Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
110 Chapter 10. Safety Framework to Assess the Impact of Highway Signs on Distraction Using Simulator Data 10.1 Introduction 10.1.1 Background An important challenge arises when law enforcement reports or other naturalistic data sources are used to evaluate the distraction potential of external objects: individual events (such as crashes or NC occurrences) only offer a single snapshot of potential distraction. That is, just because one driver is distracted by a particular feature does not mean that the majority of drivers will be, and vice versa. Determining distraction potential requires a sample of many drivers, with each driver exposed to the same set of roadway features, and a precise method for measuring visual attention in relation to external objects. Driving simulator experiments that include infrastructure elements and eye tracking can potentially be used to evaluate the link between external features and driver distraction. To date, most of the experimental research on driver distraction has focused on driversâ interactions with in-vehicle devices such as cell phones and integrated interfaces. This research has demonstrated that a driverâs focus on nondriving tasks results in significant impairment to visual attention and driving performance (Engstrom et al. 2005, Strayer and Drew 2004, Kass et al. 2007, Lee et al. 2001, Klauer et al. 2006). However, the extent to which external infrastructure elements lead to similar degradation in visual attention and driving performance is unknown. To address this question, a large archival database of driving simulator data from the University of Iowaâs NADS was utilized. An analysis of eye tracking and driving performance data was conducted to determine whether evidence exists linking external roadway features to driver distraction. 10.1.2 Scope of Framework The objective of the analysis presented in this chapter was to develop a Safety Framework to evaluate the relationship between driver distraction and external roadway features using a simulator data set. The Safety Framework provides a template for conducting similar work and demonstrates the feasibility of the approach. The analysis focused on two roadway features, billboards and route signs, along a stretch of highway driving. The feasibility of evaluating distraction from other roadway features in other roadway environments was also investigated. Examples of research questions that could potentially be addressed through an analysis of simulator data include the following: ⢠Which infrastructure features cause drivers to engage in longer glances away from the forward roadway? ⢠Does the presence of certain infrastructure features negatively impact driving performance? ⢠How much does the potential for OVD vary among drivers?
111 ⢠How does OVD vary between nighttime and daytime driving? ⢠What variables (e.g., steering wheel position, glance location, glance duration) are the best predictors of OVD? ⢠How does impairment impact OVD? ⢠What is the relationship between OVD and the presence of billboards? ⢠What is the relationship between OVD and the presence of overhead signing? The analysis presented in this chapter was designed to evaluate the following research questions: ⢠How do infrastructure features impact glance behavior? ⢠Does the presence of certain infrastructure features negatively impact driving performance? 10.2 Data Sets Utilized This study utilized a large archival simulator study data set collected by the University of Iowa. This section describes the database and elements that were and were not included in the analysis. Discussions on simulator data sets in general are provided in Section 4.2. 10.2.1 Simulator Study Data Set This Safety Framework utilized an archival simulator study data set collected at the University of Iowaâs NADS. The data set was collected as part of an NHTSA study on drowsiness detection using in-vehicle data and is hereafter referred to as the IMPACT database (Brown et al. 2014). The NADS-1 driving simulator is one of the most advanced ground vehicle simulators in the world. It consists of a 24 ft dome in which an entire car is mounted (Figure 24). Figure 24. NADS-1 driving simulator (left); cab view inside NADS-1 during a nighttime driving scenario (right). All participants in the studies conducted in the NADS-1 drive the same vehicle, a 1996 Chevrolet Malibu sedan. The motion system on which the dome is mounted provides 400 m2 of horizontal and longitudinal travel and ±330 degrees of rotation. Each of the three front projectors that display imagery on the inside of the dome has a resolution of 1600 x 1200; the five rear projectors have a resolution of 1024 x 768. The edge blending between the projectors is 5 degrees horizontal. The NADS-1 produces a thorough record of vehicle state (e.g., lane position) and driver inputs (e.g., steering wheel position), sampled at 240 Hz.
