I – Introduction

1Loss of driver alertness is almost always preceded by a period of measurable performance decrements and associated psychophysiological signs (Dingus, Hardee, and Wierwille, 1987 ; Vallet et al, 1993). Unfortunately, drivers themselves are often unaware of their deteriorating condition or, even when they are aware, are often motivated to keep driving (Itoi et al, 1993). As part of its Intelligent Vehicle Highway System (IVHS) and driver ergonomics research programs, the U.S. Department of Transportation’s (DOT) National Highway Traffic Safety Administration (NHTSA) is supporting research to develop in-vehicle Systems to continuously monitor drivers and their driving in order to provide a warning to drivers of their deteriorating alertness and performance (NHTSA, 1992). Scientific support for the feasibility of this countermeasure concept is provided by research showing that :

• Drowsiness can be detected with reasonable accuracy using performance measures such as « drift-and-jerk » steering by the driver and fluctuations in vehicle latéral lane position.
• The use of direct, unobtrusive driver psychophysiological monitoring (e.g., of eye closure) could potentially enhance drowsiness détection significantly.
• The use of secondary auditory tasks (e.g., auditory récognition tasks presented to the driver via recorded voice) could further enhance detection accuracy.

2This paper describes the NHTSA drowsy driver research program and some of its key findings, including projects on accident data analysis, driver performance monitoring, and psychophysiological monitoring.

II – Crash problem size and characteristics

3NHTSA statistics for 1990 indicate that there were an estimated 57,000 crashes in which driver drowsiness was indicated on the Police Accident Report (PAR). This was about one percent of the 6.5 million police-reported crashes occurring in the U.S. that year. This 57,000 estimate is probably conservative because it includes police-reported crashes only, which are believed to be only about one-half of all crashes. More importantly, the police or even drivers themselves may often be unaware of the role that drowsiness played in their crashes.

4In 1990, drivers of passenger vehicles represented 96 % of drowsy driver crash involvements, while those of combination-unit trucks (tractor-semitrailers) represented 4 %. Surprisingly, combination-unit truck drivers actually had a lower rate of involvement in these crashes compared to passenger vehicle drivers (4.4 versus 3.4 per 100 million km). However, other statistics show that combination-unit trucks are far more likely to be involved in these crashes than other vehicle types, and, therefore, are a very important testbed for drowsy driver countermeasures. On average, individual combination-unit trucks travel about six times more kilometres annually than do passenger vehicles. In addition, they have somewhat longer average operational lives (nearly 15 years) than do passenger vehicles (about 13 years). Thus, even though their rate of involvement per km is less than that of passenger vehicles, their expected number of involvements per vehicle life cycle is about five times greater. For vehicle-based countermeasures that last the life of the vehicle, the latter statistic is much more relevant to a détermination of potential benefits per unit cost. In addition, combination-unit truck drowsy driver crashes are generally more severe in their injury and property damage consequences.

5Important driver sex and âge patterns of involvement have also been found. In 1990, male drivers comprised 77 % of drowsy drivers involved in crashes, while they represented only 65 % of travel (km) and 51 % of driver registrations. Male drivers had an involvement rate (per km) that was 1.8 times greater and an involvement likelihood (i.e., involvements per registered driver) that was 3.1 times greater than females.

6Drivers under 30 years of âge comprised 62 % of drowsy drivers involved in crashes in 1990, while accounting for only 28 % of travel and driver registrations. Both their involvement rate (per km) and likelihood (per registered driver) were more than four times those of drivers aged 30 or over.

7Of course, male drivers and young drivers tend to be overrepresented in all crashes. However, these sex and age differences are much greater for drowsy driver crashes than for crashes in general.

8The above sex and age differences raise a concern about countermeasure deployment and market penetration. Will the drivers most in need of assistance – young males – be willing to buy and use a drowsiness countermeasure ?

III – U. S. DOT RESEARCH PROGRAMS ON OPERATOR VIGILANCE

9All modal administrations of the U. S. DOT have active research programs on operator vigilance. Across the various transportation modes, these programs generally focus on one of the following four themes :

• Operational policies such as manning requirements for ships (with implications for work and watch cycles), hours-of-service regulations for commercial vehicle drivers, and « strategie napping » policies for airline pilots on long flights. In regard to the latcr, studies sponsored by the Federal Aviation Administration (e.g., Rosekind, Gander, and Dinges, 1991) have shown that preplanned cockpit naps for each of the three crew members on extended flights can significantly improve subséquent alertness and performance on the flightdeck.
• Fitness-for-duty testing for operators prior to vehicle operation. For example, the Federal Transit Administration is assessing performance-based fitness-for-duty tests for transit operators that would identify performance déficits due to a variety of causes including alcohol, drugs, illness, and fatigue.
• Workstation or other vehicle design to reduce operator mental workload and increase alertness. For example, the U.S. Coast Guard is researching the use of greater automation and other ship control design concepts to reduce ship pilot workload. A joint Federal Highway Administration/NHTSA program is assessing the effects of longer commercial vehicle (e.g., tractor with three traders) operation on driver stress and fatigue.
• Continuous monitoring of operator status/performance. The Federal Railroad Administration is sponsoring research on innovative methods to monitor railroad engineer alertness. Conventional simple-response techniques can be defeated by drowsy engineers ; that is, they can respond correctly to secondary task probes even when they are functionally asleep. NHTSA’s principal drowsy driver research program, described in this paper, also focuses on this approach.

