Traffic Control Device Conspicuity

Glance Behavior and Sign Recall

To fully address the effects of environmental factors on messaging in the ROW, the effects of environment on all four stages of processing identified by Lay (detection, reading, understanding, and action) must be addressed.⁽¹⁾ However, because each stage is dependent on the preceding stages, the present study focused on knowledge gaps regarding the first two stages: detection and reading.

As stated in the introduction, detection relies both on visibility and conspicuity. The factors that affect sign visibility are reasonably well understood.^(1,25) However, the specific factors that affect the conspicuity of signs and other TCDs are less well understood. Conspicuity refers to the properties of an object that make it likely to capture visual attention (i.e., noticeable). However, there is no single agreed upon way of measuring visual attention or the factors that influence it. The following section reviews some of the methods that have been used to measure visual attention.

Assessing Visual Attention

Modeling Approach

Itti and Koch discussed the development of computational models of visual attention that are based on physiological and psychophysical theories and are supported by robust empirical evidence.⁽²⁶⁾ The authors relate conspicuity to perceptual saliency, asserting that some objects are inherently salient in a given context (e.g., a red jacket among black tuxedos). They suggest that such saliency is the result of bottom-up visual processing-low-level automatic processing that is not influenced by higher cognitive processes or conscious attention. The authors posit that in a scene that has multiple salient objects a saliency map will determine which objects will be attended to first at the next higher level of processing (a still relatively low-level preattentive process). Itti and Koch acknowledge that experience and background properties can influence both of these processing stages. Both are bottom-up preattentive processes that are not subject to conscious control or top-down processing. A final preattentive process posited by the authors is inhibition of return, which prevents focus from returning to highly salient stimuli that have already been processed. Beyond these preattentive processes, top-down processes play a role in visual attention. These top-down processes involve relating memory, knowledge, expectation, and conscious control to making sense and use of the low-level visual information. Itti and Koch tested various computational models of visual attention for computing the time necessary to detect traffic signs. However, this basic research does not offer guidance that might be readily applied to highway design.

It has been suggested that conspicuous objects attract attention and that attention consists of both bottom-up processes and top-down processes. However, to understand the influence of conspicuity on messaging in the ROW, a methodology is needed to measure conspicuity in a way that is consistent and repeatable.

Verbal Reports

Verbal report methods have been used to assess both object conspicuity and which objects are attended to by drivers. Renge and Cole and Hughes asked drivers to talk aloud about what attracted their attention while watching films of scenes recorded from the driver's perspective.^(27,8) Renge asked participants to sit in a driving buck and mimic the steering and braking behavior presented in the driving scenario. Participants continuously verbalized what came to their attention while simulating the drive. Renge found that 79 percent of the verbal reports mentioned driving-related objects (e.g., other vehicles, TCDs, the road). About 11 percent of the reports were of non-driving-related objects (e.g., trees, advertising), and 9 percent were non-visual (e.g., how the vehicle was operating).

Cole and Hughes attempted to improve upon the instructions used by Renge. Participants were asked to only report objects and things to which they attended. Participants were required neither to maintain continuous verbalization nor to mimic driving through the film. Hughes and Cole had a group of participants drive the route shown on the film and verbalize objects that were attended to.⁽¹⁰⁾ Overall, possibly as a result of the changes in instructions, participants in the Cole and Hughes study who watched the film reported a greater percentage of non-driving objects than participants in the Renge study. Further, when looking at both on-road and video-based reports, Hughes and Cole found that only about 50 percent of reports were driving-related (e.g., vehicles, pedestrians, TCDs) in residential areas, but the percentage of driving-related reports went up to about 65 percent in busy commercial areas. The percentage of driving-related reports was about 5-10 percent higher when participants drove the actual route than when they viewed the route on 16 mm film. Although there were somewhat more non-driving-related reports in the laboratory than in the field, the pattern of reporting was similar. Across three road classes, the percent difference in frequency of road-related to non-road-related remained about the same. Hughes and Cole concluded that laboratory studies of what drivers attend to could yield valid data for the prediction of real-world driver behavior. Thus, Hughes and Cole and Renge showed similar findings, but Hughes and Cole got more non‑driving-related verbalizations.

