National Academies Press: OpenBook

Integration of Analysis Methods and Development of Analysis Plan (2012)

Chapter: Chapter 6 - Examples of Summary Work Plans

« Previous: Chapter 5 - Work Plan Requirements
Page 21
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 21
Page 22
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 22
Page 23
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 23
Page 24
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 24
Page 25
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 25
Page 26
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 26
Page 27
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 27
Page 28
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 28
Page 29
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 29
Page 30
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 30
Page 31
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 31
Page 32
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 32
Page 33
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 33
Page 34
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 34
Page 35
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 35
Page 36
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 36
Page 37
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 37
Page 38
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 38
Page 39
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 39
Page 40
Suggested Citation:"Chapter 6 - Examples of Summary Work Plans." National Academies of Sciences, Engineering, and Medicine. 2012. Integration of Analysis Methods and Development of Analysis Plan. Washington, DC: The National Academies Press. doi: 10.17226/22847.
×
Page 40

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.

21 C h a p t e r 6 This chapter provides sample work plans for five global research areas: lane-departure crashes, intersection crashes, driver distraction, driver fatigue, and alcohol-impaired driving. Overview of Work plans Each proposed Project S08 work plan should follow this outline: 1. Overview of research topic. 2. Specific research question(s). 2.1. Crash type(s) addressed. 2.2. Proposed surrogate measures. 2.3. Rationale for research questions and use of natural- istic driving data. 2.4. Hypotheses to be tested. 3. Data analysis plan. 3.1. Data sampling, segmentation, and aggregation. 3.2. Model formulation. 3.3. Analytical approach. 3.4. Model validation. 4. Pitfalls and limitations that may be encountered and how to address them. 5. Documentation of results. 6. Expected impact or outcome of research on counter- measures or policy implications. Each work plan needs to clearly identify the global research area being addressed and how the research will contribute to the knowledge in this area. Typically, the research area will be specified by the RFP. Otherwise, the proposer needs to justify the prioritization of the proposed global research question. Proposers should articulate both the importance of the specific research question to SHRP 2 goals and why NDS data are required for their analysis. The specific research ques- tion should also identify what types of crashes the study will address (rear end, angular, intersection, roadway departures, or some combination of crash types). The intent of SHRP 2 is to improve traffic safety, and all funded projects should contribute to the goal of reducing crashes, particularly fatal or injury crashes. Research questions that are mainly relevant to low-severity or low-incidence crash types (e.g., property only) would be a lower priority for SHRP 2. If crash type is not specified, the conclusions of the research cannot direct interventions to improve safety. Accordingly, the work plans should demonstrate that the specific research question is rel- evant to a specific crash type. The S02 team has developed five work plans to demonstrate issues that may be encountered when preparing a proposal and how a team might address some of the RFP requirements. Each example work plan addresses one of the global research topic areas: 1. Lane-departure crashes; 2. Intersection crashes; 3. Driver distraction; 4. Driver fatigue; and 5. Alcohol-impaired driving. The work plans briefly illustrate what might be included in a proposal. It is important to note that these are examples only. The specific research questions used are not neces- sarily the most important questions regarding the particu- lar global topic area, and the data reduction and analytical methods may not be the best or the only tool to answer the questions posed. Project S08 researchers are not expected to use the same examples or generate similar outcomes. On the contrary, the goal is to provide enough examples to guide researchers in the development of their own innovative ideas. These work plans provide concrete examples of the chal- lenges associated with naturalistic research and how they might be overcome. The work plans vary in the level of detail they provide to most efficiently highlight the range of issues that researchers need to consider. The first research plan (the Examples of Summary Work Plans