112 For the study that produced the IMPACT database, the cab was equipped with a Face Lab 5.0 eye tracking system (developed by Seeing Machines, Canberra, Australia) that was mounted on the dash above the steering wheel, as shown in Figure 25. Figure 25. Face Lab 5.0 cameras mounted in the cab, with a separate head tracking system mounted between them. In a worst-case head pose scenario, the system could measure the driverâs gaze with an estimated root mean square error of 5 degrees. In the best case scenario, where the driverâs head was motionless and both eyes were visible, a fixated gaze could be measured with an estimated error of 2 degrees. The eye tracker recorded data at a rate of 60 Hz. The cab was also equipped with a Seeing Machines Driver State Sensor (DSS) V3.4.260101, a single-camera system that was used for head tracking. The camera installation is also shown in Figure 25. Data were collected in May and June 2011, under the approval of the University of Iowa IRB. Data for the IMPACT database were collected from 72 participants. Drivers were divided into equal groups by age (21 to 34, 38 to 51, and 55 to 68) and gender (Table 12). Participants were recruited from the NADS subject registry and provided written informed consent. Table 12. Simulator study demographics. Age Group N Mean Age (yrs) Min. Age (yrs) Max Age (yrs) Num. Female Young 24 26.5 21 34 12 Middle 24 44.3 38 51 12 Old 24 60.7 55 67 12 The original goal of the study was to collect data from drivers experiencing different levels of drowsiness (alert, mildly drowsy, severely drowsy) to develop models for drowsiness detection using vehicle data (see Brown et al. 2014 for more details). Participants completed three drives over two visits: one daytime drive between 9 a.m. and 1 p.m., one early nighttime drive between 10 p.m. and 2 a.m., and a late-night drive between 2 a.m. and 6 a.m. To avoid potentially confounding the effects of drowsiness and distraction, the analysis presented in this chapter only included the drives that occurred during the day, where participants were assumed to be alert.
113 Participants drove a scenario representative of a drive home from an urban area for a total drive time of approximately 35 minutes. The drive started with an urban segment composed of a two- lane roadway through a city with posted speed limits of 25 to 45 mph with signal-controlled and uncontrolled intersections. A second segment consisted of a four-lane divided expressway with a posted speed limit of 70 mph. The drive continued onto a rural segment composed of a two-lane undivided road with curves, ending with a 10-minute-long drive on a section of straight rural roadway. A map of the drive segments is shown in Figure 26. Figure 26. Map of the urban, Interstate, and rural drive segments in the simulator study. For the present analysis, only the second segment, which comprised a four-lane divided expressway, was included. This segment was selected for the following reasons: ⢠The segment contained several roadway elements for drivers to consider, including overhead route signs, overpass signs, navigation and exit signs, and billboards. ⢠Baseline epochs were available where the external features of interest along the roadway were absent. ⢠It was feasible to estimate glances toward the external roadway features in this segment. 10.2.2 Roadway Features Included in the Analysis Video data from the IMPACT database were reviewed. These video data included views showing images of the driverâs face, views of the vehicleâs footwell, over-the-shoulder views, and views of the forward channel of the simulator. This review was intended to determine a number of candidate external roadway features for inclusion in the analysis. Table 13 presents the roadway features that were included in the analysis; screenshots of these features are shown in Figure 27.
114 Table 13. External roadway features included in analysis. Feature Count Overhead sign before an overpass 1 Overpass sign 1 Route sign before the exit off the highway 1 Route sign 1 Billboards 4 Figure 27. Examples of two billboards (top), a route sign (middle left), an overhead sign (middle right), an overpass sign (bottom left), and a route sign for an exit (bottom right).
