Analysis of Naturalistic Driving Study Data: Roadway Departures on Rural Two-Lane Curves (2014)

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.

17 C h a p t e r 4 This chapter describes how roadway, driver, and environ- mental data were reduced. Some additional data reduction may have been conducted for a specific research question and is described in the corresponding data section for that research question. reduction of roadway Variables Roadway variables were extracted for all of the curves identi- fied for the data request, as described in Chapter 3, even though some curves ultimately did not have trips. The RID data collected as part of SHRP 2 Safety Project S04A was used to extract roadway variables when available. In some cases a variable was not collected, and in other cases the RID was not available for the study segment because the RID only covered a portion of the area where trips actually occurred. When the information was not available through the RID, other sources were used to manually extract the data. These additional sources were also used to confirm data collected through the RID, such as speed limit and advisory speed limit. ArcGIS was used to measure distances between curves using the points of curvature included in the RID. ArcGIS was also used to determine whether the curve was an S-curve or a com- pound curve based on the distance between curves. Google Earth was used to extract the roadway features not included in the RID. For instance, chevron presence was available for some of the states in the RID and was manually collected for those in which it was not. Radius was provided for most curves in the RID and was reported as radius by lane. When RID data were not available, radius was measured using aerial imagery. NDS forward video was used to determine subject measures for delineation, pavement condition, road- way lighting, and roadway furniture (which describes objects around the road that provide some measure of clutter), as well as a measure of sight distance for each curve. Variables collected are shown in Table 4.1. The methodology for reducing roadway data is provided in Appendix A. Given that the impact of countermeasures on roadway departure risk was a focus of this research, curves were selected to represent the widest range of countermeasures possible. Curve geometry is also highly relevant to roadway departure risk. Table 4.2 summarizes the number of curves with the indicated radius and countermeasures present. reduction of Vehicle, traffic, and environmental Variables Each of the traces represents one driver trip through a selected roadway segment. One spreadsheet (containing DAS data), one forward video, and one rear-view video were provided by VTTI for each trace. Each row of data represents 0.1 s, and spatial location was provided at 1-s intervals. Several other variables reported at 1-s intervals include use of cruise con- trol, air-bag deployment, date, and heading. A timestamp was also provided to link the various videos with the DAS data. A list of the main DAS variables provided includes the following: • ABS activation • Acceleration, x-axis • Acceleration, y-axis • Acceleration, z-axis • Accelerator position • Air bag, driver • Alcohol • Ambient light • Cruise control • De-identified date • Dilution of precision, position • Driver button flag • Electronic stability control • Elevation, GPS Data Reduction

18 Table 4.1. Roadway Variables Extracted and Main Source Feature ArcGIS SHRP 2 RID Google Earth SHRP 2 NDS Forward Video Curve radius  Distance between curves  Type of curve (isolated, S, compound)  Superelevation  Presence of rumble strips  Presence of chevrons   Presence of W1-6 signs  Presence of paved shoulders  Presence of RPM  Presence of guardrail  Speed limit  Advisory sign speed limit   Curve advisory sign/W1-6    Pavement condition  Delineation  Roadway lighting  Sight distance  Roadway furniture  Direction of curve  Driveways along curve  Driveways along upstream section  Sight distance  Table 4.2. Distribution of Curve Characteristics <1000 1000 to <1500 1500 to <2000 2000 to <2500 2500 to <3000 3000 to <4000 4000 to <6000 6000+ Total Radius 31 19 23 19 14 15 13 14 148 Chevrons 5 3 0 0 0 0 0 0 8 Some paved shoulder 23 18 22 14 9 15 13 13 127 Rumble strips 0 0 0 0 1 1 0 0 2 RPM 4 4 8 5 2 5 4 1 33 Markings obscured or not present 6 3 0 1 1 0 1 0 12 W1-6 4 3 2 0 0 0 0 0 9 Lighting 0 0 0 0 0 0 0 0 0 Guardrail 5 6 5 4 3 0 7 1 31 On-pavement curve signing Not present Flashing beacons or dynamic speed signing Not present

