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.
8Data Sources Data for this study were developed from a naturalistic driving data (NDD) resource obtained in a field operational test (FOT) conducted in southeastern Michigan and spatially joined with highway information and crash data from the same region. The following section describes the data sources and the devel- opment of the databases used in this study. UMTRI Naturalistic Driving Data NDD are data collected in vehicle studies using a sample of nonprofessional drivers driving test vehicles in pursuit of their normal everyday activities. Test vehicles are instrumented with a broad range of sensors to monitor the vehicle systems, driver, roadway, and environment. A data acquisition system on the vehicle collects the data, which are later transmitted to a rela- tional database and subsequently retrieved by researchers. The raw data are supplemented by derived measures, and other relevant information (e.g., solar elevation angle). In the past decade, several FOTs have used naturalistic driving studies in the evaluation of advanced vehicle-based safety technology. The NDD used in this study came from an FOT conducted to collect evidence of the interaction of lay drivers with a com- bined lane departure technology and curve speed warning system that has been termed the road departure crash warn- ing (RDCW) system (LeBlanc et al. 2006). The program was led by the University of Michigan Transportation Research Institute (UMTRI) under a cooperative agreement with the U.S. Department of Transportation. The RDCW systems were developed and integrated by Visteon Corporation and Assist- Ware in preparation for the FOT. The FOT involved exposing a fleet of 11 RDCW-equipped Nissan Altima cars to 10 months of naturalistic driving. The 78 test participants were lay drivers from southeastern Michigan, randomly selected and recruited from Michigan driver licensing records, who drove these cars as their personal vehicles for several weeks. There were 9,582 trips in the RDCW FOT. Those trips covered 133,290 km (82,773 mi) and took 2,487 h. The RDCW FOT was expressly intended to study road departure crash warning, so that the sensors and instrumentation were ideally suited to study lane-keeping behavior. Data gathered by using UMTRIâs data acquisition system included more than 400 data signals. Among them were video samples of the forward driving scene and driverâs face; differ- ential Global Positioning System (GPS) time and position; lane tracking, including boundary type (solid, dashed); forward and side radar returns; distance to lane edge; available maneu- vering distance; vehicle velocity; yaw, pitch, and roll; and sen- sor data on lights and windshield wipers. The data file (raw vehicle data and derived tables) is approximately 250 gigabytes in size and is stored in a SQL server database. The data from the RDCW FOT will be referred to in the rest of this document as the UMTRI NDD. Highway Data The major data source for highway information in this study was the enhanced Highway Performance Monitoring System (HPMS) data for the state of Michigan in 2005. The HPMS is a national-level highway information system that includes data on the extent, condition, performance, use, and operat- ing characteristics of the nationâs highways. Limited infor- mation on travel and paved miles is included in summary form for the lowest functional systems. The HPMS database contains administrative and extent of system information on all public roads, descriptive information in a mix of universe and sample data for the arterial and collector functional sys- tems, and area-wide summary information for urbanized, small urban, and rural areas. The road system is divided into individual segments. Information about the type of road and rural/urban designation is available for all road segments. The number of through lanes is available for all segments on all road types but minor collectors and local streets. Geometric information (e.g., curves, grades, shoulder and median types and widths) and traffic information (e.g., speed limits, peak C H a p T e R 3
9capacity) are available only for a sample of segments. Traf- fic volume information (AADT) is available for the sampled segments and also for segments that lie within special areas sampled for air quality monitoring. Each road segment in HPMS has a unique identifier in a linear referencing system (LRS), which allows the segment to be located spatially and joined to geospatial databases. Crash Data Michigan police-reported crash data were used to identify a set of road departure crashes for analysis. Michigan crash data were selected because they provide the best opportunity to link together crash data, roadway geometric and exposure data, and NDD from an FOT. Both the National Automo- tive Sampling System General Estimates System (NASS GES, or simply GES) and the NASS Crashworthiness Data System (CDS) were considered for the analysis. However, the road- way information in GES is severely limited, and it is not pos- sible to link in data from other files (such as HPMS) because the file is well scrubbed to ensure that specific crashes can- not be identified. Crash locations within the NASS CDS also cannot be identified, which prevents linking the crashes to sources of additional data about the roadway geometry. The Michigan data also provide an example of data sources that will ideally be available for the large naturalistic driving field study. Michigan geo-locates (identifies the crash location by using latitude and longitude coordinates) virtually all crashes. HPMS files on specific roadways are also available and these data can be linked with the crash locations using the latitude and longitude coordinates. The HPMS files include AADT information so that crash frequencies at specific locations may be normalized by a measure of exposure. Finally, FOT data are available for a set of counties in southeastern Michi- gan and provide information about driving behavior through specific locations. Five years of Michigan crash data (2001 to 2005) were used in the analysis, including records for almost 2.2 million crash involvements of passenger cars. The records were compiled from information recorded by police officers on State of