22 Proposed Surrogate Measures Because crashes are relatively rare events, other factors, such as the amount of lane deviation, will be used as a crash sur- rogate. While it is assumed that lane deviation is correlated to crash occurrence, the team is not aware of any studies that have proven this relationship. If a sufficient number of relevant crashes are available in the final data set, the team will devote some resources toward developing a model that relates lane deviation to lane-departure crash occurrence. Related measures, including time to lane crossing and yaw rate, have also been shown to be effective surrogates for lane departures. Several studies have used lane position as a surrogate mea- sure to assess the effectiveness of lane-departure counter- measures. Stimpson et al. (1977) identified lateral placement and speed as the best indicators for assessing driver behavior on horizontal curves. Donnell et al. (2006) used mean lat- eral vehicle position within the travel lane to assess the effect of wider-edge lines on horizontal curves on two-lane rural highways. Zador et al. (1987) used lane placement to assess the effect of post-mounted delineators and raised pavement markers on driver behavior at curves. Charlton (2007) used speed and lane position to compare the effectiveness of advance warning, delineation, and road marking treatments on horizontal curves. Finally, Porter et al. (2004), Taylor et al. (2005), and Hallmark et al. (2009) also used lane position to assess the effectiveness of lane-departure countermeasures. Rationale for Research Question and Use of Naturalistic Driving Data Importance of answerIng research QuestIon FHWA (2009) estimates that 58% of roadway fatalities are lane departures and 40% of fatalities are single-vehicle ROR crashes. Horizontal curves have been correlated with crash occurrence in a number of studies. Curves have approximately three times the crash rate of tangent sections. Seventy-six per- cent of curve-related fatal crashes are single vehicles leaving the roadway and either striking a fixed object or overturning; another 11% of curve-related crashes are head-on collisions (AASHTO 2008). Studies on roadway factors such as degree of curve, pres- ence of spirals, or shoulder width and type suggest that curve characteristics are the most relevant factors to crash occur- rence, but information is still lacking. In addition, little infor- mation is available that identifies which driver behaviors contribute to curve crashes. As a result, a better understand- ing of how drivers interact with various roadway features, such as curve radius or countermeasures like advance sig- nage, will provide valuable information to highway agen- cies in determining how resources can best be allocated to lane departure work plan) provides the most detail, and the subsequent plans highlight issue-specific considerations, such as alcohol-impaired driving. The research question related to crash surrogates is not a separate work plan but is incorporated into each example work plan. The crash surrogate question can of course be a research question in itself, but for the purpose of these examples, the team elected to focus on the various ways crash surrogates can be used for driver behavior and roadway improvements. With minor variations, each sample work plan follows the suggested six-part outline shown above. Explanatory com- ments from the S02 team are included in italics. These com- ments concern the main issues associated with each section of the proposal and the information that should be provided in each part of the work plan. example Work plan 1: Lane-Departure Crashes Overview of Research Topic This work plan focuses on the global topic area of “the influence of driver interactions with roadway features on lane-departure crashes.” This topic area was assigned a pri- ority one level on the Project S02 decision tree described in the Phase I report because research on driver interactions with roadway features leading to lane-departure crashes is relevant to safety, has the potential to reduce fatalities, would benefit from additional data sources, and is focused on driver behavior. Moreover, the behaviors are available in the natu- ralistic driving data set, and research outcomes may propose straightforward interventions, as well as provide broader insights into driving safety. Comment: This example summarizes one research approach to addressing a question within this topic area. It does not represent a complete or ideal work plan, nor does it advocate a preferred research methodology. Specific Research Question What roadway and driver factors influence the frequency and outcome of lane departures on horizontal curves? Crash Type(s) Addressed The ROR crash is the most likely crash type to result from right-side lane departures; however, head-on crashes and opposite-direction sideswipes can occur when a vehicle runs off to the right and then overcorrects in returning to the roadway. Head-on crashes and opposite-direction side- swipes are the typical crash types associated with left-side lane departures.

