Information As A Source of Distraction

Chapter 6. The Effect of CMS Information on Detection of Safety-Critical Events in the Roadway

Introduction

The previous chapter reported on a test of the hypothesis that driving-irrelevant CMS content would cause drivers to lose respect for the signs and then miss important traffic-related messages. The findings did not support that hypothesis. This chapter tests the hypothesis that CMS content distracts drivers from attending to safety-critical information in the roadway. A spilled load of logs was simulated in the roadway 300 ft (91 m) upstream of a CMS. The spilled load was in an area where the previous study showed that glances at the signs were most likely. That is, first glances at a CMS occurred between 901 and 555 ft (275 and 169 m) before reaching the CMS, and last glances occurred in the range 590 ft to 180 ft (180 and 55 m). The primary dependent measure was whether or not the driver avoided hitting the spilled load by changing lanes, braking, or a combination of these responses. As in the previous experiment, glance behavior and speed were also assessed.

To assess whether salient driving-irrelevant content might be more visually distracting than travel-time information or blank CMSs, the sign content that was visible when the logs came into the line-of-sight was varied among participants. To assess whether distraction effects might be greater near the beginning or end of a trip, the location of the spilled load was also varied between groups. There are several reasons that the signs might be more distracting at the beginning or end of the trip. The signs might be more distracting at the beginning of a trip if their novelty attracts attention. None of the drivers in the previous experiment participated in this experiment, so initially the faces signs might be expected to be novel. Signs might be more distracting at the end of the trip because drivers might seek additional stimulation as they begin to feel bored or fatigued. Finding a difference in spilled load response between the beginning and end of the drive would not test the involvement of novelty or fatigue as distraction facilitators, but it might indicate where to look for distraction effects in future studies.

In the experiment reported in chapter 5, the finding that faces and travel-time signs attracted similar amounts of attention suggests that the frequently changing faces displays were weak in capturing attention. In the present experiment, a face-recognition test was administered after participants finished the experimental drive. Participants were not informed in advance that they would be tested for recall of the pictures shown on the overhead signs. The purpose of the test was to provide another measure of the degree to which the face stimuli captured attention.

Method

The same driving simulator and eye-tracking system were used in this experiment as were used the experiment reported in chapter 5.

The blank, travel-time, and faces signs were used again in this experiment. However, in this experiment, all participants were presented with all three CMS content types. Each of the content types appeared on every third sign. As described in the following subsections, the content of the first sign in this sequence depended on the group to which the participant was assigned.

The Simulation

The simulated freeway was the same as that described previously except that the drive was reduced to 37 mi (59.5 km) and 72 CMSs. There were two between-group conditions: (1)whether the CMS beyond the spilled load of logs displayed the faces, the travel-time, or the blank content and (2) whether the spilled load was before the 4th CMS encountered or before the 72d CMS. An example of the appearance of the spilled load is shown in figure 42.

Figure 42. Screen capture. Driver’s view of spilled load from 128 ft (39 m).

Traffic was generated at a rate of 5,000 vehicles per hour for the first 6.7 min. Vehicles entered and exited from every other ramp intersection at a rate of 500 vehicles/h. Because the spilled load was in the second lane from the right, participants were instructed prior beginning the test to drive in that lane whenever it was safe to do so and to try to maintain the posted speed of 65mi/h. Participants were told that they would receive a $10 bonus if they drove in the instructed lane and maintained 65 mi/h whenever possible. To prevent slower vehicles from motivating participants to change lanes, the minimum desired speed of all other vehicles was set to 69 mi/h. Fifty-five percent of simulated other vehicles were programmed to seek to travel 69to 71.7 mi/h, while the remaining vehicles were programmed to seek speeds between 71.7mi/h and 80 mi/h. In the left-most lane, all vehicles were set to maintain 80 mi/h.

In the data collection zones where the spilled load was placed, other traffic began clearing the lane containing the logs 1,211 ft (369 m) before the logs. This distance, plus the average time headway of 559 ft (170 m), ensured the participants had adequate sight distance to detect the logs and respond to them by hard braking or changing lanes.