In another variation of the verbalization method, Cole and Jenkins presented static roadway scenes to participants.⁽²⁾ All photographs were taken from an Australian driver's viewpoint in the left lane and contained a target TCD 328 ft (100 m) ahead. A central fixation point was present prior to each tachistoscopic slide presentation. Each photograph was presented for 0.5 s, a duration that was assumed to allow participants only one saccade. As a result, the fixation that followed the saccade would be to the most conspicuous object. It was also assumed that the most conspicuous object would be reported first. Participants were instructed to report as many TCDs as possible in each of the photographs. After the 0.5-s viewing, a light was switched on so that participants could provide a written response. The numbers of both the identified target and non-target TCDs were recorded. On average, participants reported 1.24 objects per slide. Target stop signs were reported 62 percent of the time, but yield signs were reported only 1 percent of the time. Crossroads ahead warning signs were reported 91 percent of the time, and traffic signals were reported 51 percent of the time. When the experiment was replicated with the slide exposure time tripled to 1.5 s, the number of objects reported increased to an average of 1.68 objects per slide. Cole and Jenkins speculated that the reason the longer observation time did not result in a substantial increase in the number of TCDs reported, despite an average of four TCDs per slide, was that the TCDs were not sufficiently conspicuous to elicit foveal fixations and recognition. The authors did not discuss the possibility that 1.5 s might not be sufficient time to enable memory consolidation. That is, the act of providing written recall could have been sufficient to allow visual short-term memory to decay so that items not immediately recalled could not be rehearsed and thus would be unavailable for written recall. The authors also did not consider that the result might be related to the perceptual blink phenomenon reported by Shapiro.⁽²⁸⁾ In Shapiro's study, letter stimuli were presented one at a time at a rate of 10 letters per second. There were two target letters in the stream of letters. When the second target in the sequence was within half a second of the first, the second target was often not detected. Shapiro suggested that the sequential presentation of letter stimuli is analogous to perception of similar stimuli in a sequence of saccades. If this suggestion is correct, then the detection of a second TCD following the detection of a first TCD might be suppressed if the glance to the second TCD came immediately after the glance to the first TCD.

Verbal reports have the advantage of being relatively easy to obtain. If a driver reports attending to an object and that object is present, it is safe to say the driver attended to that object. The frequency of verbal reports of a TCD may be a valid way to assess the relative effectiveness of conspicuity treatments. However, it cannot be assumed that all objects attended to will be reported. Because drivers can shift visual focus about 3 times per second, it is unlikely that all objects that are viewed foveally could be verbally reported. Furthermore, the simple act of reporting one TCD may cause drivers to forget other objects. Especially for highly learned tasks, it is possible that much of what is viewed and acted upon may not be consciously processed, making it difficult to verbalize that information. For example, because lane maintenance is such an automatic process for experienced drivers, an effective lane marking conspicuity enhancement might not be noted by drivers even if driving performance is affected.

The verbal report method may be useful, and its application in a laboratory task may be efficient. However, it is not sufficient by itself to address all issues that relate conspicuity to the effect of roadway messaging on driver behavior. A method that is similar to the verbalization of attention method is to ask drivers to search for and report specific targets.

Target Search

Cole and Hughes compared the verbalization technique to directed search.⁽⁸⁾ The authors refer to conspicuity measured using verbalization as attention conspicuity. Conspicuity measured by directed search, a technique in which drivers are instructed to report particular targets, is referred to as search conspicuity. The investigators placed 35 circular disks at locations along a 13.6-mi (21.9-km) route near Melbourne, Australia. Each disk was white, gray, or black. The white disks were in three diameters: 11.8, 19.7, and 27.6 inches (30, 50, and 70 cm). The gray and black disks were 19.7 inches (50 cm) in diameter. There were two groups of drivers. One group was given verbalization instructions; they were asked to report any objects that attracted their attention. The other group was given directed search instructions; they were asked to report all traffic signs and disk targets that they saw.

Only the proportion of disks reported was analyzed. In the attention conspicuity condition, 20 percent of the disks were reported. In the search conspicuity condition, 63 percent of disks were reported. Not surprisingly, in both reporting conditions, the proportion of disks reported increased with disk size. Disks tended to be reported when they were within 15° of the direction of travel and more than 160 ft (50 m) ahead. That is, the disks that were reported tended to be in areas where attentive drivers would be expected to look. The most striking finding was that the report rate was twice as high in residential areas, where search conspicuity was about 80 percent, than in commercial areas, where search conspicuity was about 40 percent. The researchers suggested that the lower conspicuity in commercial areas was the result of greater visual clutter, although the authors could not rule out differences in driving demands related to traffic and pedestrians as a cause of the differences in conspicuity.

Whether or not drivers look at signs may be related to conspicuity. Certainly, if a message is inconspicuous it can escape detection. If a message is important to the driver and it is conspicuous, highway engineers hope that it will be detected and looked at.

Eye Tracking

Eye trackers have been available for on-road research since the 1970s. In one study, drivers were given verbal instructions to follow a route with which they were not familiar.⁽²⁹⁾ Because of the nature of the navigation instructions, street names and other navigation signage were important to these drivers. As a result, drivers fixated on over 90 percent of traffic signals and about 80 percent of navigation signs. Drivers fixated less often on speed limit signs (about 60 percent) and even less on lane restriction and warning signs (about 50 percent). The investigators did not discuss the circumstances in which 10 percent of traffic signals were not fixated on by drivers or the criticality of the navigation signs that were not fixated on by drivers. Thus, it is not possible to determine from the report whether the signs that were not fixated on were deemed unimportant, missed because they were inconspicuous, or seen but not fixated on. Because no red-light violations were reported, one can likely infer that drivers are able to interpret some signs and signals without fixating on them or that drivers rely on other cues such as the behavior of other vehicles.

