Naturalistic Driving Study: Development of the Roadway Information Database (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.

36 C h a p t e r 6 To ensure the collection and delivery of good quality data to support the SHRP 2 goals and objectives, the research team developed a quality assurance plan for the new data provided by the mobile data collection project. The quality assurance plan outlined the processes to ensure optimum data quality from project setup through to final data delivery, including the accuracy requirements and the tolerances for what was deemed a nonconforming product. In addition, the quality assurance plan defined the process in the event that there was a nonconforming product delivered by the mobile data con- tractor. In the event that delivered data fell outside of these tolerances, the mobile data vendor was required to take adequate corrective actions, which could have included reprocessing or re-collection of the nonconforming data. All collected data passed the quality assurance guidelines and no re-collection was required. The team used the initial phase of the data collection (pilot study) to address all the quality issues, and thus no further corrective action was necessary afterward. 6.1 Quality assurance process A major and continuous task conducted by the research team was the mobile data quality assurance process. The objective of this task was to ensure that the mobile data conformed to the accuracy requirements. Based on the pilot testing, the initial accuracy requirements (Table 3.3) were revised as shown in Table 6.1. For each NDS site, and for each project year (2011, 2012, and 2013), the project team collected information that would be used as part of the quality assurance process. The quality assurance included two processes performed simultaneously: one analyzing roadway features and the other analyzing alignment. The first process used control sites and random sites (described below) where all roadway features were collected. These roadway features were as follows (Figure 6.1): • Intersections (location, type, number of approaches, control type); • Signs (MUTCD code, speed limit value, image); • Highway lighting (presence); • Lanes (type, width); • Medians (presence, type); • Shoulders (type, width); • Rumble strips (presence, type); • Grade and cross-slope values; and • Barriers (barrier type, start/end treatment type, post material, rub rail). The second process analyzed the alignment (tangent, curve) for the accuracy of the radius. For the control sites, the curvature data was obtained from the DOTs. For the random sites, this process was completed using the GPS traces from the mobile data collection vendor to determine the chord and length of curve, which was used to determine the radius (Sec- tion 6.5 provides the details for this process). 6.2 Quality assurance Field Data Collection 6.2.1 Site Visits Two types of site visits were conducted by the team: control sites and random sites visits. Both site visits collected all the roadway features. Random site visits were typically 4-day visits to an NDS location where the team drove the majority of the data collection routes, collecting roadway attributes at ran- dom locations. Control site visits were done on pre-identified road segments. Each state contained both a rural and an urban control site, which were a couple of miles in length and had a desired combination of curves, tangent sections, and roadway features. 6.2.1.1 Random Site Visits For each NDS site location and each year of data collection (2011, 2012, and 2013), random site visits were used to collect roadway features. During the visit, the field team drove along Mobile Data Quality Assurance

37 Table 6.1. S04B Data Accuracy Requirements Data Element Minimum Accuracy Requirement (/) Curvature radius 100 ft (curves less than 1,500 ft radius) 250 ft (curves between 1,500 ft and 6,000 ft radius) Within 13% (curves over 6,000 ft radius) Curvature length 100 ft (curves less than 1,500 ft radius) 250 ft (curves above 1,500 ft radius) PC 50 ft PT 50 ft Grade (+ or -) 1.0% Cross slope/ Superelevation 1.0% Lane width 1 ft Paved shoulder width 1 ft Inventory features (signs) location 7 ft Thru Lane: 1 (21’) Thru Lane: 1 (12’) Accel. Lane: 1 Thru Lane: 2 (11’) Deccel. Lane: 1 Thru Lane: 1 (12’) Left Turn Lane: 1 Thru Lane: 1 (14’) Deccel. Lane: 1 Flush Paint. Flush Paint. Flush (Painted)Flush (Painted) 2’ Mix/Combo 0’ Mix/Combo 3’ Mix/Combo 2’ Mix/Combo N/A N/A N/A Grade, Cross Slope Unpaved Shoulder: N/A Rumble Strips: N/A Lighting: N/A Flush Paint. Flush Paint. Flush (Painted)Flush (Painted)N/A N/A N/A Unpaved Shoulder: N/A Rumble Strips: N/A Lighting: N/A Grade, Cross Slope Thru Lane: 1 (12’) Thru Lane: 1 (11’) Right Turn: 1 Thru Lane: 1 (12’) 3’ Mix/Combo 3’ Mix/Combo 4’ Mix/ComboN/APaved Shoulder Median Lanes Paved Shoulder Median Lanes Figure 6.1. Roadway features on a road segment. a majority of the data collection routes, stopping randomly to collect all the roadway-attribute data at that location. Between 50 and 100 random sites were collected for each site visit, based on the number of data collection miles during that time. The team collected a representative sample of data, in terms of coverage and roadway attributes. The total number of random sites per study site is as follows: • Florida: 206. • Indiana: 244. • New York: 211. • North Carolina: 346. • Pennsylvania: 199. • Washington: 206. Before each site visit, maps were prepared by the GIS experts on the team to be used by the field team (Fig- ure 6.2). The maps contained the data collection routes during that time and Continually Operating Reference

