Leveraging the Second Strategic Highway Research Program Naturalistic Driving Study: Examining Driver Behavior When Entering Rural High-Speed Intersections

DATA

All data used in this analysis were acquired from the SHRP2 dataset. The following sections detail the data request process, how intersections were selected for inclusion in this analysis, the contents of the static and time-series datasets, and the reduction of video data.

DATA REQUEST PROCESS

The process to request SHRP2 NDS data began with InSight, VTTI’s Web site for limited data access, message boards, and documentation. Insight’s highest user access level, qualified researcher, requires registering for the site, agreeing to the terms of service and privacy policy, and uploading a valid Institutional Review Board (IRB) training certificate. Doing so unlocks the custom query capability, which was used to assess the feasibility of the project by identifying the data components necessary for analysis of stopping and scanning behaviors.

The RID was then used to identify rural high-speed intersections.⁽⁸⁾ Geographic information system software was used to interface with the database. The RID provided detailed information on roadway features and identified each road segment with a unique link identification (ID). A set of relevant link IDs was identified and sent to VTTI in exchange for the number of crossings and unique participants who traversed them.

VTTI required three documents to establish a data sharing agreement. First, a detailed research statement was submitted, outlining the study’s objectives and proposed analyses. Concurrently, an IRB application was submitted to the principal investigator’s home institution. The application described how the study involved human subjects, the research design, expected benefits and potential risks, a risk mitigation plan, as well as how personally identifiable information (PII) would be used. Finally, a data use license (DUL) created by VTTI was filled out and submitted for review. The DUL included a project background and description, a data request scope (summary of dataset being requested and a description of how it ties into the research problem), a data specification (list of the specific data elements requested), biographies of all members of the research team, and a data security plan. All research team members and an institutional representative from the principal investigator’s home institution signed the DUL, binding them to the extensive terms and conditions. After some clarification on the data security plan, the director of VTTI then signed the DUL, thereby enacting it.

Because VTTI was handling data requests on an individual basis, a standardized pricing scheme did not exist at the time, so a subcontract was considered the correct course of action. The principal investigator’s home institution initiated the process by providing VTTI with a statement of work outlining expectations, timelines, and payments. VTTI responded with a detailed cost estimate, and the two parties agreed to a firm fixed-price subcontract with three milestones to be accomplished within 2 mo: (1) extraction of static and time-series datasets, (2) joint development of a video reduction protocol, and (3) delivery of video-reduced data (eyeglances and traffic presence).

It should be noted that the principal investigator’s home institution had two options regarding the reduction of video data. To protect PII, viewing of face videos was restricted to a facility (referred to as the “secure data enclave”) on VTTI’s campus. Reduction was accomplished entirely by VTTI staff, but the principal investigator’s home institution could have reserved time in the enclave and performed the reduction using members of the project team. The principal investigator’s home institution considered this option but ultimately chose to subcontract VTTI for the reduction because costs were comparable. Reduction by the principal investigator’s home institution staff was subject to non-completion within the constraints of the enclave time.

INTERSECTION SELECTION

Intersections were selected from the RID to be as homogeneous as possible. Through an iterative segmentation process, four Pennsylvania intersections were found to have the desired features and number of crossings sufficient for meaningful analysis. Those desired features include the following:

• A rural setting, which was determined using the U.S. Census Bureau’s 2014 TIGER/LINE® shapefiles available to the public from census.gov.
• Four approaches.
• Each approach having exactly one through lane in each direction and no turn lanes.
• A major route with a posted speed limit ≥ 50 mi/h.
• Stop-controlled minor route approaches.

After identifying the intersections of interest, a list of link IDs was sent to VTTI for extraction. In return, VTTI provided time-series data on 735 crossings through 7 intersections. However, to be useful to the analysis of intersection-approaching stopping and scanning behaviors, crossings wherein drivers began on the major route were excluded, leaving 461 relevant crossings. A total of 3 sites experienced only 5, 8 and 13 crossings, respectively. After excluding these crossings for lack of sufficient replication to test for site-specific effects and 24 others for incomplete traces (crossings originating at selected intersections), the dataset consisted of 411 crossings through 4 similar intersections. Figure 1 through figure 4 provide satellite imagery for each intersection overlaid with associated crossings.