19 • Head confidence • Head position x • Head position y • Head position z • Headlight setting • Lane marking, distance, left • Lane marking, distance, right • Lane marking, probability, left • Lane marking, probability, right • Lane marking, type, left • Lane marking, type, right • Lane position offset • Lane width • Pedal, brake • Pitch rate, y-axis • Radar range rate forward x • Radar range rate forward y • Roll rate, x-axis • Seatbelt, driver • Spatial position (lat/long) • Speed, vehicle network • Steering wheel position • Time into trip • Timestamp • Wiper setting • Yaw rate, z-axis The static driver and vehicle characteristics available include the following: • Driver 44 Driver age 44 Gender 44 Education level 44 Annual miles driven 44 Years of driving 44 Number of moving violations 44 Number of crashes • Vehicle 44 Vehicle year 44 Vehicle model 44 Vehicle make 44 Vehicle track width Smoothing Vehicle variables were either available (i.e., acceleration, position, lane offset) or reduced (i.e., distance from right or left lane line) from DAS variables, which were provided at 10 Hz (one row = 0.1 s). A macro was developed that calculated lane position, change in pedal position, and change in steering position and that smoothed offset, lane posi- tion, pedal position, side acceleration, forward acceleration, and speed. Smoothing was necessary because a certain amount of noise in the data resulted in improbable data points. Several different methods to smooth the data were investigated. The Kalman fil- ter estimates the optimum average factor for each subsequent state using information from past states. It was determined that, although the Kalman filter was appropriate, developing a model for five different variables for more than 1,000 vehicle traces was overly complicated and time consuming. A moving average method was selected because it is able to reduce random noise while retaining a sharp step response. Each of the variables listed above was smoothed using a moving average method, as follows (Smith 2003): 1y i M x i j∑ [ ][ ] = + where y[i] = the output signal; M = the number of points used in the moving average; and x = the input signal. An example of smoothed versus original data is shown in Figure 4.1 for lane offset for a vehicle trace. DAS Data Reduction A macro was developed in Microsoft Excel to calculate addi- tional columns in the DAS worksheets. They include the following: • Spatial location in reference to each curve’s point of curva- ture (PC) and point of tangency (PT) (e.g., 100 m upstream of the curve); • Change in pedal position; • Averaged pedal position (applied smoothing using a mov- ing average method over five intervals); • Forward acceleration (also calculated using the DAS accelerometer); • Averaged forward acceleration (applied smoothing using a moving average method over five intervals); • Change in steering wheel position (only available for a small portion of traces); and • Averaged steering wheel position (applied smoothing using a moving average method over five intervals). The macro also calculated distance from the right edge of the vehicle to the right edge line and distance from the left vehicle edge to the left lane line based on vehicle track width and offset.

20 Information on the make, model, class, and track width of each vehicle was also linked using the subject ID. In addition, the subject ID was used to link the driver demographics, such as the following: • Gender; • Age; • Education; • Work status; • Household income; • Number of miles driven last year; • Average annual mileage for the last 5 years; • Experience (i.e., number of years driving); • Number and type of moving violations in the last 3 years; • Number and type of crashes within the last 3 years; and • Auto insurance for at least 6 months. Vehicles traces were overlaid with the RID. For each curve, the nearest GPS points to the PC or PT were found and the position of the PC/PT located within the time series data using interpolation. Once PC/PT was established, vehicle position upstream or downstream of the curve was accomplished using speed. For some traces, there were multiple curves, so the PC/PT and upstream/downstream distances were determined for each curve. In some cases, speed was missing for multiple timestamps. In these cases, speed was interpolated assuming a constant increase or decrease. Extraction of Data from Forward View The forward video was used to reduce the environmental and other variables. The variables collected included the following: • Surface conditions (e.g., dry, wet, snow); • Lighting conditions (i.e., day, dawn, dusk, night with no lighting, night with lighting); • Visibility; • Locations of vehicles in the opposite direction passing the driver’s vehicle; • Locations where the driver’s vehicle was following another car; • Locations of curve advisory signs and locations where first visible; • Locations of chevrons and locations where first visible; and • Locations of potential roadway departure. reduction of Kinematic Driver Characteristics Initially, three data collection trips were deemed sufficient to reduce 800 to 1,000 traces for Phase 2. However, because data collection trips at the secure data enclave were conducted at the same time VTTI was processing and conducting quality assur- ance on the NDS data, some issues were present that slowed reduction of the kinematic data significantly. A fourth trip was Figure 4.1. Original versus lane offset smoothed using moving average method.