There were zones in which all traffic slowed markedly. These zones were in the same locations relative to the start of the simulation as in the previous experiment. Because the scenario was shorter than in the experiment reported in chapter 5, there were only 5 congestion locations, the first located between the 4th and 5th CMS, and thereafter there was 1 congestion zone for every 15 CMSs (e.g., the next congestion zone was between the 19th and 20th CMS). The purpose of these zones was to keep the participants engaged in the driving task. Gaze behavior was not analyzed in these zones.

Recognition Test

Upon exiting the driving simulator, participants were shown a series of 64 pictures on a laptop computer. Half of the pictures (32) were pictures of faces that had been displayed during the drive. The other half were foils (similar face pictures that had not been shown during the drive). For celebrities and other well-known persons (e.g., Barack Obama, Elizabeth Taylor, and Prince Charles), a different picture of the same individual was included as a foil for the picture in the simulation. For other images, the similarity of the foils was based on salient characteristics of the target pictures (e.g., hair color and style, nationality of dress, and facial expression). Each picture (whether target or foil) was presented individually, and participants were asked to indicate “yes” if they had seen the picture during the drive or “no” if they did not recall the picture from the overhead signs. Participants were instructed to guess when they were not sure.

Participants

Complete data, including interpretable eye tracking, were obtained from 73 participants. A total of 80 participants (51 males and 29 females) completed the drive and provided behavioral performance data. The median age of the participants who completed the test was 33.5 years (range 18 to 73 years). An additional five participants failed to complete the study, two because of simulator sickness symptoms and three because of driving simulator hardware or software failures.

Results

Throughout this report, error bars in charts and graphs represent 95-percent confidence limits around the means.

Response to the Encounter With a Spilled Load

The spilled load presented the participants with a challenging task as evidenced by 26 percent of drivers hitting the logs. However, the message content of the CMS that was visible as the spilled load was approached had no significant effect on whether the logs were avoided. Nor did the location of the logs have a significant effect on the probability of hitting the logs. Table 25 shows the frequency of participants hitting or avoiding the logs as a function of CMS content and location of the spill.

Table 25. Number of participants who hit or avoided spilled load shown as a function of load location and CMS content.

Gaze Behavior

Gaze behavior was scored for a subset of 27 of the 72 CMSs in the study. Gaze was scored for signs 1–9, 16–18, 32–34, 48–50, and 64–72. The same three measures of gaze behavior were examined: glances, looks, and fixations. All data collection zones began 10 s upstream of the point where the sign passed from view and ended when the sign passed from view. There were nine zones with each type of sign content.

Glance Results

The probability of glancing at each sign was examined as a function of sign content (faces, travel time, or blank) and time headway. Each sign was classified by whether it received a glance (no = 0, yes = 1) and by whether the mean headway in the data collections zone (i.e., 1.5 s or less) or long (i.e., greater than 1.5 s). The data were analyzed using a GEE model that assumed a binomial distribution with a logit link function. Sign content, headway, and their interaction were modeled as predictors of the probability of a glance. Only the effect of sign content was significant, χ² (2) = 41.06, p < 0.001. As can be seen in figure 43, the probability of a glance at faces and travel-time signs was about the same, and the probability of a glance at blank signs was about half that for the non-blank signs. These findings are quite similar to those for the previous experiment when headways were long.

There were two important differences between this and the experiment reported in chapter 5 that may have resulted in the failure to find a headway effect in this experiment. First, in this experiment, outside the congestion areas, other traffic always traveled at speeds greater than 65mi/h, whereas 12 percent of traffic traveled at less than 65 mi/h in the previous experiment. Second, the data collection areas in this experiment did not contain, and were not near, congested zones where headways would shrink. As a result, there were few opportunities for short headways in this experiment. Mean headway distance in this experiment was 559 ft (170 m) compared with 326 ft (99 m) in the previous experiment. Because headway had no effect on the glance results, headway was dropped from the subsequent look and fixation analyses.

Figure 43. Chart. Predicted probability and confidence limits for a glance at a CMS as a function of sign content.