115 10.2.3 Roadway Elements Considered but Not Included As part of the video review, other segments of the drive and other external features were considered for inclusion in the analysis. Specifically, the urban segment of the drive contained many external features that might capture a driverâs attention (Figure 28). Figure 28. Potential features of interest along the urban roadway segment. Upon further investigation, however, it was determined that measuring potential distraction in the other segments of the drive and for other features was not possible for a number of reasons: ⢠It was not possible to isolate any single external feature to measure glances. ⢠No suitable baseline section of the drive was available against which to compare epochs that contained an external feature. ⢠Drivers were instructed to follow a set of auditory navigation cues, meaning that they had incentive to look away from the road toward the external environment. That is, as part of the design protocol, participants were instructed to make turns at particular roadways or locations, which meant that they need to scan the environment. 10.3 Analysis 10.3.1 Data Reduction The simulator recorded a large number of variables related to vehicle position, driver inputs, the external environment, and driver visual behavior via the in-cabin eye tracker. These data were reduced to summary measures of interest using custom MATLAB scripts. For each external roadway feature, an epoch was created consisting of the 1,000 ft preceding the roadway feature. At 65 mph, this would translate into a 10-second window leading up to the external feature. For comparison purposes, a baseline epoch was also created corresponding to each event epoch. For instance, for the billboards, the corresponding baseline epochs occurred immediately after the driver passed each respective billboard. The baseline epochs for the navigation signs were clustered on a sparse segment of highway between two billboards. The locations of the signs and billboards were obtained from the visual database, and their dimensions were used to define a bounding box for each feature. Glances within the bounding
116 box for a feature were an indicator that the driver looked at that feature. Gaze pitch and yaw angles that corresponded to glances at the feature were determined, and gaze locations were tracked to estimate the occurrence and duration of each glance. The driver approached and passed each feature at highway speeds, approximately 65 to 70 mph. During the approach, the location of the feature shifted to the right with respect to the driver (billboards also shifted up relatively), and this progression was exponential until the driver finally passed the feature. Because of this behavior, it was possible to track driversâ glances at each of the features, and tracking became easier the closer the driver came to the feature. 10.3.2 Variables Included Of primary interest for this analysis was driversâ visual behavior in the event and baseline epochs. A glance was defined as a sequence of gaze values directed at a feature that amounted to at least 0.5 seconds in a 1-second period. This definition allowed for brief fixations outside the bounding box of the feature as long as a sufficient amount of the driverâs gaze was targeted within its box. From these gaze data, two sets of measures were computed to understand different components of visual behavior as they relate to the external roadway features in this data set. The first set of measures was used to estimate the visual attention devoted to the external feature: ⢠Number of glances: Total number of glances to each feature. ⢠Average glance time: Average duration of glances to the feature. ⢠Total glance time: Total eyes-off-road time associated with each feature. The second set of measures was calculated for both the event epochs (i.e., the leadup to each feature) and baseline epochs (i.e., the period after passing each feature). The goal of these measures was to compare standard glance metrics that might speak to whether visual attention differed with and without the features present. These measures included the following: ⢠PRC gaze: Percentage of time where gaze was directed to the forward roadway within the epoch. ⢠Horizontal gaze dispersion: Horizontal dispersion of gaze (in degrees). ⢠Vertical gaze dispersion: Vertical dispersion of gaze (in degrees). In addition to measures of visual attention, measures of driving performance were also calculated to determine whether there was evidence of performance degradation in the event epochs compared to baseline epochs. These measures included the following: ⢠Standard deviation of lane position: A measure of lateral vehicle control. ⢠Standard deviation of speed: A measure of longitudinal control. ⢠Number of lane departures within the event and baseline epochs. 10.4 Modeling Approach and Results An exploratory analysis of the data consisted of two parts. The first focused on visualizing glance measures related to the external event epochs. The second compared glance and driving
117 performance measures between the event and baseline epochs. These analyses are described in detail below. 10.4.1 Glance Measures For each roadway feature, the number of glances, average glance duration, and total glance time were plotted as a way to estimate how glances in the event epochs compared to established measures of distracted glance behavior. Individual glances longer than two seconds and long periods of total eyes-off-road glance time are associated with significant increases in crash risk. Figure 29 shows a histogram of the number of glances to each roadway feature. Figure 29. Number of glances to different external features. Note that the four billboard events were aggregated together. For the most part, drivers made few glances to any of the external elements included in this analysis. Billboards received the most glances, and even then, very few drivers made more than one total glance to the billboard. Drivers glanced at billboards 40 times, which represents 11.1% of drivers. Drivers glanced at the highway exit sign eight times, which also represents 11.1% of drivers. These numbers indicate both the number of drivers (11.1%) and the total number of glances (40 glances and eight glances). Other signs received fewer glances, with 6.9% of drivers glancing at the overpass sign and 4.2% of drivers glancing at the both the overhead and route signs. Figure 30 shows the average duration of glances directed to the different external features.