Original image: ©2016 Google®; map annotations provided by Leidos.

Figure 1. Map. Crossings through intersection 1.⁽⁹⁾

Original image: ©2016 Google®; map annotations provided by Leidos.

Figure 2. Map. Crossings through intersection 2.⁽¹⁰⁾

Original image: ©2016 Google®; map annotations provided by Leidos.

Figure 3. Map. Crossings through intersection 3.

Original image: ©2016 Google®; map annotations provided by Leidos.

Figure 4. Map. Crossings through intersection 4.⁽¹²⁾

The 411 extracted crossings of 4 intersections were performed by 31 unique drivers. Table 1 lists the number of crossings in this final dataset by participant and intersection and highlights the unbalanced nature inherent to naturalistic data.

Table 1. Number of crossings in final dataset by participant and intersection.

STATIC AND TIME-SERIES DATA EXTRACTION

For each crossing through the selected intersections, numerous static and time-series variables were requested for their relevance to stopping and scanning behaviors.

Static Data

The static data used in this analysis consisted of variables that remained constant throughout the individual crossings. Table 2 lists these variables and their definitions. All static data elements (with the exception of maneuver) were collected via questionnaire prior to participation in the NDS.

Table 2. Static variables and definitions.

Both genders (male and female) were well represented in the data. Of the 411 crossings, 47.7 percent were made by males, and 52.1 percent were made by females, with the remaining 0.2 percent unspecified. Of the 31 participants, 41.9 percent were male, 54.8 percent were female, and 3.2 percent were unspecified.

Age has been shown to affect both braking distance and scanning patterns.^(9–11) Age group was originally quantified in 5-year increments, but because of the scarcity of the data, this was aggregated into two groups: younger (ages 16–44) and older (ages 45–84). Though drivers aged 16–19 years are likely to drive differently than any other age group, only one such participant crossed a qualifying intersection during the study period. The aggregation into two age groups divides the participants almost exactly in half, with 14 participants classified as older and 12 as younger (with two participants missing age data altogether). No effort was made to update age throughout the study because it would change by no more than 2 years, which would likely not result in any change to the binomial aggregation.

AAM was similarly aggregated from 5,000- to 10,000-mi increments. Mileage may reflect driving experience better than age, and greater experience has been shown to correlate with longer glance durations.⁽¹⁶⁾

The number of crashes was transformed from a count variable (with levels {0, 1, 2 or more} and frequencies {360, 13, 38} respectively) to a binary indicator with 1 used to indicate that the participant had experienced at least 1 crash in the prior 3 years and 0 otherwise, resulting in levels {0, 1} with respective frequencies {360, 51}.

Prior to beginning the study, participants were asked to indicate the risk they associated with performing a rolling stop. The level of risk associated with performing a rolling stop was originally captured on a 7-point scale, with 1 and 7 corresponding to “no greater risk” and “much greater risk,” respectively. These responses were aggregated to low (1–2), medium (3–5), and high (6–7). The tendency to perform a rolling stop originally had four possible responses (never, rarely, sometimes, and often) but was aggregated to two: never/rarely and often/sometimes. In a survey of 4,010 American drivers in 2002, 58 percent of respondents considered rolling stops a major threat, while 42 percent admitted to performing them.⁽¹⁷⁾

Drivers’ maneuvers were also extracted as a static variable. Because all chosen intersections consisted of exactly four approaches, the three possible values consisted of left turn, right turn, and straight ahead. These values were not aggregated or manipulated in any way.

Several variables described the scene of the crossing, such as weather conditions (raining, clear, etc.), road surface conditions (wet or dry), and the presence of construction. However, these variables exhibited too little variation to warrant analysis.