Glance duration was computed as the sum of all 0.0083-s glance vectors to the sign ROIs. The GEE models with glance duration as the predicted variable and sign content, headway, and the content by headway interaction were tested. These models assumed a gamma distribution with identity link function. Because the gamma distribution does not include zero, two durations were computed for each ROI: (1) the first method included all ROIs regardless of whether they received a glance, and where no glances were recorded, glance duration was assigned a duration of 0.00001 s; and (2) the second measure included only ROIs that received a glance. Both models led to similar conclusions. Only the sign content main effects were significant: (1) with zero durations coded as 0.00001 s, χ²(2) = 41.2, p < 0.001, and (2) with zero durations coded as missing, χ²(2) = 27.96, p < 0.001. Figure 44 shows predicted glance durations and their confidence limits by both methods of computing glance duration. By either method, the durations to faces and travel-time signs were not significantly different from each other, and the durations of glances at blank signs were significantly less than at the non-blank signs.

Figure 44. Chart. Expected mean duration of glances at CMSs and confidence limits as a function of sign content.

The expected distance, in feet, for the end of a glance is shown in figure 45. The effect of sign content on end-of-glance distance was significant, χ²(2) = 31.18, p < 0.001. Post hoc tests indicated that glance-end distance to blank signs was significantly greater than glance-end distance to travel-time or faces signs, which were not significantly different from each other.

1 ft = 0.305 m

Figure 45. Chart. Expected glance-end distance as a function of sign content.

Look Results

The probability that a participant would look at a CMS was modeled with sign content as the predictor. A binomial response distribution and logit link function were assumed. The sign content main effect was significant; χ²(2) = 31.83, p < 0.001. Post hoc comparisons indicated that the probability of looking toward a blank sign was significantly less than the probability of looking toward a faces sign, χ²(1) = 27.39, p < 0.001; or travel-time sign, χ²(2) = 30.66, p < 0.001. The difference between the probability of looking at faces signs and at travel-time signs was not significant. The predicted means and respective confidences limits are shown in figure 46.

The number of looks at each CMS was modeled as a function of sign content. The GEE model assumed a Poisson response distribution and log link function. Sign content was significant; χ²(2) = 25.55, p < 0.001. Post hoc comparisons indicated that the number of looks toward a blank sign was significantly less than the number of looks toward a faces sign, χ²(1) = 25.52, p < 0.001, or a travel-time sign, χ²(1) = 21.40, p < 0.001. The difference in the number of looks at faces signs and at travel-time signs was not significant. The predicted means and the confidence limits for those means are shown in figure 47.

Figure 46. Chart. Probability of at least one look at a CMS as a function of type of sign content.

Figure 47. Chart. Predicted number of looks at a CMS as a function of type of sign content.

The number of looks at signs that received at least one look and the average duration of those looks were also modeled. GEE was used to model the number of looks assuming a Poisson response distribution and log link function. Given that there was at least one look at a sign, knowing the content of the sign did not add additional predictive information. The predicted mean number of glances given at least one look is shown in figure 48. The duration of individual looks to the CMSs was modeled to determine whether sign content had a significant effect on look duration. Only signs that the participant looked at (i.e., for which at least one look was recorded) were included in the analysis. A gamma response distribution and identity link function were assumed. The resulting expected mean look durations are shown in figure 49. Sign content was a significant predictor of look duration, χ² (2) = 6.58, p = 0.037. Post hoc comparisons suggest that the main effect was the results for looks at blank signs, which were significantly shorter than looks at faces signs, χ² = 6.50, p = 0.011. There was no significant difference between the duration of looks toward faces signs and travel-time signs.

Figure 48. Chart. Number of glances at each CMS that received at least one look.

Figure 49. Chart. Predicted mean duration of individual looks as a function of sign content.

Fixation Results

GEE were used to model the probability of a participant fixating on each CMS. A binomial response distribution and logit link function were assumed. The predictor was the sign content. Sign content was a significant predictor of fixation probability, χ²(2) = 84.30, p < 0.001. Post hoc comparisons indicated that the probability of fixating on a blank sign was significantly less than the probability of fixating on faces or travel-time signs and that the difference between the probabilities of fixating on travel-time and faces signs was not significant. Figure 50 shows the probability of fixating on individual CMSs as a function of their content.