In a 1973 eye-tracking study reported by Bhise and Rockwell, drivers were instructed where to enter and exit a freeway but otherwise to drive as they normally would. Eye glance data were recorded while the participants drove on interstate highways.⁽³⁰⁾ The study focused on characterizing eye glances to navigation signs. The researchers reported that drivers moderated glance duration in a rational way. Signs with a lot of information received more glances but did not receive longer glances. Glance durations tended to be 1.5 s or less. The authors proposed a system for evaluating signs that takes into consideration legibility time-distance (based on individual driver visual acuity), first glance time-distance, and the distance at which the sign can no longer be read (based on rate of change in angle). According to this system, better signs use less of the total available time between first glance time and when the sign can no longer be read. The researchers pointed out that sign reading is shared with required driving tasks, so the time needed to read navigation signs must also accommodate traffic conditions, roadway geometry, and weather.

In a more recent study, Luoma used an eye tracker to examine driver glances to a limited selection of TCDs and advertising signs.⁽³⁾ In the analysis, eye fixation data were combined with memory for the selected objects. Participants were asked to drive as they normally would on a 31-mi (50-km) route. They were not informed of what objects were of interest beforehand. After an object was passed, participants were queried to determine if they had perceived the object. Perception was defined as correctly recalling something about the specified object. The results of the eye-tracking and perception queries are presented in table 3, which shows whether an object was fixated on or not and whether it was correctly recalled or not. The mean fixation times for TCDs that were subsequently recalled ranged from 644 ms for a speed limit sign to 130 ms for the location where crosswalk lines were not painted. The mean fixation time on billboards for which the subject of the advertisement was correctly recalled was 2,310 ms. Mean fixation time for billboards for which advertisements were not recalled ranged from 185 ms to 444 ms. Perhaps the most important message from the Luoma study is that fixation on an object and awareness of that object are not interchangeable. TCDs that are not fixated on by drivers may be recalled and those that are fixated on may not be recalled even moments after the fixation. In the case of the pedestrian crossing warning sign that was fixated on by participants, the majority of drivers showed no verbal awareness of it. Thus, fixation may not be sufficient to ensure awareness of signs, at least when verbalization is the metric used to validate awareness.

Table 3 . Percent fixated on and recalled for selected objects in Luoma study.⁽³⁾

Memory for Signs

Memory for signs has been used as a measure of conspicuity. The largest and most cited study of memory for traffic signs is by Johansson and Backlund.⁽³¹⁾ In this Swedish study, more than 5,000 drivers were stopped soon after passing a traffic sign. Because of a curve in the road, the traffic blockade that required drivers to stop was not visible until after the sign had been passed. Six signs were tested at the same location: a speed limit sign and five different warning signs.

When drivers were stopped, they were asked, "What was the last road sign you passed?" If the first question was not answered, drivers were asked, "Do you remember if the last road sign you passed was [sign name]?" The authors estimated that 60-70 s elapsed between passing the sign and these questions. The percentages of correct responses (i.e., the answer to either the first or second question was correct) are shown in table 4.

Table 4 . Percent correct recall of signs from Johansson and Backlund.⁽³¹⁾

Based on these findings, the investigators drew the conclusion that the road sign system does not achieve its purpose-that highway signs often fail to communicate their message to drivers. However, the authors also proposed that the probability of a sign being "registered" with a driver is, in part, determined by its importance to the driver. That is, the investigators proposed that the speed limit sign and the presence of police controlling the road are registered because they are important to drivers, whereas drivers on rural highways do not view nonspecific warnings or warnings of a pedestrian crossing as important. The researchers asserted that pedestrians are rare in the environment where the test was conducted and that pedestrians in this environment typically yield to vehicles rather than the reverse. Such consideration might make a crosswalk warning sign irrelevant to most drivers. How drivers distinguish important and unimportant warning signs without noticing and evaluating them was not discussed. However, the investigators' analysis implies that "registering" a TCD occurs after a driver detects and analyzes the sign for its importance.

A similar motorist intercept study was reported by Shinar and Drory.⁽³²⁾ The study was conducted in Israel, and vehicles were stopped at an army checkpoint. The checkpoint was 56 mi (90 km) from the nearest town. Traffic in both directions was intercepted. The experimental signs were placed 658 ft (200 m) before the checkpoint, which became visible about the same time as the signs. Two experimental signs were used: a warning for an intersection ahead and a general warning. In addition, there were preexisting warning signs 1,300 ft (400 m) before the checkpoint, one on the northbound approach and one on the southbound approach. These were a stop ahead warning and a winding road warning. Drivers traveling in each direction were asked to recall the last two signs they had passed. It was hypothesized that the signs would be better recalled at night because during the day drivers could see the actual hazard at about the same time they could see the warning signs.

