contrast to the results of EOG, processing of ETS signals has led to lower errors for the horizontal than for the vertical component. This can be explained by the reduced amount of information in ETS signals compared to EOG. Rooted in the measurement principle, the ETS measures eyegaze movements and pupil size changes, but cannot acquire signals during eye closures and cannot represent information of eyelid movements. Both aspects seem to have a large importance for the detection task, because errors were lower in EOG than in ETS. It turns out that also the pupil diameter (D) is an important signal for microsleep detection. The performance of ETS signals for microsleep detection was in the same shape as of EEG signals. This is noticeable because ETS suffers from the problem of missing data. Compared to the EOG, the EEG signals performed inferior, among them the Cz signal performed best. Relatively low errors were also found for other central (C3, C4) and for occipital (O1, O2) electrode locations, whereas both mastoid electrodes (A1, A2), which are considered as least electrically active sites, show lowest classification accuracies (highest errors), as expected. Similarities in performance between symmetrically located electrodes (A1-A2, C3-C4, O1O2) meets also expectancy and supports reliance on the chosen way of signal analysis. DVV features alone turned out to lead to relatively high classification errors (Fig. 4) despite additional numerical effort in optimizing the internal DVV parameters. This is surprising because DVV was successfully applied to sleep EEG [Gautama et al. 2004b]. Processing EEG
7
during microsleep and drowsy states and, moreover, processing of shorter segments seems to be another issue. The performance of PSD was much better. The fusion of DVV and PSD features (DVV+PSD) gained slight improvements, especially when SVM with Gaussian kernel function was utilized to find a discriminant function. OLVQ1 was always outperformed by SVM (Fig. 4), but only if Gaussian kernel functions were utilized and if a regularization parameter and a kernel parameter has been optimized which takes considerable computational costs.
Figure 4: Mean and standard deviation of test set errors for different single signals. Two different feature sets (DVV, PSD), their fusion (DVV+PSD) and two different classification methods (OLVQ1, SVM) are compared.
4.3 Data fusion A pronounced improvement of the classification accuracies was achieved by feature fusion of all three signal sources (Fig. 5). Compared to the best single channel of each signal type (three leftmost groups of bars in fig. 5), the feature fusion of vertical EOG and central EEG gained a more accurate classification, and was also more successful than the fusion of features of both EOG (EOG all) or of all seven EEG signals (EEG all). Feature fusion of nine signals (EOG + EEG all) and feature fusion of all fifteen signals (all ETS + EOG + EEG) resulted in slightly higher accuracies when OLVQ1 was applied for discriminant analysis. But, classification errors were considerably lowered if SVM was utilized. For the latter mentioned case best results were achieved; the fusion of features of both types (PSD + DVV) and of all 7 EEG, of both EOG, and of all 6 ETS signals utilizing SVM resulted in test errors lower than 10%. A comparison of more classification methods and a report of some more details on discriminant analysis, their parameters and their computational costs can be found elsewhere [Golz et al. 2007]. All different types of signal sources, namely brain activity reflected by the EEG, eye and eyelid movements reflected by the EOG as well as by the ETS and pupil size changes reflected by the ETS are meaningful for fusion of their features in order to get an optimal detection of MSE.
8
Figure 5: Mean and standard deviation of test set errors for three single signals and several examples of feature fusion, e.g. fusion of all EEG channels (EEG all), of all six ETS signals (ETS all) and of both EOG channels (EOG all). The rightmost group of bars shows the best performing fusion of all signals available (all ETS + EOG + EEG). As in Fig. 4 two different feature sets, their fusion and two different classification methods are compared.
5 CONCLUSIONS A way to assess many examples of clear-cut microsleep events has been presented. Especially young subjects suffer from unintentional and very abrupt ongoing MSE. If they are requested for a monotonous lane tracking task then in most subjects a large amount of MSE are occurring overnight. In addition, the cumulative time-on-task and the time-since-sleep were chosen relatively long. Subject were selected such that all had to drive in the simulator at a time-ofday when they are normally at sleep. All factors have contributed to get a relatively large data set. This is an advantage of simulator studies compared to studies on the real roads. Nevertheless there are a lot of disadvantages. The subjects are for example aware that they do not experience the real risk of accidents in a simulator. As a consequence the operators have to look after if the subject is driving correctly between consecutive MSE, because Non-MSE recorded in between have to be assigned to fatigue but active driving. The visual scoring of the behavioural events under interest is a critical point. There are numerous states of the driver where it is not clear to say if she / he performs still sufficiently. The video recordings as well as all biosignals recordings contain no unique signs of MSE and of Non-MSE. Nevertheless, rating of the video recordings is an excellent tool [Lal & Craig 2002]. Please note that the operators are tracking all recordings (video, biosignals, vehicle variables) the whole time online, are getting informed when the car drifts out-of-lane and stay in contact with the subject when an accident has occurred or warnings have to be given in case of bad driving. Therefore, it is not so much difficult to give a score if a clear MSE, a clear Non-MSE, or an unclear episode actually happened. In addition we have checked all
9
events by two independent offline scorers. They seldomly found other scores, but had to correct the temporal starting point of MSE due to delays of the online scorers. The correct definition of the starting point is an important issue, because it turned out that the detection of MSE is feasible in a very short temporal window. It is optimal when the biosignals segment under analysis starts 3 sec before the onset of MSE and ends 5 sec afterwards. A shift of, say, ± 4 sec already leads to much higher detection errors. The case of MSE prediction, i.e. the temporal offset has to be lower equal -8 sec, leads to errors of about 34 % which is not acceptable. Please notice that also a shift in the reverse direction shows the same decline in detection; this suggests that the biosignals are not remarkably changed during the seconds after MSE onset. Results support the hypothesis that MSE are spontaneous events which the subject is not aware of some seconds before and after their onset. On the other hand, is has to be considered that recently several authors [Ingre et al. 2006] found that several biosignals during strong fatigue show high inter- and intra-individual variability which give request to more experimental investigations to clarify this issue. Following this, we conclude that only in a small temporal window the large variability of the biosignals is diminished such that a relatively precise detection is possible for all subjects. Future research should also be concerned about the large inter- and intra-individual variability in the characteristics of all types of biosignals which we have observed also in our previous studies. To date, the required amount of microsleep examples is not available to conduct such data analysis. Our methodological framework for adaptive signal processing has turned out to perform well also in more complex cases, e.g. when many feature sets of different types of biosignal sources has to be fused. All signal sources had importance when seeking for an optimal solution, but no one was predominant. Even the contactless measured ETS variables showed good utility, but unfortunately, the pupil diameter is largely influenced by other processes like ambient light adaptation which may complicate the detection in real driving situations. Issues of future research should be a further diversification of feature extraction to include a larger variety of features which is likely to improve accuracy and robustness of MSE detection. This could be a valuable contribution to future online driver monitoring technology, because for their improvement and validation it will be necessary to establish a reference standard of drowsiness and microsleep detection.