23 question. Specific tasks are not broken out, but the components could easily be divided into tasks. Data Sampling, Segmentation, and Aggregation Data reQuest Since the SHRP 2 NDS will generate a significant amount of data and because a disproportionate number of lane-departure crashes occur on two-lane roadways, the focus of the research will be curves on rural, dry, paved, two-lane roadways. Gårder (2006) indicated that two-thirds of all fatal crashes in Maine take place on rural, two-lane collector or arterial roads. Studies from other states have also indicated that a large number of lane-departure crashes occur on rural two-lane roads (ETSC 1995; Fitzpatrick et al. 2002; KTC 2006). FHWA (2009) also found that most lane-departure crashes occur on two-lane rural roadways. It is also relevant that a review of the most recently available information on the lane-positioning system that will be included with the SHRP 2 data acquisition sys- tem (DAS) suggests that the system does not perform well on unpaved roadways (Dingus et al. 2008a; Dingus et al. 2008b). As a result, the research team does not expect to be able to include unpaved roadways in the analysis. As to the focus on dry roadways, although adverse weather conditions increase the likelihood of lane-departure crashes, inclusion of varying weather and road conditions increases the complexity of the model and would require much larger sample sizes than can be included in the scope of this research. A review of data that will be collected at the NDS sites sug- gests that several of the sites are likely to include a large amount of rural road driving. (This is based on information available as of December 2009.) The team will therefore request data from the following sites: • Central Indiana will have 150 DAS units. The study area has 192 miles of rural principal arterial and 202 miles of rural minor arterial roadways. • Durham, North Carolina, will have 300 units. Although they do not distinguish between rural and urban in their study site description, they list 605 miles of primary road without limited access that are expected to include a large number of rural two-lane roadways. • Erie County, New York, will have 450 units. Although they do not list by rural and urban, the scheduled study site has 185 miles of primary road without limited access and 1,117 miles of secondary state and county highways that are expected to include a significant number of rural two- lane roadways. • Central Pennsylvania will have 150 units and has 568 miles of principal arterial and 734 miles of minor arterial road- ways, the majority of which are expected to be rural road- ways given the location of the study site. maximize driver performance and reduce the incidence and severity of crashes. The study results related to this research question should provide agencies with additional information to implement curve countermeasures or policies that will allow them to make better decisions to target resources in order to improve safety. Information that leads to a reduction in crash severity is a high priority for state highway agencies. ratIonale for use of nDs Data Crash data can be used to evaluate some factors related to curve crashes, but such analysis is limited by the amount and type of data provided in crash reports. Usually only aggregate information about roadway features is requested on crash forms. Even when more specific information might be avail- able, police officers may not choose to spend time collecting details on all roadway variables. Thus such data are incon- sistently reported. For example, an officer may code a crash as occurring on a curve, but will most likely not include any information about curve geometry or signage. In addition, minimal information on driver distraction is requested on police crash reports, and what is reported can be highly sub- jective. For example, an officer may or may not report that the driver was distracted by cell phone use, depending on whether that officer attempted to ascertain whether this was the case and whether it was a contributing factor. Other driver factors, such as driver forward attention, are not included in crash report data. A driving simulator could be used to assess how drivers interact with roadway features, since simulator data can pro- vide information on a specific driver’s performance. However, simulator studies do not represent normal driver behavior (e.g., natural engagement in driver distraction). In addition, addressing the wide variety of curve radii and varying road- way features that would be necessary to have representative samples would result in a very large and costly simulation study requiring a large number of drivers. Moreover, simula- tor studies, like crash data, do not yield exposure data. Thus, the use of NDS data was determined to be the best method to address this research question. Hypothesis to Be Tested Relationships exist between driver and roadway factors and the frequency and outcome of lane departures. To assess these relationships, the study will use lane deviation as a crash sur- rogate as described in the crash surrogate measures section. Data Analysis Plan Comment: The following sections outline the major components that would need to be included in a work plan for this research