The investigators' prediction was confirmed: drivers recalled the warning signs 4.5 percent of the time during the day and 16.5 percent at night. However, it is unclear whether the effect of time of day was the result of the content carrying more weight at night or if the signs were simply more conspicuous. After all, roadway signs mostly consist of retroreflective material. When these signs are illuminated with headlights, a large contrast is created between the bright sign and the dark, visually stark environment. As a result, it is important to determine if the time of day effect is because many warning signs are more useful at night or because these signs are more conspicuous at night. Nonetheless, the Shinar and Drory study confirmed Johansson and Backlund's conclusion that warning signs are not well remembered by drivers.

Luoma compared the validity of the intercept/recall method of measuring conspicuity with the eye glance tracking method.⁽³³⁾ In the study, recall was tested following a minimum delay (about 2 s) and following a delay comparable to that earlier intercept studies (mean = 49 s). A speed limit sign and an animal crossing warning were used. All signs were fixated on an average of 2.9 times, with a mean duration of 484 ms for each fixation. Independent of immediate or delayed recall, 94 percent of participants recalled the speed limit sign. However, recall of the animal crossing sign was less: 71 percent with immediate recall and 31 percent with delayed recall. This result appears to support the findings of Johansson and Backlund that signs regarded as important by drivers are more likely to be recalled. However, the results also suggest that recall of signs may not be indicative of sign conspicuity or perception. Luoma also measured speed changes in proximity to the signs. The speed limit sign was associated with substantial speed reduction (about 6 mi/h (10 km/h)) whether or not the speed limit sign was recalled. A much smaller speed reduction occurred with the animal crossing warning (about 1.2 mi/h (2 km/h)) whether or not the sign was recalled. There was no speed reduction in the control condition (no sign) at the same location.

Luoma's study cast doubt on all measures of conspicuity discussed to this point. Drivers may fixate on signs multiple times but not recall them. Furthermore, signs may result in measurable driver control responses, suggesting they were perceived at some level, yet drivers may not be able to verbalize the stimulus that triggered the responses. Luoma did not assess how long drivers maintained the lower speed upon passing a sign that they could not recall. Nor was he able to assess whether drivers who slowed in response to the animal warning sign would be more likely to detect an animal presence. It is also unclear if increasing the conspicuity of these signs to the point where they would be recalled would increase their effectiveness. These unknowns are candidates for future research.

Some research shows that signs are more conspicuous in environments with less clutter (e.g., residential areas, rural roads) than in highly built areas (e.g., business districts and industrial areas).^(8,34) Obviously, signs are necessary in both low- and high-clutter areas. As such, guidance is needed on how to maximize the conspicuity of individual signs and to ensure that necessary messages are communicated to drivers. Increased size, careful selection of sign location, and added visual cues are all potentially effective strategies that have been used in isolated instances. The literature reviewed suggests that measures such as eye glance fixation, recall, and verbalization are all related to conspicuity, but none of the measures unambiguously indicate whether a TCD is conspicuous and correctly influences driver behavior.

Like the Luoma study, the present study employed an eye tracker and asked participants to recall TCDs within seconds after they were passed. In addition, a small number of participants were not asked about the TCDs so that the effect of asking for recall could be evaluated. The focus of the present study was speed limit and warning signs. The objectives of the study were as follows:

• Replicate the Luoma studies in an urban/suburban setting in the United States. (The Luoma studies were conducted in rural Finland.)
• Assess the influence of recall requests on glance behavior, a factor Luoma did not address.
• Assess whether trends in glance and recall could be related to the environment in which the sign is located.

The demands for signing in urban areas are quite different from those in rural areas, which makes the extension of earlier study findings to urban areas of interest.^(35,36) The assessment of the effects of prompting drivers to recall signs could influence the design of future studies in which eye glances to signs are measured. If environmental influences on TCD conspicuity can be identified, guidance on TCD design and placement could be more specific.

Method

Participants drove a 34-mi (55-km) route that consisted of arterial roadways. These included two-lane roads and four-lane divided and undivided roads. A small early portion of the route was rural in character. The remainder of the route included residential, commercial, urban, and suburban areas.

At 21 locations along the route, participants were queried about a recently passed TCD. Generally, the queries came within 2 s of passing the TCD. In the case of crosswalks, the query came after the crosswalk in question could no longer be viewed in the rearview or side-view mirrors. The participants' gaze direction was obtained with a dashboard-mounted eye-tracking system. This approach enabled the researchers to correlate gazes to TCDs with subsequent recall of those TCDs. In addition, an attempt was made to characterize the visual environment of each TCD.

Participants

Prior to participation in the study, volunteers provided consent for motor vehicle departments in their States of residence to release their driving records. Only drivers with no recent violations on their records were asked to participate further. Twenty-six licensed drivers from the Washington, DC, metropolitan area (Maryland, northern Virginia, and the District of Columbia) provided usable data. Seven other drivers were recruited but did not complete the study either because they could not be calibrated to the eye tracker or because of data loss during the drive. All volunteers had at least 20/30 vision in both eyes as assessed with a Snellen eye chart. Of the 26 drivers that provided usable data, 10 were male and 16 were female. The mean age of the male drivers was 38 years (range 23-67), and the mean age of the female drivers was 36 years (range 23-55).