21 made, but only 515 total traces could be coded. These included one crash and three near crashes. This was significantly less data than planned and limited the amount of data that could be used in the different analyses. Hawkeye, the VTTI-developed video data reduction tool that allowed analysts to simultaneously observe multiple camera views and use preset key strokes to code driver char- acteristics, was used to code data for this project. Driver attention was measured by the location on which a driver was focused for each sampling interval. Scan position, or eye movement, has been used by several researchers to gather and process information about how drivers negotiate curves (Shinar et al. 1977). The majority of studies have used simulators to collect eye tracking information. Because eye tracking is not possible with NDS data, glance location was used as a proxy. Glance locations, shown in Figure 4.2, represent practical areas of glance locations for manual eyeglance data reduction. Note that Figure 4.2 does not show “over the shoul- der,” “missing,” and “other” eyeglance locations. Those three locations were determined based on the UI team’s extensive eyeglance reduction experience. Glance locations were coded using the camera view of the driver’s face, with a focus on eye movements, but taking into consideration head tilt when necessary. Potential distractions were determined by examining both the view of the driver’s face and the view over the driver’s right shoulder, which showed hands on/off the steering wheel. Dis- tractions were identified when drivers took their eyes off the forward roadway. The coding process was developed by the UI team. They are experts in the field of human factors and have used similar methodologies in other NDS studies. Potential distractions include the following: • Route planning (locating, viewing, or operating); • Moving or dropped object in vehicle; • Cell phone (locating, viewing, operating); • IPod/MP3 (locating, viewing, operating); • Personal hygiene; • Passenger; • Animal/insect in vehicle; • In-vehicle controls; • Drinking/eating; and • Smoking. Glance location and distractions were coded for each trace. The data reductionist indicated each time the glance location changed, and the data reduction tool recorded the timestamp. Similarly, the start and end times for distractions were also recorded. The data reduction method used to code driver glance location and distraction is provided in Appendix B. Glance location and distractions were manually merged with the trace files using timestamp as a reference. Once this was completed, glance location was indicated for each row in the trace file. As a result, the time series analysis has glance location and distraction at the same resolution as the DAS variables. A number of issues were noted during reduction of the driver face and steering wheel/hand position videos, based on the UI team’s experience in reducing other data sets: • Bright sunlight caused the camera to “wash out” the entire face, especially at certain times of the day when the sunlight Figure 4.2. Glance locations.