Figure 50. Chart. Expected mean probability of a fixation on a CMS as a function of sign content.

GEE models were used to model the number of fixations on the CMSs. A Poisson response distribution and log link function were assumed, and sign content was the predictor. Figure 51 shows the predicted mean number of fixations as a function of sign content. Sign content was a significant predictor, χ²(2) = 35.90, p < 0.001. Post hoc comparisons indicated all three predicted means were significantly different from each other.

Figure 51. Chart. Predicted mean number of fixations across all signs as a function of sign content.

Mean fixation duration was examined for those signs that received at least one fixation. GEE models were used to evaluate fixation duration. A gamma response distribution and identity link function was assumed. Sign content was the predictor variable. Sign content significantly predicted fixation duration, χ²(2) = 15.10, p = 0.001. Post hoc comparisons indicated mean fixations on travel-time signs were significantly longer than those on blank or faces signs, and the latter two were not significantly different from each other. Figure 52 shows expected mean fixation durations for blank, travel-time, and faces signs along with the 95-percent confidence limits for those means.

Figure 52. Chart. Expected mean fixation duration as a function of sign content.

The average fixation duration was 0.5 s or less. However, the average duration may not reflect the existence of long fixations on CMSs that might represent an unsafe driver distraction. For in‑vehicle device distraction, it is suggested that display devices that capture visual attention for more than 2 s represent a safety risk.^(55,56) Although no such guidance is available for display devices located above the roadway, it is doubtful whether displacement of gaze only a few degrees above the forward roadway would be unsafe. However, should generalization of the 2-s capture rule be considered appropriate, then a few unsafe fixations were observed. Table 26 shows that 1.8 percent of fixations on the simulated CMSs in this study were greater than 2 s.

Table 27 shows that more than 50 percent of the durations greater than 2 s were for travel-time messages.

Table 26. Distribution of fixation durations.

Table 27. Frequency of fixations greater than 2 s.

In chapters 5 and 6, beginning and end of glances were examined. It was found that glances began at about the same distance regardless of sign content and that glances at faces signs ended closer to the sign than did glances at blank or travel time signs. Similarly, mean fixation distance was examined. The overall test and post hoc tests were all GEE models, with sign content as the predictor for mean fixation distance. Mean fixation distance was assumed to be gamma distributed, and an identify link function was used. The main effect of sign content was significant, χ² (2) = 23.01, p < 0.001. The mean fixation distance results are in line with those for glances: expected mean distance was greatest for blank signs and least for faces signs, with travel-time signs falling between. Post hoc comparisons showed all three expected means, as shown in figure 53 are significantly different from each other.

1 ft = 0.305 m

Figure 53. Chart. Expected mean distance of fixations on a CMS as a function of sign content.

Driving Performance Measure

Travel speed was analyzed with a GEE model that assumed a gamma distribution and identity link function. The sign content of the data collection zone and location of the logs served as predictor variables. The travel speed dependent measure was based on mean travel speed for each data collection zone for each participant. Data collection zones 4 and 72 (the zones with logs in the roadway) were excluded from the analysis. Figure 54 shows the resulting expected mean speeds as a function of the content on the CMS and log location. The interaction of log location and sign content was significant, χ²(2) =8.94. p = 0.011. This unexpected interaction resulted because the group that encountered the logs late in the trip exhibited a nearly constant speed regardless of CMS content, whereas the group the encountered the spilled logs early in the trip drove significantly slower than average when approaching blank signs and significantly faster than average when approaching faces signs. Also striking is that all groups drove below the instructed and posted speed in all three data collection zone types.

No predictor variable (sign content, log location, or data collection zone order) was significantly related to speed variability.

1 mi/h = 1.6 km/h

Figure 54. Chart. Expected mean speed as a function of sign content.