Equipment

The route was driven in a 2007 Jeep Grand Cherokee equipped with a Smart Eye eye-tracking system.⁽³⁷⁾ The eye-tracking system is vehicle-mounted (not head-mounted) and non-invasive to the driver. The system uses two infrared light-flashers and three dash-mounted infrared cameras to detect head and eye position. Eye position was sampled at 60 Hz. The dashboard installation is shown in figure 16. Three additional forward-view cameras were mounted on the vehicle roof directly above the driver's head. The view from these cameras was digitally recorded so that the direction of the driver's gaze could be overlaid in postprocessing on a 76° view of the road ahead. The external cameras recorded at 25 Hz.

Figure 16 . Photo. Eye-tracking system with three dash-mounted cameras and two infrared flashers.

Procedures

After informed consent was obtained and visual screening was completed, participants were shown a map of the route and asked to rate their familiarity with each of 10 route segments on a 5-point scale, with 1 indicating "not at all familiar" and 5 indicating "very familiar." Participants were also shown pictures of regulatory signs (black and white signs), warning signs (yellow diamond signs), navigation signs (green and blue signs) and informational signs (blue signs). They were told that the researchers were interested in what drivers looked at while driving and that they might be asked questions about regulatory, warning, navigation, and informational signage, among other things. All participants were given these instructions; however, 9 of the 26 participants were not subsequently asked to recall any TCDs. These nine participants served as a control group to enable assessment of the effect of the recall requests on glance behavior. The primary goal of this on-road study was to observe the relationship between glances to TCDs and recall of those TCDs. The possibility that the limited number of queries to TCDs would cause participants to pay more attention to all TCDs was of secondary interest; this is why recall was requested of the first 17 participants. After it was determined that adequate data were obtained for the comparison of glance behavior with recall behavior, nine participants were asked to drive the same route without requests for recall so that the possibility of a priming effect could be assessed. Nine participants were sufficient to show a significant priming effect.

Participants were seated in the research vehicle so that the eye-tracking system could be calibrated. Calibration took 15-30 min. If the calibration was successful, the drive began, and the participant was reminded of the overall route to be taken.

During the drive, a researcher in the front seat provided turn-by-turn route guidance and queried the participant at the 21 locations listed in table 5. There were two planned stops along the route where participants could stretch their legs and the eye tracker was recalibrated.

Table 5 lists the TCDs for which glances and recall were tabulated. The TCDs of primary interest were speed limit and warning signs. When recall of speed limit signs was requested from participants, they were asked, "What is the speed limit?" Speed limit signs were only queried at locations where the speed limit had changed or where the speed limit sign was the first to be encountered on the current road. This was done to reduce uncertainty as to whether correct recall was of the last sign passed or some previous speed limit sign that the driver may have thought was the last sign passed (participants were not told that the signs being queried had been passed in the preceding 2 s). The roadway locations where queries were made did not differ markedly from the immediately preceding roadways so that contextual cues to the change in speed limit would be minimized.

Table 5. TCDs for which recall was requested and glances were tabulated.

The request for recall of warning signs was, "What was the last warning sign that you passed?" Any response that indicated the target sign was recalled was accepted. This criterion was particularly important with respect to the slippery when wet warning as most participants did not know the intended meaning of the sign and gave responses, scored as correct, such as "curvy road ahead" and "skidding vehicles."

Crosswalks were queried to provide a comparison with the Luoma study in which stop lines were targets, as well as to reduce potential priming of recall for speed limit and warning signs by increasing the uncertainty as to which roadway elements might be queried.⁽³³⁾ The request for crosswalk identification was, "Was there a marked crosswalk at the previous intersection?" If the answer was affirmative, the driver was asked to describe the crosswalk. Responses were scored as correct if the type of lines (e.g., transverse, longitudinal, ladder) were correctly described.

Identification of two roadway name signs and the one informational sign with the location of a Health Department office were also requested. The request for street names was, "What was the name of the street just passed?" Responses were scored as correct if they approximated the correct street name. For example, "Hayward" was scored as correct identification of "Harvard."

A researcher in the back seat of the vehicle monitored the quality of the eye tracking. On some occasions, the researcher would request that the driver find a safe place to stop so that the eye tracker could be recalibrated. Following the initial calibration, recalibration was a brief procedure that required only about 1 min.

Following completion of the drive, participants were debriefed and reimbursed for their time.

