In this study, researchers Liang (Liberty Mutual Institute for Safety), Lee (U of Wisconsin) and Yekhshatyan (U of Iowa) compared 24 algorithms intended to estimate driver distraction and predict crash and near-crash risk as it relates to driver glance behavior.
They provide the following example statistics ion their introduction:
- 16% of fatal crashes and 21% of injury crashes were attributed to driver distraction
- driver inattention and distractions were associated with 80% of crashes and 65% of near-crashes
- risk of crash/near-crash was approximately 3 times greater when drivers engaged in complex secondary tasks requiring multiple glances away from the road and/or multiple button presses
- off-road glances greater than 2 seconds lead to slower breaking response and greater lane deviation
- glances to in-vehicle displays located farther than the center of the road result in slower response to hazardous events
- cognitive distractions increase risk, especially when combined with visual distractions
The researchers postulate that algorithms that quantify drivers’ eye-glance patterns in terms of glance duration, glance history, and glance location, could be a useful way to assess risk related to visual driver distraction, as well as some cognitive distractions, since the two are often intertwined (e.g., glancing at a keypad while texting). However, some cognitive distractions, such as talking on a hands-free, do not involve eye-glance distractions.
Other researchers have already investigated and found evidence for these visual distractions.
- Glance durations: the longer an off-road glance, they postulate, the more likely it will be detrimental to driver reaction times, and may lead to a complete miss of critical events.
- Glance history: driver glancing behavior over time, they postulate, could suggest some level of distraction, where a pattern of glancing may indicate driver attention to a secondary task, for example.
- Glance location: the farther away from the road a glance is, they postulate, the more it may reduce drivers’ awareness of roadway situations.
Methods
Liang et al. conducted their analysis using the "100 Car Naturalistic Driving Study," a data set collected from over 241 drivers who were unobtrusively monitored over approximately 2 million miles of actual driving conditions. This data set was collected by other researchers, and has been used for a variety of driver behavior studies. For this study, they relied on two of the 100-Car databases:
- Events, which included 68 crashes and 760 near-crashes; and
- Baseline data allowing them to match those events to such things as driver, time-of-day, GPS location, etc.
They only included events that contained appropriate eye-tracking data and in which the driver was determined to be at fault, resulting in a total of 359. Building on the work of other researchers, the researchers then tested a variety of algorithms for their ability to predict increased crash/near-crash risk as defined by an increased odds-ratio. They present a great deal of detail on their methodology in the full article, referenced below, and interested readers are encouraged to read it for a complete understanding of their methods.
Results
Their key findings include:
- driver eye-glance patterns can indeed indicate driver distraction and crash/near-crash risk;
- the most sensitive estimations of risk were algorithms that accounted for instantaneous changes of off-road glance duration (vs. glance history), suggesting that the effects of eye glancing and crash risk are instantaneous, and not as well predicted by a historical pattern of eye glancing;
- surprisingly, glance location was not predictive of risk in this study, which is inconsistent with previous research findings and theory; they postulate that this is due to less precise data available to them in the 100-Car data set, and suspect that this variable would have been more predictive had they had access to precise eye-glance location data;
- surprisingly, glance history was not predictive of risk in this study, which is also inconsistent with previous research findings and theory; the researchers again suggest limitations in the data set may explain the lack of correlation, and they also suggest that short-glance histories may be a better measure than long glance histories. They also note that the 100-Car data set was collected before mobile phone and texting devices became prevalent.
What this Might Mean for Ergonomists
I have to admit, when I first saw this study I was very hopeful it would contain more detailed and precise findings that we could begin to readily apply. However, as I dissected the entire article it became apparent that it left more questions unanswered than answered. Essentially, they found that glancing away from the road does indeed increase crash/near-crash risk, which is no surprise. They also found that the 100-Car data set, which predates the influx of mobile phones and texting devices, does not have the level of detailed eye-tracking data needed to achieve more precise predictive algorithms. So, why did they even rely on such "old" and "incomplete" data in the first place? Presumably because there is no better data available for real drivers in real driving conditions that involved crashes and near-crashes, and by spending the time to analyze this existing data set, they can now make a strong claim that "additional research is needed," the siren song of all research studies.
The biggest take-away from this study, then, is that we still don't know enough to really understand the underlying mechanisms of driver distraction as they relate to eye-glancing behavior and the likelihood of a crash. Previous studies have tied glance duration, glance history and glance location to increased distraction and risk, but the precise relationships between these variables remains unclear.
Does that mean we're all safe to start phoning, texting, etc. while we drive? Of course not, but it does mean that if we're going to come up with design solutions that address the root-causes of driver distraction related crashes, which is the ultimate goal of ergonomics in this case, we're going to need better data. Sigh. The pace of research is slow indeed.
Reference
Yulan Liang, John D. Lee, and Lora Yekhshatyan, (2012). How Dangerous Is Looking Away From the Road? Algorithms Predict Crash Risk From Glance Patterns in Naturalistic Driving. Human Factors, Online First, DOI:10.1177/0018720812446965, avaialble to Human Factors subscribers at the time of this wrighting at http://hfs.sagepub.com/content/early/2012/06/15/0018720812446965.full.pdf+html.
This article originally appeared in The Ergonomics Report™ on 2012-06-27.