118 Figure 30. Average glance time across the external features, with points representing individual drivers. Average glance duration ranged from less than 0.5 seconds to just over 1 second for each of the billboard events. Across all events, very few average glance durations exceeded the 2-second distraction threshold. This suggests that the external features included in this analysis did not, for the most part, result in glances indicative of visual distraction. The average glance time was 2.21 seconds for the overpass sign, 2.00 seconds for the highway exit sign, 1.54 seconds for the route sign, 1.37 for the billboards, and 1.20 seconds for the overhead sign. Figure 31 shows total glance time across the different external events. Figure 31. Total glance time across the external events, with points representing individual drivers.
119 Again, the point of reference for classifying distraction was long periods of eyes-off-road time. Drivers who glanced toward the external features in this data set clearly fell below this off-road glance threshold. The median total glance time was around 2 seconds, and the maximum total glance time was under 10 seconds in all cases. This further indicates that the external elements included in this analysis were not associated with significant visual distraction. The average total glance times were 3.22 seconds for the overpass sign, 3.18 seconds for the highway exit sign, 2.54 seconds for the billboards, 1.84 seconds for the route sign, and 1.51 seconds for the overhead sign. 10.4.2 Comparison of Event and Baseline Epochs The second component of the analysis compared visual behavior in the event epochs versus the baseline epochs. For each external feature, measures of glance behavior and driving performance were compared between the event and baseline epochs. Measures of PRC gaze and gaze dispersion were computed to determine whether there were differences between the two epochs. Data from these events were analyzed using linear mixed-effects models with the lme4 package in R. Linear mixed-effects models have the advantage of adjusting effects for differences in sample size and missing data. Participants and epochs were entered as random effects. To control for differences across signage, each event was entered as a fixed effect. In all cases, p-values were obtained by likelihood ratio tests comparing the full mixed-effects model to a partial model without the effect of epoch. The results of these tests are presented in Table 14. Table 14. Results (p-values) of Ï2 tests comparing models with and without epoch. Measure Billboard Overhead Sign Overpass Sign Highway Exit Sign Route Sign PRC Gaze 0.44 0.08 0.69 0.03* 0.12 Horizontal Gaze Dispersion 0.84 <0.001* 0.04* 0.01* <0.001* Vertical Gaze Dispersion 0.64 0.90 0.03* 0.09 <0.001* Standard Deviation of Lane Position 0.94 0.77 0.92 0.03* 0.46 Standard Deviation of Speed 0.64 0.18 0.22 0.08 0.07 Number of Lane Departures 0.65 0.66 0.58 0.74 0.71 * Difference was statistically significant. For the analysis of billboards, data from all four billboard events were combined. There was no evidence that either the glance measures or driving performance metrics differed between the event and baseline epochs. For the overhead sign analysis, there was no evidence of an effect of epoch on PRC gaze, but epoch had a significant impact on horizontal gaze dispersion. Gaze dispersion measures the breadth of glances to different regions of the driving scene, with a larger gaze dispersion reflecting a broader scanning of the environment. Significant reductions in gaze dispersion are commonly associated with in-vehicle distraction and suggest a reduction in visual scanning, in
120 effect âtunnel vision,â when drivers are distracted. This could lead to drivers missing critical information in the driving scene (Figure 32). Figure 32. Horizontal gaze dispersion for the overhead sign before the bypass event. There was no evidence of an effect of epoch on other gaze measures, nor was there evidence of a difference between the event and baseline epochs for any of the driving performance measures. For the overpass sign analysis, there was, again, no evidence of an effect of epoch on PRC gaze. There was a main effect of epoch on both horizontal and vertical gaze dispersion, again driven by a decrease in gaze dispersion in the event epochs compared to the baseline epochs. For the route sign analysis, both PRC gaze and horizontal gaze dispersion were significantly impacted by epoch. PRC gaze was lower in the event epochs compared to the baseline epochs (Figure 33), and horizontal gaze dispersion was higher in the event epochs compared to the baseline epochs. There was also a main effect of epoch on standard deviation in speed, with lower speed deviation in the event epochs compared to the baseline epochs.