Time-Series Data

Whereas the static data provide one data point for each crossing, the time-series data consisted of a large variable amount of observations per crossing. Most data were recorded by the onboard data acquisition systems at a rate of 10 Hz (one observation every 0.1 s). Table 3 lists the raw data provided by VTTI used in this analysis as well as the corresponding definitions.

Table 3. Time-series variables and definitions.

The VTTI timestamp counted milliseconds within each trip and was used to calculate glance times. Latitude and longitude were used in conjunction with intersection center coordinates to calculate how far drivers were from the intersection at any given moment. The moment when the distance between the driver and the intersection center was minimized was considered the point of arrival at the intersection. Coordinates were provided at a 1-Hz frequency and interpolated to 10 Hz using provided speed data. Interpolated coordinates were not simply linear extensions of existing coordinates but calculated based on minute changes in speed (provided at 10 Hz). The brake use indicator, speed, and acceleration variables were left unaltered and were analyzed using their original definitions.

Eyeglance locations and traffic conditions are data of the time-series variety but not readily extractable like the variables listed in table 3. The following section details the reduction process for these data points.

DATA REDUCTION

In addition to static and time-series data, VTTI staff were commissioned to reduce video from four camera angles (forward, driver face, hands/dash, and rear) to produce useable quantitative data regarding eyeglance locations and traffic presence.

Eyeglance Locations

Due to its very nature, video of participating drivers’ faces is considered PII and was therefore not viewable outside of VTTI’s secure data enclave. Reductionists viewed the video feeds of the 411 crossings frame by frame and noted when drivers glanced to the 10 regions of interest (ROIs). Examples of glances to each ROI are provided in figure 5 through figure 11 along with descriptions of each in table 4. Note that the example photos shown here depict a VTTI employee and as such do not violate PII protection agreements.

©Virginia Tech Transportation Institute.

Figure 5. Photo. Example of glance to far left.

©Virginia Tech Transportation Institute.

Figure 6. Photo. Example of glance to near left.

©Virginia Tech Transportation Institute.

Figure 7. Photo. Example of glance to road ahead.

©Virginia Tech Transportation Institute.

Figure 8. Photo. Example of glance to rearview.

©Virginia Tech Transportation Institute.

Figure 9. Photo. Example of glance to near right.

©Virginia Tech Transportation Institute.

Figure 10. Photo. Example of glance to far right.

©Virginia Tech Transportation Institute.

Figure 11. Photo. Example of glance to cell phone.

Table 4. Definition of each ROI.

The video’s refresh rate was 15 Hz, which resulted in a dataset describing eyeglance locations approximately every 0.07 s. All assigned reductionists had been previously trained in VTTI’s eyeglance methodology and tested for accuracy. At the start of this project, assigned reductionists were familiarized with the updated glance location definitions that apply to this project. All reduced data were reviewed by a second-level quality assurance data reductionist; no one performed reviews of their own work. When corrections were identified, the original reductionist would go back to make the changes unless they disagreed with the suggestion. Any remaining disagreements were resolved by a supervisor.

Traffic Presence

Because traffic was considered likely to influence stopping and scanning behaviors, the presence and path of other visible vehicles was also coded by VTTI personnel. Each vehicle was assigned a two-letter code for each frame during which it was visible: the first letter designated the vehicle’s approach direction, and the second letter designated the vehicle’s departure direction. Table 5 shows the construction of these codes, which are illustrated in figure 12 through figure 16. Note that all directions are from the participant driver’s perspective upon arrival at the intersection. These data were later used to indicate the presence of cross traffic (CT) and vehicle queues.

Table 5. Construction of traffic presence codes.

Figure 12. Illustration. Traffic vehicle path approaching from left.

Figure 13. Illustration. Traffic vehicle path approaching from the driver.

Figure 14. Illustration. Traffic vehicle path approaching from right.

Figure 15. Illustration. Traffic vehicle path approaching from ahead.

Figure 16. Illustration. Traffic vehicle path approaching from unknown direction.