Results

Eye-Tracking Data Reduction

The eye glance data were analyzed with MAPPS™ software.⁽³⁸⁾ With this software, an analyst marked regions of interest (ROIs) on the road-ahead video captured from the roof-mounted cameras. Each target TCD was marked with an ROI from a point 240 ft (72 m) upstream of the TCD until the TCD passed out of video view. The marking was done with a drawing tool on individual video frames. Only a small fraction of the video frames needed to be marked because the software interpolated the ROI area on intervening frames. The analyst then verified the interpolation and inserted corrections as necessary. The ROIs around signs were drawn such that they included an area the width of the sign on each sign side. An example of an ROI is shown in figure 17. At a distance of 85 ft (26 m), the ROI halo surrounding a TCD represented approximately 2° of visual angle. In a moving vehicle environment, the accuracy of the eye-tracking system is limited to a radius of about 2° of visual angle. Thus, ROIs were drawn so as to capture glances within this margin of error about the TCD. In the case of crosswalks, this approach was problematic because vehicles ahead often occluded large parts of the crosswalk. Therefore, visible areas of crosswalks or visible areas where a crosswalk would be marked were circumscribed without including an additional border. The center of the green circle in figure 17 indicates the calculated gaze location for that frame. The analysis software used 60‑Hz data, and the green circle location was calculated at 25 Hz. The green circle was not used in the data reduction process.

Figure 17 . Photo. Analysis software ROI around a speed limit sign.

Because of the challenges of recording eye movements in a dynamic vehicle environment and because the eye tracker used in this study was limited to 60-Hz sampling, no attempt was made to identify eye fixations, saccades, or smooth pursuit movements. Rather, the location of direction of gaze in each 60-Hz sample was scored as either within or outside a marked ROI. If the sum of 60-Hz gazes (each frame representing 0.0167 s) on an ROI (whether continuous or not) exceeded 100 ms, then a look to the TCD within the ROI was recorded. Models of eye movement generally define five types of movement: saccades, microsaccades, fixations, smooth pursuits, and nystagmus. The many issues regarding the measurement of these movements are beyond the scope of the present report. Several books that deal with these issues are available, including one by Duchowski and another by Holmqvist et al.^(39,40) The look or glance referred to in this report is the sum of the times between gaze-coordinate samples that fell within an ROI. This sum may have included portions of one or more fixations and the saccades between those fixations. As Holmqvist et al. point out, gaze locations represent points within a 0.0167-s interval, not the entire interval.⁽⁴⁰⁾ That is, any fixation may have started or ended any time within 0.0167 s. Most looks recorded in this study were the result of contiguous 0.0167-s intervals, and those that were not were separated by either missing data or coordinate locations that did not last more than two records (0.0334 s). Therefore, it is highly likely that these looks represented one or more physiological fixations. The minimum duration of fixations in visual tasks is open to some theoretical dispute and is likely task-related. Holmqvist et al. argue that the minimum duration is at least as small as 30 ms.⁽⁴⁰⁾

Speed limit signs 6 and 7 in table 5 were actually a series of two signs. The ROIs for these signs moved from the first to the second sign as the first sign passed from view. In the case where a single ROI consisted of two signs, a look was recorded if the total of gazes to either or both signs exceeded the 100-ms criterion. Thus, these pairs of signs were treated as a single TCD in the glance and recall analyses. Figure 18 shows the ROI for the second speed limit sign that was paired with the sign depicted in figure 17. Speed limit sign 10 in table 5 also represents a series of two signs. However, the first of the pair was occluded by foliage so that it was not visible from 240 ft (73 m) upstream. Therefore, the upstream sign in that pair was marked as an ROI as soon as it became visible in the recorded video. Subsequent analysis revealed that this upstream sign never garnered a look. It was necessary to score these two pairs of signs as if they were one sign because the second sign of the pair was legible when the first sign was passed. Thus, the sign recall query would not represent a sign recall request until the second sign was out of view.

Figure 18 . Photo. ROI shift from the first to the second sign in a sequence of speed limit signs.

TCD Recall

Initial analyses examined sign identification recall and included only those participants who were asked to identify the signs.

To enable easy comparison to the Luoma findings, table 6 shows the number of times that drivers correctly and incorrectly identified a TCD they had just passed as a function of whether the TCD had received a look.⁽³³⁾ Because some participants may have been familiar with some parts of the 34-mi (55-km) route, the sections that follow consider each TCD type as a function of familiarity as well as whether drivers looked at the TCD.

Table 6. Percent of TCDs of each type correctly and incorrectly identified as a function of whether drivers looked at them.

As in the Luoma study, there were a substantial number of cases of TCDs that were identified in the absence of looks as well as a substantial number of TCDs that were looked at but not correctly recalled only a few seconds later.⁽³³⁾ A look was defined as an eye glance dwelling on an ROI for a minimum of 100 ms. As discussed later in this report, it is possible to read TCD text and recognize symbols with near peripheral vision.⁽⁴¹⁾ Therefore, it should not be assumed that TCDs that were not looked at by the current definition of a look were not seen or read. Of the 118 opportunities in the present study where warning and speed limit signs did not receive looks, 56 percent nonetheless resulted in correct sign identification. Of the 119 occasions where speed limit and warning signs received looks, 21 percent resulted in recall failure.

A generalized estimating equations (GEE) model was used to perform the statistical analyses discussed in the following sections. This type of model is appropriate for analysis of binomial data with repeated measures. Correct sign identification (0 or 1) was the binary response measure, and driver was the subject variable. For all analyses, the repeated measure was individual TCDs on the route. The Chi-square statistic is used to evaluate significance in GEE models.