22 was more direct. External light sources at night, such as street lights, created the same effect. • Night videos had a grainy quality, making it difficult to dis- tinguish facial features and almost impossible to code eye glances. For a large portion of the files, it was not possible to code glances without the use of head movement. Coding was especially difficult for glances within the cabin that require little head movement (e.g., console, steering wheel). • Many drivers wear sunglasses, which completely obscure the eye, or prescription glasses, which create problems associated with glare. Traces were used unless it was not possible to code glance location or distraction. It should be noted that glance and dis- traction were more likely to have been accurately coded for traces with clearer views of the face and eyes. However, discard- ing data that had some issues would have entailed removing almost all nighttime data and significantly reducing sample size. Glance location was further reduced to indicate time spent with “eyes-off-roadway” engaged in roadway-related tasks or eyes-off-roadway engaged in non-roadway-related tasks based on data coding used by Angell et al. (2006). The authors define roadway-related glances or situation awareness (SA) as glances to any mirror or speedometer. Glances to other locations are defined as not roadway-related (NR). Roadway-related glances (SA) included left mirror, steer- ing wheel, and rearview mirror. The data reductionists could not distinguish between a glance to the right mirror and a glance to the right for other reasons (e.g., to converse with passenger). Additionally, on a two-lane roadway, glances to the right mirror are not likely to be as common because drivers are not expecting vehicles to the right. Consequently, all glances to the right were consid- ered to be non-roadway-related. Additionally, when glances to roadway-related locations were also associated with a distraction, it was decided that these glances were likely to be non-roadway-related. For instance, a driver who was texting and glancing at the steering wheel was likely to be looking at the cell phone rather than the speedom- eter. As a result, non-roadway-related glances included center console, up, right, or down. Data reductionists also indicated characteristics that applied to the trace in general, such as when the driver appeared to be drowsy or emotional. Weather conditions that add to the driving demand were also noted. Summary of Data Limitations About 1,000 traces were identified for Phases 1 and 2, and a data file with DAS variables, a forward-view video, and a rear- view video were provided in-house. Roadway, environmental, and static driver characteristics were reduced or provided for all of the available traces. As noted in this chapter, key DAS variables were not present or reliable for some traces, so not all traces could be used for all of the research questions. Initially, three data collection trips were deemed suffi- cient to reduce 800 to 1,000 traces for Phase 2. However, because data collection trips to the secure data enclave were conducted at the same time VTTI was processing and con- ducting quality assurance on the NDS data, some issues were present that slowed reduction of the kinematic data significantly. A fourth trip was made, but only 515 traces total could be coded, which limited the amount of data that could be used in any of the analyses. Consequently, the main limitation to this study was a smaller than expected sample size. Another limitation was that some types of data were not available and could not be included. Surface friction and pavement edge drop-off are important factors in roadway departure crash risk, but neither could reasonably be col- lected and were not available in the RID. It would have been ideal to intentionally select a range of driver states, such as distraction or drowsiness, to ensure a reasonable sample of certain driver characteristics. However, there was no available method to detect whether driver dis- tractions were present from the time series data so that traces with distraction could be preselected. Distraction could only be identified by viewing the driver face video, which was the last step in the data reduction process. Initially, the team iden- tified a method to preselect traces in which drowsy driving may have occurred by using steering wheel reversals. How- ever, steering wheel position data was only captured for a sub- set of vehicles. As a result, it was not possible to identify drowsy driving using the time series data. The team did target drowsy driving by intentionally including nighttime driving when available. Nighttime conditions were present for 124 traces (~25%), and 36 traces were at dawn/dusk. Additionally, because the accuracy of the alcohol sensor was not known at the time data were collected, potential impairment could not be targeted. Another limitation was that newer vehicle technologies could not be targeted. Vehicles with electronic stability con- trol (ESC) or collision warning systems made up only a small fraction of the vehicle fleet. Because other factors had a higher priority, vehicles with advanced technologies were not spe- cifically targeted. Curve characteristics were described earlier in his chap- ter. Other characteristics represented in the final available data set are described in Figures 4.3 and 4.4. The distribu- tion of driver age and gender (n = 202) represented in the

23 viable traces is shown in Figure 4.3. A number of drivers had multiple trips. Speeding was common, as indicated in Figure 4.4, which shows the percentage of drivers who entered the curve over the advisory speed limit if present or posted speed limit if not present by a certain threshold. Almost all drivers entered the curve 5 mph or more over the advisory speed, and a large fraction of drivers entered the curve 20 mph or more over the advisory speed. Data are summarized from the data reduced for Research Question 3 (sample size = 583). A summary of the crash/near-crash events is provided in the sections below. Summary of Crash 1 Event: 14950079 Driver: Male Age: 21 Passengers: None Location: WA Figure 4.3. Distribution of driver age and gender. Figure 4.4. Percentage of drivers exceeding advisory or posted speed limit.