REFERENCES Buysse, D.J.; Reynolds, C.F.; Monk, T.H.; Berman, S.R.; Kupfer, D.J. (1989). The Pittsburgh Sleep Quality Index: A New Instrument for Psychiatric Practice and Research. Psychiatry Res., Vol. 28, pp. 193-213. Cortes, C.; Vapnik, V.N. (1995). Support Vector Networks. Machine Learning, Vol. 20, pp. 273-297. Gautama, T.; Mandic, D.; Van Hulle, M. (2004a). The Delay Vector Variance Method for Detecting Determinism and Nonlinearity in Time Series. Physica D, Vol. 190, pp. 167-176. Gautama, T. ; Mandic, D. ; Van Hulle, M. (2004b). A Novel Method for Determining the Nature of Time Series. IEEE Trans Biomed Engin, Vol. 51, pp. 728-736. Galley, N.; Andrés, G.; Reitter, C. (1999). Driver Fatigue as Identified by Saccadic and Blink Indicators. in Gale, A. (ed.) Vision in Vehicles - VII; pp. 49-59, Elsevier, Amsterdam. Golz, M.; Sommer, D.; Chen, M.; Trutschel, U.; Mandic, D. (2007). Feature Fusion for the Detection of Microsleep Events. J VLSI Signal Proc Syst, in press. Golz, M.; Sommer, D.; Seyfarth, A.; Trutschel, U.; Moore-Ede, M. (2001). Application of Vector-Based Neural Networks for the Recognition of Beginning Microsleep Episodes with an Eyetracking System. In: Kuncheva, L. (ed.) Comput Intell: Methods & Applic, 130-134. Hammer, B.; Villmann, T. (2002). Generalized Relevance Learning Vector Quantization. Neural Networks, Vol. 15 (8-9), pp. 1059-1068.
10
Ingre, M.; Akerstedt, T.; Peters, B.; Anund, A.; Kecklund, G. (2006). Subjective Sleepiness, Simulated Driving Performance and Blink Duration: Examining Individual Differences. J Sleep Res., Vol. 15, pp. 47-53. Joachims, T (2002). Learning to Classify Text Using Support Vector Machines. Kluwer, Boston. Kantz, H.; Schreiber, T. (2004). Nonlinear Time Series Analysis. Second edn. Cambridge University Press. Kohonen, T. (2001). Self-Organizing Maps. 3rd ed., Springer, Berlin. Lal, S.K.L.; Craig, A. (2002). Driver Fatigue: Electroencephalography and Psychological Assessment. Psychophysiology, Vol. 39, pp. 313-321. Percival, D. B.; Walden, A. T. (1993). Spectral Analysis for Physical Applications. University Press, Cambridge. Sagberg, F.; Jackson, P.; Krüger, H-P.; Muzet, A.; Williams, A.J. (2004). Fatigue, Sleepiness and Reduced Alertness as Risk Factors in Driving. Project Report, Transport RTD. Sommer, D.; Golz, M.; Trutschel, U.; Mandic, D. (2005). Fusion of State Space and Frequency-Domain Features for Improved Microsleep Detection. in Duch, W. et al. (eds.): Int Conf Artificial Neural Networks (ICANN 2005); LNCS3697, 753-759, Springer, Berlin. Sommer, D.; Golz, M. (2006); A Comparison of Validation Methods for Learning Vector Quantization and for Support Vector Machines on Two Biomedical Data Sets; in Spiliopoulou, M et al. (eds.): From Data and Information Analysis to Knowledge Engineering. Springer, Berlin, pp. 150-157. Sommer, D.; Golz, M. (2007). Feature Reduction for Microsleep Detection. Proc. 52nd Int. Sci. Koll., Technical Univ. of Ilmenau, Germany, to appear. Trutschel, U.; Guttkuhn, R.; Ramsthaler, C.; Golz, M.; Moore-Ede, M. (1998). Automatic Detection of Microsleep Events Using a Neuro-Fuzzy Hybrid System; Proc 6th Europ Congr Intellig Techn Soft Comput (EUFIT98), Vol 3, 1762-66, Aachen, Germany.
11