24 possible to collect eye-tracking information. Since eye-tracking data are not available in naturalistic driving studies, forward scan position will be used as a proxy. DAS is expected to have some eyeglance positioning capabilities. Information on driver face position will be used to infer driver scan location. Bao and Boyle (2009) used driver scan behavior as a metric to assess age-related differences in how drivers perform various turning maneuvers at rural expressway intersections. They divided the forward view into seven scan locations (see Figure 6.1). A pro- tocol to measure location of driver forward scan position for each sampling period will be developed based on the method used by Bao and Boyle (2009). Comment: At this point the final accuracy and resolution of data from the various data sources have not been finalized. As a result, the accuracies desired for this sample work plan may not be available in the final data sets. If accuracies for certain elements are lower than desired, a determination will have to be made as to whether the accuracy is sufficient to answer the research question. The expected accuracy should be available for the S08 studies, and researchers should demonstrate that they understand what will be available and whether it is adequate to answer the specific research question they have posed. Database structure The database will be set up so that it can be shared with other researchers. Shared information will include a description of the data extraction, reduction, and processing methods used, as well as a data dictionary with an operational definition for each term or variable used. Irb reQuIrements Although the final requirements are not available, it is antici- pated that the team can meet the IRB requirements in order to obtain and use forward video, driver face video, and GPS data for the sections sampled. segmentatIon approach The sequential block data segmentation approach was selected as the most appropriate method for sampling data for clas- sification and regression tree (CART) analysis. Data will be sampled at 30-second intervals on tangent sections and at four points for each curve. Campbell et al. (2008) indicated that the driving task through a curve can be divided into four areas (approach, curve discovery, entry and negotiation, and exit). Each area requires different levels of attention and involves different driving tasks, so every curve will be sampled at each of the four points. Driver (e.g., distraction type, head posi- tion), vehicle (e.g., speed, acceleration, lane position), road- way (e.g., lane width, shoulder type and width, curve radius), and environmental (e.g., day, night) factors will be reduced from the corresponding data for 1 second for each sampling point. A 30-second sampling period for the tangent sections • Seattle, Washington, will have 450 units with a total of 174 miles of rural principal and 444 miles of minor rural arterial roadways. The Tampa, Florida, site has 450 units scheduled, but the site description specifies only 23 miles of rural principal arte- rial and 37 miles of rural minor arterial roadway. As a result, data will not be requested from the Tampa site. Comment: More detailed information about the amount and location of data to be collected and the schedule for data col­ lection will be available once NDS data collection begins and researchers from Project S04A prioritize collection of roadway data elements. S08 researchers should indicate that they under­ stand where roadway and NDS data will occur and what data collection schedule constraints may exist that will affect their ability to obtain data in a timely manner. extractIon of Data elements The data elements necessary to answer the research ques- tion will include roadway, vehicle, driver, and environmental characteristics. Table 6.1 indicates the necessary data elements and their expected source. The accuracy necessary for each data element is also provided. The majority of roadway data elements will be collected by using mobile mapping or will be gleaned from existing state databases. Depending on where mobile mapping data are col- lected and what other data sets are available, some data will need to be reduced from sources such as forward images or aerial images. Vehicle factors (e.g., speed, acceleration, spatial position, and lane position) will be provided by DAS. Driver face video, passive alcohol sensor data, and potentially some face tracking will also be available from DAS. All other driver factors will have to be reduced from the video. Because the analysis will include only dry road conditions, it will be necessary to determine a method for selecting only the desired environmental conditions, since roadway surface or ambient environmental conditions will not be provided by any of the data sources that are expected to be available. If possible, archived roadway weather information system (RWIS) data or other meteorological records may be used. Number and type of driver distractions will be extracted for each sampling interval. To ensure consistency among research team members, a protocol for extracting and cod- ing driver distractions will be developed based on the driver distraction coding system developed by VTTI for the 100-car NDS (Dingus et al. 2008a; Dingus et al. 2008b). Driver forward attention will be measured by the location of driver focus 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; Suh et al. 2006). The majority of these researchers conducted simulator studies in which it was

25 Table 6.1. Necessary Data Elements for Lane-Departure Work Plan Data Element Data Stream Minimum Vehicle Factors Latitude, longitude In-vehicle DAS ±6.6 ft Distance between vehicle and nearest strikeable object In-vehicle DAS ±6.6 ft Vehicle position from lane center DAS lane position tracking system ±0.1 ft Forward and lateral acceleration and speed In-vehicle DAS ±0.1 ft/s2 and 0.1 ft/s Pitch, roll, yaw In-vehicle DAS NA Roadway Factors Lane and shoulder widths Mobile mapping ±0.25 ft Roadway and shoulder surface types, number of lanes, presence and type of edge and centerline rumble strips Mobile mapping NA Horizontal and vertical curve lengths and radii, distance between successive curves, type and characteristics of curve spirals, curve start and end points Mobile mapping ±25 ft Superelevation, grade Mobile mapping ±0.5% Lane cross slope Mobile mapping ±0.1% Curve direction Will be extracted using DAS forward imagery NA Type and location of signage (e.g., chevrons), type and location of roadside objects Mobile mapping ±6.6 ft Pavement marking type and condition Extracted from DAS forward imagery NA Location and type of roadside objects Mobile mapping ±6.6 ft Speed limit and curve advisory speed Mobile mapping NA Exposure Factors Traffic volume (annual average daily traffic) State databases NA Time into trip Extracted from DAS NA Traffic density Extracted from DAS forward imagery NA Lane-departure crash data State databases NA Percentage of time driving on various roadway types under different conditions Extracted from DAS NA Driver Factors Age and gender Driver questionnaire NA Driver distraction Extracted from DAS driver videos NA Alcohol use Inferred from DAS NA Driver fatigue Extracted from driver video NA Driver forward attention Inferred from driver face tracking NA Note: DAS = data acquisition system; NA = not applicable.