24 Month: August Time of day: Midnight to 3:00 a.m. Type of crash: Run-off-road Description of site: This crash occurred on the second curve of an S-curve in rural Washington State. The driver was travel- ing from east to west. The curve on which the ROR crash occurred had a radius of approximately 50 m, with a lane width of approximately 3 m and no shoulder. This curve had guardrail on the inside portion of the curve and did have chevrons. The curve also had a curve advisory sign and a curve advisory speed of 20 mph, a reduction from the 45 mph speed limit on the rest of the roadway. Data on seatbelt usage, ESC activation, and the status of traction control and cruise control were missing from the DAS data, but from the driver video it was determined that the driver was wearing a seatbelt. From the driver face video, it appears that the driver may have been tired, as he is looking forward and resting his head on his right arm/hand. Description of crash: The driver took the right-hand curve too fast, departed the lane to the left at 4082869, and departed the roadway at 4084070 after crossing over oppos- ing lanes of travel. The impact was with the ditch, bushes, and trees along side of roadway. The driver appeared to be unhurt. Suspected main contributing factor: Speed appears to be the main contributing factor. The driver entered the curve travel- ing approximately 54 mph, 9 mph over the posted speed limit and 34 mph over the curve advisory speed limit. The only glance outside of a forward glance that occurred was a glance to the steering wheel, which lasted 0.6 seconds and occurred 7.7 seconds before the driver left his lane. This glance was not associated with a distraction. Summary of Near Crash 1 Event: 11543161 Age: 18 Gender: Female Passengers: Three (one front, two rear) Location: VA Month: July Time of day: Noon to 3:00 p.m. Type of crash: Near rear-end Description of site: This near crash occurred on the tangent downstream of a curve. The roadway had lanes approximately 3.33 m wide, paved shoulders, centerline rumble strips, and raised pavement markings. The speed limit was 55 mph. Data on seatbelt usage, ESC activation, and the status of traction control and cruise control were all missing from the DAS data. From the driver face video, it appears that the driver may have been tired, as she is looking forward and rest- ing her head on her right arm/hand. Description of near crash: The driver came up over a vertical curve while reaching to pick up a drink from the cup holder. The driver noticed that the vehicles ahead had come to a stop and had to brake hard to avoid a rear-end collision. Suspected main contributing factor: The driver being distracted coupled with the presence of a vertical curve that limited sight distance appear to be the main contributing factors for this near crash. The driver was traveling under the 55 mph speed limit. Summary of Near Crash 2 Event: 15483160 Age: 25 Gender: Female Passengers: One front Location: IN Month: February Time of day: 3:00 p.m. to 6:00 p.m. Type of crash: Near rear-end Description of site: This near crash occurred near the PC of a curve with a radius of 1,637 m. The speed limit on the road- way is 50 mph, with lanes approximately 3.2 m wide. There are paved shoulders and a curve advisory sign alerting drivers to the curve. Data on seatbelt usage, ESC activation, and the status of trac- tion control and cruise control were all missing from the DAS data. The driver is talking to a passenger throughout the trip. Description of near crash: The driver looked away and was operating in-vehicle controls when the vehicles ahead came to a stop. The driver swerved to the right shoulder to avoid a rear-end collision. Suspected main contributing factor: The driver is talking to the passenger during most of the event and that appears to be the main contributing factor for this near crash. The driver is operating the in-vehicle controls while looking at the center console from 1297157 to 1298625. The brake lights of the vehicle ahead are visible at 1297290. The participant vehicle does not begin braking until 1.3 s later. The driver was travel- ing at the 55 mph posted speed limit. Summary of Near Crash 3 Event: 22512290 Age: 24 Gender: Male Passengers: None Location: NY Month: August Time of day: 3:00 p.m. to 6:00 p.m. Type of crash: Near rear-end Description of site: This near crash occurred near the PC of a curve with a radius of 570 m. The speed limit on the roadway

25 is 45 mph, with lanes approximately 3.5 m wide. There are paved shoulders and a curve advisory sign alerting drivers to the curve. Data on seatbelt usage, ESC activation, and the status of traction control and cruise control were all missing from the DAS data. From the driver face video, it appears that the driver may have been tired, as he is looking forward and rest- ing his head on his right arm/hand. Description of near crash: The vehicle ahead applied its brakes at 392583. At this same time, the participant driver glanced at the center console. The driver braked hard to avoid hitting the vehicle ahead. Suspected main contributing factor: The driver glancing away appears to be the main contributing factor to this near crash. The driver is looking at the center console from 391983 to 392850. The driver ahead slams on his/her brakes starting at 392583, so the participant driver does not notice the vehicle braking ahead for 0.3 seconds after the braking. The driver was only going about 1 mph over the speed limit of 45 mph, so speed did not appear to be a factor.