Mich- igan Traffic Crash Report forms (UD-10). The form is optically scanned and data to populate the crash file variables are cap- tured from the scan. Crashes are geolocated from the descrip- tion of the crash location entered by the reporting officer. The Michigan crash data have limitations that apply to all police-reported data. Certain items are probably recorded with reasonable accuracy and completeness, such as time of day, road type, roadway alignment, vehicle type, weather, and so on. Other more transient conditions are inherently more difficult to identify accurately, such as driver fatigue and dis- traction. Certain important pieces of data such as travel speed, lane position, roadway radius of curvature, and superelevation are not recorded at all. Moreover, the system of quality control on the data is not rigorous. The crash reports are supposed to be reviewed for accuracy by supervisors at the enforcement agency before being submitted to the state. But once a report is scanned, the only check is to ensure that the scanner accu- rately captured what was entered on the form, not whether the data were accurate in the first place. It is important to recognize the limitations in the Michigan crash data. However, in practice, the limitations are shared with other crash data files and, on balance, the ability to link the data with the HPMS files and the availability of the NDD for the same areas as the crashes make for a very powerful combination for the analysis. The crucial data element in the Michigan crash data is the location information. Previous experience with the location information for a project on signal optimization in Michigan has shown that the accuracy of the location information is good (Green and Blower 2007). The top image in Figure 3.1 shows an aerial photo of an intersection and the bottom is a pin map of crashes at that intersection, using a map gen- erated by the geographic information system (GIS) package Maptitude. The longitude and latitude coordinates are accu- rate enough to discriminate between the different directions of travel on the two roads. After crash locations were reviewed for 130 intersections in southeastern Michigan, few were found to be inaccurate in a gross way (e.g., the location entered on the UD-10 indicated that the crash was not at the intersection specified by the longitude and latitude coordinates). Moreover, 12 Figure 3.1. Crashes geographically located at Hall Road (M-59) and Schoenherr Road and at Hall Road and the eastbound crossover west of Schoenherr Road.
10 missing data rates for location were low. Only about 2.5% of the crashes could not be located. Road departure crashes are identified from a set of events variables. The variable records âwhat happenedâ in the crash, and include both harmful events (such as collisions) as well as nonharmful events (e.g., crossing the centerline or running off the road). Events include collisions with specific types of fixed and nonfixed objects, rollovers, roadway departures, loss of control, crossing the centerline or median, and reen- tering the roadway. This system in principle permits quite a complex series of events to be captured. Up to four events may be coded, though in practice only one event is coded in about 83.8% of cases. For the purposes of identifying road departure crashes, cases were flagged as roadway departure if a collision with a fixed object or run-off-road event occurred before a collision with a motor vehicle or a nonmotorist. In other words, any time a vehicle went off the road before a collision with an on-road vehicle or a nonmotorist, the case was counted as a road departure crash. Thus, road departure crashes may include on-road collisions with motor vehicles, as long as a roadway departure occurred before the collision. The crash data include a large sample of road departure crashes for the analysis. The Michigan crash files for 2001 through 2005 include records for almost 2.2 million vehicles involved in a traffic crash, of which 192,512 records involved a roadway departure before any harmful event. The UMTRI NDD were collected in eight counties in southeastern Michi- gan: Lenawee, Livingston, Macomb, Monroe, Oakland, St. Clair, Washtenaw, and Wayne. Restricting the roadway departure crashes to those counties produced 73,135 involvements. Longitude and latitude were missing for 1,827 crashes (2.5%), leaving 71,308 road departure crash involvements for analysis. The road departure crashes in the sample are more severe than other crash types. About 0.6% of the road departure crash involvements included a fatality, compared with 0.2% of other crash involvements. About 27% included one or more injuries, compared with 23.3% of other involvements. The distributions of these road departure crashes by envi- ronmental and driver conditions are shown in Appendix B. analysis Data The UMTRI NDD were spatially joined to the highway and crash data from eight counties in southeastern Michigan by using GIS software tools from ArcMap Version 9.2 from the Environmental Systems Research Institute, Inc. (Esri). A spa- tial base map of Michigan from the Michigan Center for Geo- graphic Information (CGI) of the Michigan Department of Information Technology provided the key layer for the GIS tools. CGI also provided the research team with a digital map for all public roads in Michigan. The data for each road include jurisdiction, a physical reference number, and the roadâs func- tional class among other descriptors. The project used Version 6 of the base map, which represents some 8,765 mi in southeast- ern Michigan. An illustration showing the four layers used to join the UMTRI data sources is shown in Figure 3.2. In addition, CGI provided the project with ortho-imagery (aerial maps) for the southeastern Michigan region. Spatial data layered on the aerial images helped to identify pavement markings, roadside obstacles, and other road features not available via spatial data sets. The database for subsequent analyses was developed from the spatially joined data. Only HPMS-defined segments that were also in the NDD were included. Because traffic volume Figure 3.2. Illustration showing the spatial layers of the data used in the UMTRI analysis. Sources: Google Earth and MapQuest Transportation Display.