38 Station (CORS) locations, which were used to ensure accu- racy with the GPS data. These maps were used in the field to select random points for data collection and verify com- plete coverage of the data collection routes. These maps were also loaded on a Trimble GeoExplorer 6000 Series GPS data collection device, which was used in the field to enter roadway-feature locations, selected attributes, and images (e.g., sign faces). The process at each random point consisted of capturing the location from the GPS receiver by averaging 100 posi- tion readings within Esri ArcPad. All roadway features were also documented. Each random point collected was given an ID to allow for joining the GPS data and roadway-feature information. Images could also be taken with the GPS- enabled device for reference during the quality assurance process. A form, shown in Figure 6.3, would then be filled out referencing all the roadway features’ attributes. During the process of selecting a random point, the field team made sure to get a representative sample of each of the roadway attributes. 6.2.1.2 Control Site Visits Each NDS study location had two control sites, which were used for all the data collection years. The control sites were selected in cooperation with the DOT in each study area. The control sites had complete curve information (PC, PT, radius, and length). These pre-identified road segments, which are a couple of miles in length, were located in an urban and rural environment. The sites included a combination of curves and roadway features for the quality assurance pro- cess. The control sites were collected by the mobile vans before every deployment to the specified data collection routes, at the end of data collection, or if the equipment was nonoperational for an extended period of time. This resulted in a detailed quality assurance process of the data for every year. Figure 6.4 shows a map of the control sites in Washing- ton. Detailed curve information was also collected as part of the control visit. The control site visits collected data in a fashion similar to that used for the random sites with the GPS-enabled device and a form for documentation of the roadway fea- tures. The difference between control and random sites is the number of data collection points. Points were collected at every sign, intersection, and barrier, as well as random locations along predefined sections within each control site. At each of these points, all other roadway features were col- lected as well. This gave detailed data for the control site that would be compared with the mobile data collected. The col- lected data were then linked to mobile data for quality assurance. No horizontal alignment data were collected along the con- trol or random sites by the Center for Transportation Research and Education (CTRE). For all control sites, horizontal curve parameters, including PC, PT, curve radius, and curve length, were obtained from the respective state departments of trans- portation in the form of as-built plans and preconstruction surveys. Figure 6.2. An example map prepared before a site visit.

39 6.3 Linking Field Data to Mobile Data Before linking field and mobile data, the field GPS data were differentially corrected using Trimble GPS Pathfinder soft- ware. Automatic carrier and code processing corrections were made using the nearest available base provider (CORS site). Esri shapefiles representing the corrected positional data were also created. All field attribute data were entered into a database and joined to the corrected shapefiles. Then, the data collected at the random and control sites were linked to the mobile data. The research team developed a process to link the data collected at random site visits to the data provided by the mobile data collection project. A cus- tom GIS module was developed and used to link site visit data points and attributes to mobile data layers, which are presented by each roadway feature (Figure 6.5). The GIS module used a semiautomated process in which the user (quality assurance analyst) would select the closest mobile data roadway features to the random point for each layer and then run the script to join those attributes in a Micro- soft Access database. Attributes for the mobile data could then be compared to random site data to verify accuracy requirements (Figure 6.6). Figure 6.3. An example form used in random site visit.

40 Figure 6.4. Control sites in Washington. Figure 6.5. GIS module to link data. Only one roadway feature would be assigned to each ran- dom point for all layers, when present, except rumble strips and lighting, because they presented segments and not points. Rumble strips were selected along both sides of the lane, if present, and given a prompt to define the rumble strips as a set. Lighting was selected along both sides of the road for analysis. For the control sites, multiple runs were conducted by the mobile data collection vans. In these cases, one road- way feature for each layer present was assigned for every run. This verified the accuracy for each of the runs through the control site. The GIS module also had input features to flag roadway fea- tures that had ambiguity and needed further review. Ambiguity occurred when a discontinuity occurred in the roadway feature, and the research team was notified to go back and investigate if there were differences. This review was used when the QA ana- lyst was unsure about what roadway attribute to select. The research team would then further investigate before moving forward with the quality assurance. 6.4 accuracy analysis With the field data linked to the mobile data, the quality assurance analysis was conducted for each of the NDS sites for each year roadway data were collected. The same process was used in the analysis for the random sites and the control

41 sites. The only difference was that the control sites compared each of the runs through the control site. The linked data allowed for easy comparison of the ran- dom sites data to the mobile data. Each of the roadway fea- tures was analyzed independently to verify the accuracy of the designated attributes. If any differences occurred, the ID Figure 6.6. Linking random site control data to S04B data. for that attribute would be noted and confirmed using the ROW images that were collected at 21 ft intervals (Fig- ure 6.7). If an error was confirmed, a report was created with a detailed location and the discrepancy. A final report from each site was provided to the contractor so that the issues could be investigated and addressed. The majority of the Figure 6.7. An example of an ROW image.