121 Figure 33. PRC gaze for the route sign before the exit event. Finally, for the highway exit sign analysis, there was no evidence of an effect of epoch on PRC gaze. There were, again, significant main effects of epoch on both horizontal and vertical gaze dispersion. Gaze dispersion was, again, lower in the event epochs compared to the baseline epochs. 10.5 Outcomes and Discussion 10.5.1 Summary A preliminary analysis was conducted to evaluate the impact of five external roadway features on visual attention and driving performance using an archival simulator data set. The features included four billboards, a highway exit sign, an overhead sign, an overpass sign, and a route sign. Two sets of analyses were conducted on a sample of 72 drivers across different age groups. First, glances made toward the external roadway features were evaluated. The results of this analysis suggest that drivers, on average, made few glances to these roadway features. Furthermore, these glances tended to fall below established thresholds for distraction, both in terms of the average glance length and the total glance duration. Drivers were most likely to glance at the billboards and the highway exit sign. Additionally, the overpass sign received the longest average glance time of 2.21 seconds, followed by 2.0 seconds for the highway exit sign. A second set of analyses focused on comparing measures of visual attention and driving performance between epochs that contained the external roadway features and baseline epochs absent of the external roadway features. PRC gaze was statistically different for only the
122 highway exit sign. Horizontal gaze dispersion was statistically different for all of the features except billboards. Vertical gaze dispersion was statistically different for the overpass sign and route sign. No differences in standard deviation of speed were noted for any of the signs, while standard deviation of lane position was statistically different for the highway exit sign. While the differences identified in these two analyses were small, they do indicate that differences in driver behavior for different roadway features can be determined. For instance, drivers spent more time looking at and had more frequent glances toward certain features. Other metrics such as PRC gaze and gaze dispersion also showed differences between features, though the differences were small. However, this result was not unexpected because drivers may have focused on the locations of certain features to guide route planning decisions within the drive. Further, a reduction in gaze dispersion alone does not represent evidence of visual distraction. The results of this exploratory analysis suggest that the roadway features evaluated have limited distraction potential, especially when viewed in comparison with in-vehicle, visual-manual distractions such as texting on a cell phone. It should be noted that the types of features selected for evaluation were not expected to have a large distraction potential. Signs along the Interstate segment were selected because they could be isolated from other features and were available for a number of drivers. The results do indicate, however, that differences in surrogate measures of distraction can be isolated for roadway features in simulator studies, particularly if the study is designed for that purpose. 10.5.2 Performance Metrics Performance metrics suitable for measuring the success of the research activities and outputs were identified and are described below. Feasibility of the Data Sets Utilized This analysis used a reduced data set from an NHTSA study on driver drowsiness. The analysis focused on the alert (control) drives from that study. Driving simulator studies have a number of advantages for analyzing the potential impact of external roadway and other features on driver attention and performance: ⢠Repeatability across multiple individuals. ⢠Precise experimental control. ⢠High-grained data collection and reduction. ⢠Eye tracking data. The selection of the data set included in this analysis was constrained by a number of factors: ⢠Ability to isolate roadway features of interest. ⢠Availability of baseline epochs for the features. ⢠Visibility of the features. ⢠In some cases, the presence of multiple events for certain sections of the drive.