26 The sequential block method was selected for several reasons. Continuous data segmentation would represent all driving situations and would provide a high level of confidence that meaningful patterns in the data could be detected. However, reduction of driver data at the continuous level is not prac- tical given the amount of data that is expected to be avail- able for rural two-lane paved roadways. A sample-based segmentation approach would result in data that were overly aggregated. An event-based approach could also be used. However, the purpose of data mining is to uncover patterns in the data, and an event-based approach focuses on pre- determined events that may prejudice the results. In addition, unless combined with another approach, an event-based approach may exclude drivers who did not engage in a pre- defined event. As a result, it was decided that the sequential block data segmentation approach provided sufficient ran- domness to uncover data patterns and was achievable given the time and resource constraints. Development of analytIcal tools The team will develop a processing tool since none is expected to be available. Model Formulation CART analysis, a data mining approach, will be used to address this research question and evaluate the data. This approach iteratively generates a tree structure by splitting the sample data set into two subsets based on a predictor variable and the value of that variable that produces the maximum reduction in variability. The algorithm will continue creating splits until some minimum criteria are met. Tree-based models are used for both classification and regression. A tree-based analysis uses a response variable (Y) that can be either quantitative or qualitative and a set of clas- sification or predictor variables (Xi) that may be a mixture of ordinal or nominal variables. For classification trees the response is categorical, and for regression trees the depen- dent variable is quantitative (Nagpual 2009). Classification and regression trees use algorithms to determine a set of if– then logical split conditions that divide the data into subsets. One of the advantages of regression tree analysis over tradi- tional regression analysis is that because it is a nonparamet- ric method and does not require assumptions of a particular distribution, it is more resistant to the effects of outliers, since splits usually occur at nonoutlier values. Tree models are nonlinear; that is, there are no assumptions about the underlying relationships between the response and explana- tory variables. In addition, independent variables do not have to be specified in advance. A regression tree selects only the most important indepen- dent variables—and values of those variables that result in was selected because reduction of driver video is expected to be time consuming and resource intensive. Sampling on tangent sections will be used to account for exposure. The oversampling of curves compared with tangent sections will be accounted for in the analysis. Vehicle variables can be extracted automatically by using an extraction tool that will be developed. Most roadway vari- ables will also be extracted from existing data (i.e., roadway data sets and mobile mapping data). Some roadway data may need to be extracted from the forward video or other sources such as aerial photos. However, at this point the amount of data reduction is unknown. Data reduction can be time consuming and depends on the time frame, the number of variables, and the amount of data filtering needed. For data sampled at 30-second intervals, data reduction time can be anywhere from 2 weeks to 2 months depending on the amount of data. The amount of data available on rural, dry, two-lane roadways in the NDS is unknown at this point, so an estimate cannot be made of the total resources that would be required to reduce the data. Depending on the amount of data available and project resources requested, a subset of the data meeting study conditions can be reduced. Comment: Although information on possible sample size and resource needs cannot be estimated at this point given the avail­ able information, S08 proposers should demonstrate that they understand what data are likely to be available and the resources that will be required to reduce sufficient data to ensure a statisti­ cally representative sample. JustIfIcatIon for samplIng approach Several different sampling approaches (e.g., continuous, sample based, or event based) could be used to extract data. 3 FAR RIGHT 1 FAR LEFT 2 CLOSE LEFT 4 CLOSE RIGHT 6 REARVIEW MIRROR 5 STRAIGHT AHEAD 7 OTHER Driver Figure 6.1. Seven visual scanning areas as defined by Bao and Boyle (2009).