IV – Research on driver performance monitoring

10In 1991, NHTSA initiated a three-year cooperative agreement with Virginia Polytechnic Institute and State University (VPISU) to develop a vehicle-based capability for unobtrusively monitoring driver performance. The envisioned System entails continuous measurements of driver performance, data processing to « decide » whether the driver is drowsy, and an appropriate warning System interface with the driver (Wierwille, Wreggit, and Mitchell, 1992). Unlike most IVHS crash avoidance Systems, this would be a driver status warning System as opposed to a warning of a spécific, imminent collision threat.

11A successful System will require very high detection accuracy and, even more importantly, a low false alarm rate (Knipling, 1993). The false alarm rate is critical because, for example, a hypothetical System with 100 % accuracy of detection (drowsiness is always detected) and a 1 % false alarm rate (i.e., warning sounds 1 % of time when driver is not drowsy) would still yield more false alarms than « hits. » For example, if drivers are actually drowsy 0.1 % of ail time driving, this hypothesised System would still have 10 false alarms for every « hit. »

12In the current experiments, sleep-deprived subjects are tested on a driving simulator ; the test scenario is a desolate rural highway at night. Data from these experiments will be used to refine multiple regression prediction models in which performance measures and their derivatives are the opcrational predictor measures. These measures are each « baselined » against initial performance. During the monitoring, measures are initially computed over one-minute intervals and then averaged across six consecutive intervals. They are then combined through multiple regression to predict « actual » drowsiness during that six-minute period. One sensitive and reliable definitional measure of « actual » drowsiness is the proportion of time that the driver’s eyelids are closed 80 % or more. This definitional measure was named « PERCLOS. » The predictor measures are potentially obtainable in vehicles during actual driving, whereas definitional measures like PERCLOS need be obtained only in the research setting (Wreggit, Kim, and Wierwille, 1993).

13Table 1 presents a few minutes of sample data from a single subject to illustrate the extreme variations in alertness seen in a typical simulator session and the corrélation of performance measures such as latéral lane deviation with psychophysiological measures such as PERCLOS. During minutes 10, 11, and 12, the subject is still alert as indicated by the low mean Thêta (MNTHETA) amplitude of the EEG and zéro values for PERCLOS (i.e., eyes open). The subject is lane tracking precisely and shows few large or medium steering reversais (instead, an alert driver makes many small steering reversais to guide the vehicle). Eighty minutes later (minutes 90, 91, and 92), the subject is dangerously drowsy. Mean Theta amplitude has more than doubled, and PERCLOS increases to around 50 %. All five of the driving performance measures shown in Table 1 exhibit dramatic increases.

14Table 2 lists and defines six operational measures used to predict PERCLOS in one optimized détection algorithm. Multiple régression Bcta (B) weights and levels of significance are provided for each. All six operational measures are computable from steering, lateral accelerometer, and lane tracking variables. The lane tracking measure LANDEV is the strongest single factor. Multiple régression based on these factors yields a multiple régression coefficient (R) of + 0.872 between predicted drowsiness and actual drowsiness as measured by PERCLOS.

Table 1. Sample Data for Driving Subject When Alert and Drowsy

15As noted, a false alarm problem is inhérent in any attempt to predict low probability events such as driver loss-of-alertness (low probability when compared to ail time spent driving). However, the false alarm problem appears less daunting from the perspective of multiple degrees of alertness and intensities of warnings/advisories.

16Figure 1 shows a theoretical relation between « actual » drowsiness level (and thus actual risk of loss-of-alertness) and « operational » drowsiness level as measured/derived by a détection system. Three levels of « actual » and « operational » drowsiness are shown. Since the System is not perfect, its data points form an ellipse rather than a straight line. Within this scheme, zones G, E, and C represent perfect classification ; zones D, B, H, and F represent small misclassifications ; and zones A and I represent large misclassifications. Drowsy driver détection algorithms must be refined to a point where zones A and I are minuscule compared to the « universe ». The effects of small misclassifications (Zones D, B, H, and F) on crash prévention, driver performance, and driver acceptance must be determined through further research. For example, the « partial-false alarm » zones D and B could be a source of irritation to drivers or they could have the positive effect of reassuring drivers that the System is functioning continuously.

17Using a 3-level decision matrix as shown in Figure 1, Wreggit, Kirn, and Wierwille (1993) assessed the accuracy of the Table 2 algorithm for 25 six-minute intervals for each of 12 sleep-deprived subjects (300 trials total). They obtained the following classification accuracy levels : 79 % « perfect » (i.e., Zones G+E+C), 19 % small misclassifications (Zones D+B+H+F), and 2 % large misclassifications (Zones A+I).