Faces Recognition

The d' measure from signal detection theory was used to assess the ability of participants to distinguish between pictures they had been exposed to during the drive and similar pictures not previously displayed.^(57,58) The obtained estimate was quiet low, mean
d' = 0.25, SD (standard deviation) = 0.37. A d' of zero would indicate no ability to distinguish between targets and foils. This finding indicates that participants showed very little ability to distinguish the new pictures from the ones displayed during the drive.

Discussion

Distraction From Monitoring the Road Ahead

A total of 59 of 80 drivers were successful in detecting and avoiding the spilled load of logs. The detection task was difficult enough that many of the remaining 21 drivers struck the logs because they were indecisive in reacting, such as deciding to seek a gap in the adjoining lanes and then failing to find that gap in time to avoid the logs. A few drivers showed no indication of detecting the logs. In any case, failure to detect or react successfully to the threat did not appear to vary with the content of the sign that was being approached. The log avoidance measure yielded no evidence that driving-irrelevant stimuli, even stimuli considered highly salient in other contexts, would distract drivers from their primary (driving) task more than would a blank overhead sign.

Eye Gaze Distraction Evidence

The eye gaze data from all three metrics (glance, look, and fixation) converged on the same general conclusion: drivers were about equally likely to shift their visual attention to travel-time messages as they were to colorful, changing, and driving-irrelevant content (e.g., faces), and less likely to shift visual attention to a blank CMS than to a sign with information content.

There were few fixations on the CMSs longer than 2 s, and the majority of these were on travel-time signs. It was thought that faces of celebrities or faces displaying strong emotions might attract attention more than predictably structured text messages. To enhance the saliency of the faces, they changed every 3 s. Nonetheless, the faces did not attract gaze more than travel-time messages.

Recognition of Faces

The finding that pictures not shown before could not be reliably distinguished from pictures that were repeatedly presented on the CMS may suggest that the foils chosen for use in the test were too similar to the targets. However, the findings also suggest that participants did not devote much attention to studying the pictures on the overhead signs (and the gaze data would seem to confirm this). Because the participants were not told that recognition of the face pictures would be tested, there was no motivation for them to study the pictures beyond whatever intrinsic value the pictures might hold. A stronger test of the ability of drivers to avoid distraction from the primary task—driving—might be to offer some incentive to drivers to attend to the CMS messages. The challenge of a study with incentives to attend to the CMSs would be to avoid making the driving task secondary because of unrealistic contingencies.

Summary and Conclusions

These findings suggest that messaging on CMSs that is not related to driving would be no more distracting than traffic-related messaging or blank signs. As in the previous experiment, these results apply to a single relatively long trip. Should drivers habituate to frequently occurring CMSs, then any distraction away from detecting road hazards should be less.

The findings indicated that when a message is displayed on a CMS, drivers move their center of gaze to the sign about 50 percent of the time. When the CMS is blank, drivers still have a 25‑percent probability of shifting their gaze to it. These gazes are generally short, about 0.5 s. The number of short glances is generally small, with an average of one fixation per sign with content.

The aims of this and the experiment reported in chapter 5 were limited to a strong test of whether travel-irrelevant content on CMSs would distract drivers. The answer is that under the conditions tested, irrelevant messaging, even if designed to be salient in other settings, would not be more distracting than traffic-related messaging, which itself did not appear to stress the visual attention capacity of drivers.

The present tests did not present the CMSs in environments with high amounts of visual clutter, as might be present in an urban environment with tall buildings, billboards, and overpasses, in addition to other critical highway signs. The tests simulated daytime conditions with signs that were not brighter than static signs or the simulated sky. The contrast ratio of white text to the black background, measured in a previous study in the simulator, was 14.6. This ratio is somewhat greater than that recommended by the MUTCD (8 to 12) (see section 2L.04, paragraph 11) but much less than contrast ratio of the CMS reported in chapter 3.⁽²⁾ Aside from the logs in the road in the experiment reported in this chapter, eye gaze behavior was not assessed in areas with hazardous driving conditions. The participants in these experiments were not young novice drivers. For these and other reasons, it is possible that in some conditions, with some driver populations, some CMS content might be distracting and pose a safety risk. These experiments suggest that it would be challenging to present visual information that would compel drivers to shift their attention from the primary driving task.