42 issues were identified during the pilot testing, and no further corrections were needed. The grade and cross-slope data set (location layer) was the first step of the quality assurance evaluation. The location layer is present along all the roadways collected. The first step was to verify that location data were present for all the ran- dom sites. Any point without location data was flagged and brought to the contractor’s attention. After the coverage was evaluated, the absolute difference in the grade and cross slope were compared. Intersections were analyzed to verify that the intersection was identified correctly and within a reasonable distance (150 ft or less) from the field-collected location. The control type and number of approaches were also verified for accu- racy. Highway lighting and rumble strips were both evaluated for presence. The type of rumble strip was also verified for accuracy. The lanes data set was evaluated to determine any difference between the lane widths, assessing the required 1 ft accuracy. The lanes data were also evaluated for total number of lanes and correct classification of all lane types. For medi- ans, presence and the type were analyzed. Shoulders were divided into two roadway features: paved and unpaved. Paved shoulders were evaluated for the width of the shoulder being within the 1 ft acceptance, as well as the shoulder type. Width for unpaved shoulders was not a data type required for this project. They were evaluated for correct classification of shoulder type only. All signs were evaluated to verify that the sign was identified and within 7 ft of its field-collected loca- tion. Speed signs were analyzed to verify that the speed limit text was correct and whether the sign was regulatory or advi- sory. All other signs were evaluated for correct MUTCD code, as well as any other additional text. Guardrails were evaluated for accuracy of the begin/end treatment type, the guardrail type, and the post material. Charts were developed for each roadway feature discussed above, presenting the number of attributes collected at the ran- dom sites in comparison to the linked mobile data. Example charts are shown in Figure 6.8. The figure shows an accuracy comparison for speed limit signs (message and location). As can be seen from Figure 6.8, both items meet the minimum accuracy requirements (95% passing). A detailed example of the quality assurance process for Pennsylvania is provided in Appendix E. 6.5 Curve Quality assurance The second part of the quality assurance process involved the curve analysis. The alignment data set contained several curve attributes, including whether the record (roadway extent) was a horizontal curve or tangent. For all horizontal curves, the begin/end points, direction, length, and radius were provided. The curves were separated into three radius ranges: less than 1,500 ft radius, 1,500 ft to 6,000 ft radius, and greater than 6,000 ft radius. Each of the ranges had different accuracy requirements, as presented in Table 6.1. For the control sites, mobile-collected curve PC, PT, radius, and length were compared to the corresponding DOT-provided alignment attributes. For the random curve evaluation, the GPS traces from the mobile data collection were used to identify ran- dom curve locations, and then the radius was calculated (using the method described below) and compared to the mobile alignment data for quality assurance. The process to determine the radius of a curve was devel- oped by Hans et al. (2012) using the curve length and chord length to calculate the radius of the curve (Figure 6.9). This technique, known as the long chord method, was reported to yield a root mean squared error (RMSE) of 19.5%, and the coefficient of determination was 0.90. This provided an accu- rate measure of the radius, in comparison with the ground truth, with 95% of the radii values with an error less than 30% and 68% of the radii with an error less than 5%. Within Esri ArcGIS, points within a horizontal curve were selected from the mobile data collection GPS traces. The selected points were joined together, forming a polyline. The length of the polyline was calculated, as well as the straight-line distance from the curve begin and end (i.e., chord). These values were used to compute the curve radius. 18 18 0 1818 18 0 17 0 5 10 15 20 Signs identified Speed limit text Advisory speed limit Sign location (GPS coordinates) Speed Signs Random Sites Vendor Figure 6.8. Comparison charts example.

43 Figure 6.9. Radius calculation for quality assurance (Hans et al. 2012). Curves were selected randomly, with equal frequencies of curves in each radius range. After initial selection of the curves, the radius values were calculated. Multiple iterations were completed until a sufficient number of curves were selected in each radius range. The resulting curves, with calculated radius values, were spatially joined to the mobile alignment data for quality assurance. A statistical analysis was completed for the curves to deter- mine the distribution of the differences between the calculated Figure 6.10. Curve radius quality assurance check. radius and the mobile-collected radius. Any major differences in radius were further investigated to determine any errors in the calculation process or spatial assignment of curves. The final summary of the three radius ranges were then compiled similarly to the method shown in Figure 6.10 to verify that the curves met the radius accuracy requirements. As can be seen from Figure 6.10, the minimum accuracy requirements were met for all radius values (less than 1,500 ft and between 1,500 ft and 6,000 ft).