Warning Signs

A preliminary analysis was conducted to determine whether there was a difference in warning sign recall as a function of sign type (i.e., a difference between text and symbol warning signs). Variables in this preliminary analysis were whether the sign received a look, whether the drivers rated themselves familiar with the section of road where the sign was posted, and whether the sign contained text or a symbol. In this and all subsequent analyses, familiarity was reduced from five levels to two. Familiarity ratings of 1 or 2 were reclassified as unfamiliar (0) and ratings of 3, 4, or 5 were reclassified as familiar (1). Recall did not differ as a function of warning sign type, X²(1) = 0.01, p = 0.94, nor were there any interactions with sign type. Therefore, the distinction between text and symbol warning signs was dropped in subsequent analyses.

Speed Limit and Warning Sign Recall

A GEE model was run in which the independent variables were sign type (speed limit or warning), whether the driver looked at the sign (yes or no), and whether the driver was familiar with the road segment on which the sign was located (yes or no). One of the 17 drivers was excluded from this analysis because familiarity ratings were not obtained from that driver. The statistical model was a full factorial (i.e., it included all three main effects and all interactions among the variables).

Warning signs were correctly identified about a third of the time when no looks to the signs were recorded. This was true despite the rather large area (about 2°) around each sign that was scored as a look to the sign.

The most striking finding was an interaction effect in which sign identification was good for speed limits regardless of whether the speed limit sign received a look, whereas warning sign recall was highly dependent on the sign receiving a look. This effect is shown in figure 19, where warning sign results are displayed in yellow and speed limit sign results are displayed in white. When drivers looked at the TCD, the identification difference between speed limit and warning signs was not significant (p = 0.62).

Figure 19 . Graph. Probability of identification as a function of sign type and receiving a look.

Driver familiarity with the roadway segment had an interesting effect on sign recall. Recall of speed limit signs was reasonably high regardless of rated familiarity. However, recall of warning signs was better when drivers were unfamiliar with the roadway segment. This led to a two-way interaction between familiarity and sign type, X²(1) = 4.0, p = 0.046, as shown in figure 20. This finding seems reasonable if drivers who are familiar with a roadway feel that they are familiar with the hazards and do not need to attend to warning signs. Drivers who are not familiar with the road may value hazard information more and therefore attend to these warning signs, thus facilitating later recall.

Figure 20 . Graph. Probability of identification as a function of sign type and rated familiarity.

A notable finding not specifically related to conspicuity or glance behavior was that the slippery when wet sign, shown in figure 21, was not well comprehended by the participants. Of the 17 participants who were asked to identify this sign after it was passed, 12 "correctly" described it, but 8 of those 12 used descriptions such as "curvy road ahead" or "skidding vehicles ahead." Of the four participants who used the wording "slippery when wet," two were civil engineering interns who were likely to have had exposure to the sign in the course of their academic training. Table 7 summarizes identification response frequencies for the slippery when wet warning sign.

Figure 21 . Illustration. Slippery when wet warning sign.

Table 7 . Identification responses to the slippery when wet warning sign.

Probability of Look to Signs

Speed Limit and Warning Signs

Asking drivers to identify or recall a sign they just passed has the potential to prime drivers to search for highway signs. To assess the extent to which this type of priming affected glance behavior, nine participants drove the route but were not asked to identify signs. The probability of a glance to the target signs was then compared to the same probability from the 16 drivers that were asked to identify signs and had provided segment familiarity ratings.

Whether drivers were asked to identify signs (yes or no), familiarity with the road segment (yes or no), and the type of sign (warning or speed limit) were included as independent variables in a GEE analysis. The probability of a look to the sign as a function of the independent variables is shown in figure 22.

Figure 22 . Graph. Probability of a look to a sign as a function of familiarity, sign type, and whether the driver was asked to identify signs.

There was a substantial and statistically significant priming effect, X²(1) = 8.2, p < 0.01. Overall, participants were about twice as likely to look at a target sign when there was a possibility they might be asked to identify the sign. Nonetheless, the mean probability of looking at a target sign never exceeded 0.44. Priming was not consistent across conditions; there was a significant three-way interaction, X²(1) = 4.8, p = 0.03. Drivers who were not being queried but were unfamiliar with a segment of road were about as likely to look at warning signs as drivers who were being queried. However, in the case of speed limit signs, a lack of familiarity was not associated with an increase in the probability of a look. Participants who were not asked to identify signs looked at about 20 percent of the speed limit signs.

