On 31 August 1997, Diana, Princess of Wales died in a car crash in Paris. The crash was ruled to be the consequence of her driver losing control over the vehicle because he was intoxicated and under the influence of prescription drugs. Her death brought home a message that has been at the center of road safety campaigns for decades: Do not drive while intoxicated. Fortunately, this message has been heeded, and fatalities from drunk driving have declined dramatically: In the U.S., fatalities from drunk driving have declined by 64% during the period 1982-2016.

So far, so getting better.

However, there is another hidden killer behind the wheel. It’s called fatigue. In a recent post, we showed how cognitive performance can be impaired by prolonged mental exertion. And it is pretty obvious that falling asleep behind the wheel is a recipe for disaster. It has been estimated that between 19% and 24% of all fatal crashes involve driver drowsiness.

No wonder, then, that road safety authorities around the world are running campaigns against drowsy driving:

Unfortunately, however, to date those campaigns may not have been as effective as the fight against drunk driving. According to the U.S. National Highway Traffic Safety Administration (NHTSA), fatalities from drowsy driving have remained relatively stable for the last decade.

Fortunately, technology is being developed that may help detect when a driver is drowsy or fatigued, and may then alert the person that it is time to pull over for a break. To date, the drowsiness-detection technology has explored a number of physiological measurement techniques, from electroencephalogram (EEG) to electrocardiography (ECG) and electrooculography (EOG). Other technologies have relied on performance-based measures such as steering behaviour.

There are at least two problems with potential technological solutions to detecting driver fatigue: first, any solution that requires attachments to the body are unlikely to be accepted by drivers—who would want to wear various electrodes all over their head on a holiday trip? Second, given that self-driving vehicles are on the horizon (with fleets being rolled out maybe as early as 2021), how can driver-fatigue technology work when drivers, whether fatigued or not, do not do much of the driving?

A recent article in the Psychonomic Society’s journal Behavior Research Methods tackled those questions by testing camera-based fatigue detection in both manual and conditionally-automated driving modes.

Researchers Jürgen Schmidt, Rihab Laarousi, Wolfgang Stolzmann, and Katja Karrer-Gauß measured and analyzed drivers’ eye blinks as a potential proxy measure of drowsiness while participants drove in a simulator for nearly 3 hours at the end of their work day (starting at 6pm or 10pm). The simulated conditions were in the evening, with a dark, heavily overcast sky and a monotonous roadside to induce drowsiness. Participants were asked to rate their fatigue on a self-report scale every 15 minutes.

In the manual condition, participants drove a (simulated) car entirely on their own. In the conditionally-automated condition, participants could engage the automation (i.e., switch to “autopilot”), which then took over all lateral and longitudinal movements of the vehicle at a speed of 100 km/h (around 60mph). When the car was on autopilot, the driver could take back control at any moment by operating the brake or accelerator or by moving the steering wheel. The automation was “conditional” because it could only be engaged under certain conditions—that is, when a certain speed had been reached and the vehicle was centered in a lane and so on. For that reason, the autopilot could never take over completely and the driver had to remain ready to resume control of the vehicle at all times.

The figure below shows the driving simulator that was used in the manual condition. (Hint: the pod contains an entire car plus 360-degree surround screens. You may guess the marque.)

Schmidt and colleagues were primarily concerned with validating various automatic algorithms that could detect blinks from the video-recordings made during the driving session. This examination revealed that accuracy of the automatic blink detection exceeded 90% in the manual condition when drivers were classified as awake or alert (based on their own self-ratings). When drivers in the manual condition were drowsy (again identified based on their own ratings), blink detection accuracy remained high but declined to somewhere around 90%.

A further analysis of the details of eyelid behaviour, which went beyond simple blink detection to specification of the eyelid movement, revealed that alert drivers do not close their eyes during a blink as far as drivers who are drowsy, pointing towards a possible avenue towards automatic detection of drowsiness.

In addition, it was found that blink-detection accuracy rates declined notably in the conditionally-automated condition. When drivers under automation were alert, blink detection was reduced to somewhere between 70% and 80% (depending on the details of the algorithm). When drivers under automation were drowsy, the performance was reduced further, to around 60%-70%.

These results indicate that people’s eyelid movements change under automation, irrespective of their degree of alertness. It follows that algorithms for drowsiness-detection during manual driving cannot be automatically applied to autonomous (or semi-autonomous) vehicles that still require the driver to remain alert.

So how close are we to reliable automatic drowsiness detection? It is clear that much work remains to be done before we can hand over monitoring of our own alertness to our car. We may well be able to hand over driving altogether sooner and with greater confidence than monitoring of our alertness. This will not put cognitive scientists out of business because the ethical implications of autonomous vehicles are immense.

Psychonomics article focused on in this post

Schmidt, J., Laarousi, R., Stolzmann, W., & Karrer-Gauß, K. (2018). Eye blink detection for different driver states in conditionally automated driving and manual driving using EOG and a driver camera. Behavior Research Methods, 50, 1088-1101. DOI: 10.3758/s13428-017-0928-0.