27 were reduced from UMTRI’s lane-departure and collision- warning system FOT data. Odds ratios were calculated using logistic regression. A time–series analysis that used continu- ous data from UMTRI’s FOT data was also used to predict vehicle position according to a vehicle’s previous position and some roadway characteristics. Comment: CART analysis was selected for this sample work plan because the S02 team wanted to showcase a different analy­ sis method for each work plan presented in Chapter 6. The goal was to allow potential task force members and S08 proposers to see a variety of segmentation and analytical approaches that could be used to answer research questions. Other work plans presented in this report showcase time–series analysis, risk ratio, odds ratio, and crossover design. Data mining using CART analysis was selected for this first work plan because it was thought that data mining might not be as familiar to task force members as other methods. The authors acknowledge that this is certainly not the only valid approach for analysis of lane departures, and might not even be the best approach to answer the research question posed. Model Validation To assess how well the data mining models perform against real data, the data will be randomly partitioned into training and testing sets. The training set will be used to develop the model, and the testing set will be used to validate the model’s accuracy (how well the model correlates with the attributes in the data provided), reliability (how well the model works for multiple situations), and usefulness (whether the model provides useful information for the stakeholders). Pitfalls or Limitations That May Be Encountered and How to Address Them In order to represent a large number of lane-departure crashes and surrogates so that patterns in the data can emerge, a large amount of data is required, and reducing the data required at the level of 30-second intervals will be quite time consuming. When possible, the team will automate the process. Another limitation lies in whether the research team can demonstrate that lane deviation is a reliable surrogate for lane-departure crashes. In addition, data mining using CART analysis is not as commonly used as methods such as calculation of odds ratios or generalized linear models, and as a result it may not be well understood by the highway agencies that will need to use the information. Care will be taken in the dissemination of the project results so that stakeholders will have a general idea of how data are derived from data mining using CART analysis and specific knowledge of how such data can most reasonably be used. the maximum reduction in deviance—and does not require an assumption of best fit (Roberts et al. 1999). Regression tree analysis allows patterns in data to emerge that may not be uncovered using other approaches. Regres- sion tree analysis also reveals relationships between variables that may not be determined using other methods (StatSoft 2011). This method only allows variables to split at the value at which a correlation exists. For instance, shoulder width may only be relevant in determining whether a right-side lane departure results in a lane-departure crash on curves of a certain radius, and may be completely irrelevant for tangent sections or curves with larger radii. The response variable will be the amount of right- or left- side deviation from the lane center. Explanatory variables will include driver, roadway, vehicle, and environmental factors as discussed in the sections above on data extraction and segmentation approach. A sample analysis is presented in Figure 6.2; the data are plausible but hypothetical and are provided to illustrate how the method works. As shown, the hypothetical probability of having a right-side lane-departure conflict is related to curve radius, driver distraction level, and presence of edge-line rumble strips. Analytical Approach Various methodologies could be used to answer the research question and identify relationships between lane depar- tures on curves and driver and roadway factors. Jovanis et al. (2009) evaluated data from the VTTI 100-car study by using both event-based and driver-based approaches and generalized linear and Bayesian models. Gordon et al. (2009) used NDS data from existing FOTs to capture the associa- tion between highway factors, crashes, and driving behavior. They used Bayesian multivariate generalized models, SURs, and extreme value theory to test this association. Hallmark et al. (2009) used an event-based approach to model the rela- tionship between lane departures and roadway factors. Data Figure 6.2. Example of regression tree analysis.