When speed limits are reduced, many agencies repeat speed limit signs at relatively short distances, substantially less than the distance normally specified for separation where there is no change or an increase in speed limit.⁽⁴²⁾ The probability that the speed limit will be looked at and remembered was examined as a function of whether there was one speed limit sign or a sequence of two signs. In two areas where the 25-mi/h speed limit sign was repeated within 500 ft (153 m), the two signs were treated as one sign when scoring looks and recall; identification was requested after the second sign in the pair was passed. Looks and identification in those two areas were compared to two areas with 35-mi/h speed limit signs where looks and identification were scored for only one sign. The combined probability of a look to the speed limit increased when there were two signs, X²(1) = 7.1, p < 0.01. With two signs, the probability of a look to at least one sign was 0.43, whereas with one sign, the probability of a look was 0.25. This is exactly the same result that would be obtained if the probability of a glance to either 25-mi/h sign was the same as the probability of a look to a 35-mi/h sign (i.e., (0.25 + 0.25) - 0.25² = 0.437). As in the previous analysis, the identification was high whether or not speed limit signs received looks. Mean probability of identification for the sign pair was 0.84 and for a single sign was 0.95. This difference was not significant (p = 0.29).

Crosswalks

The mean proportion of correct identification of whether crosswalks were marked or unmarked was 0.49. There were no significant differences in identification as a function of looks, marking type, or familiarity with the roadway segment.

Street Names

Because the street names that were queried were at cross streets where no navigation decisions were required, there was no explicit reason for drivers to attend to the street name signs. Correct identification of the street names varied only as a function of whether drivers looked at the sign, X²(1) = 4.4, p = 0.04. If the sign received a look, identification probability was 0.42, whereas if the sign did not receive a look, identification probability was only 0.09.

Discussion

There were numerous cases where drivers looked at TCDs yet failed to identify the TCD and many cases where TCDs were correctly identified in the absence of a look. This suggests that conspicuity measures that rely on either glance or recall alone are not adequate for evaluating the attention-getting qualities of TCDs.

Although 80 percent correct recall by drivers who glanced at speed limit and warning signs is impressive, unfamiliar drivers looked at only 35 percent of warning signs when not being queried and only slightly more when they might be asked to identify the signs. Overall, warning sign recall by both drivers familiar and unfamiliar with the roadway was about 50 percent. If it is assumed that recall of a warning sign reflects a driver's need for the sign information, as argued by some researchers, then warning signs appear to have more perceived utility to drivers who are unfamiliar with roadway segments.^(31,32)

The reason that warning signs were infrequently looked at or remembered cannot be determined from the present data. It could be that drivers were relying on other elements of roadway context and seldom felt the need to attend to warning signs. Alternatively, it could be that warning signs are difficult to attend to because of other demands on driver attention. This study cannot distinguish between these two circumstances or other explanations for failure to recall TCDs.

Drivers in the study were able to correctly identify the posted speed limit about 80 percent of the time regardless of whether the drivers looked at the speed limit sign or were familiar with the segment of roadway. There are at least two possibilities for this finding. One is that drivers do not need to fixate on speed limit signs to read them. Another possibility is that drivers are skilled at inferring speed limits from roadway context. The present research cannot distinguish between these two explanations. However, the sign identification conspicuity results presented later in this report suggest that speed limit signs can be read with 80 percent accuracy when the point of fixation is 9° away from the center of the sign and the sign subtends 2° of visual angle. Thus, there is evidence from a non-driving task that speed limits can be read in the absence of fixation.

Other studies have also reported that the slippery when wet warning sign is not well comprehended. Dewar, Kline, and Swanson found 44.6 percent recognition of this sign among drivers interviewed in Texas, Idaho, and Alberta, Canada.⁽⁴³⁾ Charlton reported similar comprehension difficulties with a similar international symbol sign.⁽⁴⁴⁾ A study conducted by the Texas Transportation Institute found only 62.5 percent comprehension of the slippery when wet sign among 747 participants who took a multiple-choice test.⁽⁴⁵⁾

Participants in the present study were prompted to identify only a small proportion of signs along the 34-mi (55-km) route. This was done to minimize priming participants to pay special attention to signs. The selective sampling of signs for recall was not completely successful, as participants who were never prompted to identify signs were about half as likely to look at the target signs as participants who were prompted to identify signs. It is interesting that even when participants expected to be asked to identify signs, they glanced at less than half of the target TCDs. Although there were other demands on the driver's attention (e.g., other traffic, curves), the signs were visible to the driver for most or all of the last 240-ft (73.2-m) approach to the sign in all cases. In no case were the driving conditions such that safe opportunities to look at the signs were unavailable. If it is assumed that most, if not all, TCDs included in this study were intended to capture driver looks, then it might be concluded that these signs failed between 65 and 85 percent of the time. The present data do not warrant that conclusion. Only one of the warning signs indicated an existing hazard (i.e., parked vehicle ahead). The remaining warning signs were for occasional hazards that are usually not present (e.g., deer, ice on bridge). In these cases, the drivers may have realized that the warning was not intended for them and therefore elected not to look at or process the warning for future recall. Furthermore, Luoma has provided evidence that drivers respond to warnings by slowing even though the nature of the warning is not available for immediate recall.⁽³³⁾ The present data indicate that sign effectiveness evaluations need to go beyond glance and recall methodologies.

However, drivers in this study may have failed to look at TCDs because the TCDs lacked sufficient conspicuity to allow them to be reliably detected. The next study in this report specifically addresses the detectability of signs by measuring how far a person can focus his/her gaze away from a sign and still detect its presence.