Figure 1. « Actual » and « Operational » Drowsiness Level – A 3-Level Decision Matrix

18System accuracy could be further enhanced using various techniques currently under research. For example, this algorithm (perhaps aided by a direct psychophysiological measure) could perform an initial classification. Then, when an initial indication of drowsiness is detected, the driver could be requested to perform a subsidiary task that is presented auditorially. The driver might respond by means of pushbuttons on the steering wheel spokes or orally if a voice récognition capability is présent. If the driver fails the initial performance/psychophysiological assessment and then fails the subsidiary task also, there would be prima facie evidence of impairment. In addition to serving as a « final stage » alertness test, the auditory secondary task may alert the driver somewhat and delay recurrences of drowsiness.

V – Research on direct psychophysiological monitoring

19Direct, unobtrusive psychophysiological driver measures (especially of eye activity) could also be integrated into the measurement/decision regimen to assess driver status. Potential psychophysiological measures include heart rate variability, respiration rate, hand grip pressure on steering wheel, head inclination, measures of whole-body posture, electroencephalograms (EEGs), electrooculograms (EOGs), and various measures of eyelid activity (especially « slow closure »). The R & D challenge here is to develop unobtrusive or « minimally-obtrusive » devices that drivers are willing and able to use regularly and which do not interfere with normal driving performance. Device cost is also critical due to cost-benefit and marketability concems.

20NHTSA is supporting two R & D efforts to develop direct measures of eye closure. One approach, under investigation by Systems Technology, Inc. using a device developed by SRD Shorashim Medical Ltd., involves measurement of electroocular and neuromuscular potentials (EOG and EMG) associated with eye closure using a head-band/headset device. A second approach, under research by MacLeod Technologies, involves détection of eyelid closure using opto-electronic techniques and miniaturized emitters and sensors. Both of these approaches are minimally obtrusive, employ established technologies, and hâve the potential to be very low-cost (e.g., less than $100). Current and future R & D will determiné whether they are comfortable and unobtrusive enough to be worn for extended limes and whether they provide reliable data in a real vehicle setting.

21Another approach to eye closure detection involves the use of a dashboard-mounted video camera and sophisticated image processing. This approach has the potential to be completely unobtrusive, and could be adapted for applications other than drowsiness détection. For example, it could discern the driver’s point of regard (i.e., where the driver is looking) and thus be used for additional forms of driver monitoring. PC-based prototype Systems exist, although at présent they may be too expensive for widespread commercial use. Extensive image processing is required to deal with problems such as driver head movements and the partial obstruction caused by eye-glasses. Nevertheless, a number of U.S. vendors are actively exploring and promoting this technology. The agency is currently evaluating this technology for use in research vehicles. Commercial applications may corne later as device cost decreases.

VI – Research needs

22Numerous research needs must be addressed before a vehicle-bascd drowsy driver countermeasure can be deployed successfully. These research needs include the following :

23– Measurement technologies must be advanced. For example, continuous lateral vehicle lane position data will be an essential input. Several machine vision technologies for lane position monitoring are under development, but currently none has reached fruition in ternis of reliability and cost.

24– False alarms must be minimized. In addition to approaches already described, another way to increase detection and reduce false alarms might be to consider not just the current measurement time interval but also trends evident from preceding intervals. Were there early signs of developing drowsiness based either on the overall operational measure or among specifie indicators ?

25– The warning signal or other System response (e.g., automatic control) must alert drivers but not startle them, causing a panic response (Vallet et al, 1993). The System must also prevent recurrences of drowsiness – a major deficiency of some current Systems (Haworth and Vulcan, 1991). NHTSA is beginning to develop human factors guidelines for future drowsy driver warning Systems, although there are many challenges to the specification of an optimal design. For example, there are large individual differences in acoustic stimulus intensity (i.e., loudness) necessary for arousal from sleep (COMSIS, 1993). It may prove difficult to develop an idéal « standard » warning intensity for all drivers.

26– Extensive test and évaluation will be necessary to refine System algorithms and responses. This may include advanced simulation research, instrumented-vehicle research, and operational field tests. The National Advanced Driving Simulator (NADS) and the Portable Driver Performance Data Acquisition System (DASCAR) will be two new tools for performing this research. The NADS will provide a high-fidelity facility for simulating crash situations and countermeasures without endange-ring subjects. DASCAR will be a portable, unobtrusive, and inconspicuous human factors instrumentation suite for installation in production vehicles. The System will thus provide data on driver performance and behavior in actual traffic situations – for example, with and without the aid of a drowsiness countermeasure.

27– « Crosstalk » with other Systems must be considered. For example, the target crash populations of the drowsy driver détection System and a roadway departure warning System might overlap considerably. These two Systems would require compatible driver interfaces and also could share common data on driver performance to maximize accuracy.

28– Device cost should be low enough to ensure a favorable ratio of average benefits to cost. Since the « per-vehicle » crash problem size is much greater for combination-unit trucks than other vehicle types, their « break-even » device cost will be much greater. Of course, an imperfect relation exists between monetary cost-benefits and actual marketability and driver acceptance. For example, young males as a driver group would reap the greatest benefits from drowsiness countermeasures, but they may not be the drivers most willing to purchase it.