11 curve information for the segments not in the HPMS sample. Because only road segments that had been traversed by an instrumented vehicle are included in the analysis database, information about the vehicleâs path can be used to obtain additional roadway information. The team found that it was possible to obtain the degree of curvature and length of horizontal curves from the vehiclesâ path and yaw rate infor- mation. However, the procedure was labor-intensive and time-consuming. For example, drivers do not follow the same path through a curve, drivers turn off onto driveways, or a long horizontal curve appears as a series of short curves interspersed with tangent sections. While this clearly was a way of obtaining curve information, the team did not spend time and effort to develop a more efficient algorithm because there still was the problem of describing the curves if there was more than one curve in the segment. It was clear from this experience that a different way of defin- ing road segments for future studies is needed; for example, dividing the roadway into segments so that there is only one horizontal curve in a segment, or in a way that each horizon- tal curve is a unique segment, would have advantages over the use of HPMS-defined segments. For the purposes of this study, the team used the yaw rate in the NDD for the segment and defined a variable indicating the presence or absence of at least one curve in the roadway segment. Exposure measures for each segment were also developed. Because the team used 5 years of crash data, the exposure for crashes was based on the volume of vehicles entering the seg- ment over 5 years, as well as the segment length. The exposure for surrogate events was based on the number of traversals of instrumented vehicles and the segment length. These expo- sure measures are to be included in the statistical modeling, and this is considered further in Chapter 5. The team was not able to code roadside information into its analysis database. Although the ortho-imagery provided visual information on the roadside, this information was not in a form that could be used for formal analysis. Another option would be to conduct labor-intensive manual coding of the roadside, but this option was not considered feasible. This data limitation precluded the teamâs ability to address research questions concerned directly with roadside features. The team recognizes that methods do exist for distinguishing between features that are found in the roadside (e.g., spectral analysis) and sees this possibility as something that could be explored in future work. information was needed for exposure, HPMS segments with- out AADT information were excluded. Examination of the segments without volume information showed them to be minor collectors and local streets. HPMS segments are not directional, that is, the data are for both directions of travel if the road is two-way. Because the direction of a curve and the direction of road departure are relevant in the teamâs analyses of road departures, the road segments in the teamâs analysis data are the HPMS road seg- ments plus the direction of travel. While it is possible that the close spatial proximity of the opposite sides of the same segment of road could cause unmodeled correlations (which might be analyzed in future studies), the two travel directions were only considered to be correlated via the coincidence of the explanatory highway variables used. Thus, the research teamâs basic unit of analysis was the directional road segment for which traffic volume informa- tion was provided, that had been traversed by at least one of the drivers in the NDS. There were 9,526 direction road seg- ments in the analysis database. Of the 71,308 road departure crashes in the eight counties of southeastern Michigan that were recorded in the Michi- gan crash data file from 2001 to 2005, 21,340 crashes were on the traversed directional segments in the teamâs analysis database. Of those crashes, 7,562 vehicles departed the road to the right and 4,372 vehicles departed to the left; in 9,406 cases, the direction of the departure was not known. Selected attributes from the HPMS database were spatially joined to the analysis road segments. Road type information was taken directly from the functional classification of the road segment in the HPMS database. The rural/urban desig- nation was also taken directly from the HPMS codes. Infor- mation on types and widths of shoulders was taken directly from the HPMS data, if available for the segment. Other- wise, the shoulder variables were assigned the median value obtained from the HPMS sample road segments of the same functional class, with the same number of through lanes, in the same county. Horizontal curves in sample HPMS segments are clas- sified into six ranges of degrees of curvature, and the total length of curves in each of these categories is reported in the HPMS database. This form of information was not useful for the research teamâs analyses because there was no simple and meaningful way to summarize the curve information for the segment, and also there was no credible way to impute