123 While the data set for this analysis ultimately included data from an Interstate segment only, the initial plan was to examine other aspects of the study drive, including an urban segment with unique signs and roadway features. For the reasons mentioned earlier in this chapter, however, it was determined that the analysis undertaken for this Safety Framework was not feasible for those sections of the drive. In future efforts, simulator drives designed specifically around the types of research questions that informed this analysis may be better suited to answer specific questions about roadway design and infrastructure elements. Another issue that should be addressed regarding the application of simulator research to identify the effects of external roadway features is that of generalization to real-world situations. The simulator used to collect the data for this study is one of the most immersive in the world. Still, future efforts are needed to validate the visibility conditions within the simulator against those of real-world environments, particularly in order to draw inferences about the absolute impact of roadway features on attention and performance. Sample Size The analysis presented here included data from 72 drivers who experienced the same drive. This sample should have been sufficient to detect any negative impacts of external roadway features on driving performance. Indeed, this sample was originally designed to isolate differences in performance due to drowsiness, another cause of driver impairment. One limitation of the present analysis was the inability to focus on differences across groups of drivers or to examine the interaction between external features and other independent variables. Some variables of interest that may be relevant and should be included in similar modeling efforts include the following: ⢠Driver experience. ⢠Time of day. ⢠Visibility conditions. ⢠Surrounding traffic congestion. ⢠Presence of in-vehicle distraction, passengers, etc. ⢠Speed limit. ⢠Roadway type. Performance of Statistical Models Because the analyses were used as proofs of concept to guide the development of the Safety Framework, sample sizes were limited to existing data that could be obtained within project resources. Due to sample size limitations, only simple metrics such as the p-value could be used to assess model performance. Additionally, due to the small sample size, the impact of covariates could not be determined. 10.5.3 Discussion The objective of this Safety Framework was to assess the efficacy of using a legacy driving simulator data set to evaluate distraction due to external roadway features. The five roadway
124 features evaluated in this analysis were all common roadway signing. The results indicate that there was some evidence, although limited, of visual distraction due to the five roadway features included in the analysis. There was also limited evidence that driving performance differed in the presence or absence of the external roadway features. Other segments of the drive and their unique roadway features could not be included in the analysis for the reasons explained earlier in the chapter. The analysis of average glance duration suggested that drivers glanced 2 or more seconds at the overpass sign and highway exit sign, longer than drivers glanced at the other highway features (route signs, billboards, and overhead signs). In the comparison of driver behavior between event and baseline epochs, PRC gaze was statistically different for only the highway exit sign, horizontal gaze dispersion was statistically different for all of the signs except billboards, and vertical gaze dispersion was statistically different for the overpass sign and route sign. No differences in standard deviation of speed were noted for any of the signs, while standard deviation of lane position was statistically different for the highway exit sign. While the differences identified in the two analyses of the simulator data set were small, they do indicate that differences in driver behavior for different roadway features can be determined through simulator studies. For instance, drivers spent more time looking at and had more frequent glances toward certain sign types. Other metrics such as PRC gaze and gaze dispersion also showed differences between signs. The advantages of driving simulator experiments make them well suited to evaluating the impact of roadway features (and other variables) on driver attention and performance. This Safety Framework can help guide future efforts to design and evaluate simulator studies assessing the impact of roadway features. A disadvantage of using driving simulator experiments is the lack of real-world consequences (e.g., crashes, citations) for drivers that commit errors or unsafe driving acts.