28 of left- and right-turn lanes, signal phasing, roundabouts, pedestrian crossings, and signage. Specific Research Questions The research outlined in this proposal will answer the follow- ing specific research questions: How do intersection geomet- ric and operational features influence driver scan behavior and response? What is the effect of those influences on RLR violation and crash risk? Comment: This sample work plan is based on the assump­ tions that either the forward video includes the signal head suf­ ficiently far in advance so that the researcher can see if a driver ran a red light or that some other method will be available to identify instances of RLR. A review of the latest version of the forward video as shown in Figure 5.8 suggests that data reduc­ tionists will be able to determine signal head state, but at present the S02 team does not have sufficient information to determine whether a data reductionist would be able to identify RLR using only the forward video. A proposer who addressed the research question in this sample work plan would need to determine how RLR would be identified in the database, estimate the amount of resources to identify and reduce RLR events, and determine whether that method would be practical. Crash Type(s) Addressed Right-angle crashes are the most common crash type in RLR. Left-turn oncoming crash type, and opposite-direction side- swipe can also result from RLR. Some countermeasures, such as RLR camera enforcement, are believed to contribute to an increase in rear-end crashes. Proposed Surrogate Measures Even with the amount of data that will be provided by the full-scale NDS, it is expected that crashes will be rare events. RLR violation has been used as a safety surrogate for RLR crashes in a number of studies evaluating the safety impact of camera enforcement. Although a robust crash study requires several years of data after installation of the cameras, agencies often wish to evaluate the immediate impact of the cameras in order to justify their investment. As a result, a reduction in the number of RLR violations has been used by agencies as a safety performance measure (Bonneson et al. 2002; Fitzsim- mons et al. 2007; Retting et al. 2007). While it logically fol- lows that a strong correlation exists between RLR violations and RLR-related crashes, no studies were found that estab- lished a direct relationship between the two. RLR violations will be the basic crash surrogate for RLR crashes. An RLR violation is defined as a vehicle crossing the stop line after onset of the red phase. Time to conflict between Documentation of Results Project outcomes will be presented in a final report that will include a description of the data, data reduction, model for- mulation, analysis, results, and conclusions. The team will also prepare several two-page technical briefs geared toward nontechnical persons who may benefit from project results. Expected Impact or Outcome of Research on Countermeasures or Policy Implications This research is important because a large number of fatal crashes occur on curves. The results will aid agencies in understanding the relationship between driver behavior, including distractions, and curve negotiation. The results will also allow agencies to better understand which curve treatments result in fewer and less severe lane departures and will provide insight into which distractions have the most significant impact on the likelihood of a lane departure on a curve. Most highway agencies are proactive in implementing a range of countermeasures to reduce lane departures on curves, but they are hampered by having only limited infor- mation about the effectiveness of different countermeasures. The results of this research will provide more information about which specific roadway features are correlated to increased risk of lane departure. Study outcomes will also provide valuable information about how drivers interact with roadway features and how those interactions demon- strate the effectiveness of countermeasures. Understanding how drivers approach the task of negotiating curves, for example, will help to explain why certain countermeasures work. Increased understanding of the interactions of drivers with roadway features will allow agencies to make better deci- sion about selection of countermeasures. The research has implications for roadway design, selec- tion of sign type and placement, sight distance, and selection and application of countermeasures. It is expected that more appropriate application of countermeasures to mitigate ROR or head-on crashes on curves will result in fewer fatal crashes. example Work plan 2: Intersections and Crash Likelihood Overview of Research Topic Several issues associated with the interaction between driv- ers and the configuration and operation of intersections were observed in the S01 reports and the report for the S05 proj- ect, Design of the In­Vehicle Driving Behavior and Crash Risk Study (Antin et al. 2011). The questions related to the role of intersection characteristics include the effects on crash risk

29 information necessary to make the correct decision about whether to slow down or continue through the intersection during the yellow interval. The study hypothesis is that visual scanning behavior is a function of intersection geometry, operational factors, and driver factors and can be correlated to RLR crash risk. Data Analysis Plan Comment: The following sections provide a brief example of how a researcher might approach populating the work plan for this research question. A time–series analysis will be used to evaluate the data. Data will be sampled using a continuous segmentation approach. The data needed to answer the research question are listed, and how the data will be sampled, extracted, and aggregated for the time–series analysis is discussed. Data Sampling, Segmentation, and Aggregation Data neeDs The data variables necessary to answer the research question include intersection geometry, intersection operation, vehicle, and driver factors. Intersection factors include lane width and approach grade. Most roadway factors are expected to be avail- able from either existing databases or from mobile mapping data. Some data elements, such as signal head configuration and sight distance, may need to be reduced from the forward video. It will also be necessary to develop a metric that subjec- tively measures intersection level of service from the forward or other outward-facing videos. It is assumed that a database that indicates location of signalized intersections will be available from either existing data sets or from mobile mapping data sets. Signal timing, particularly clearance interval length, is an important factor in RLR. However, most agencies do not maintain a database of intersection timing, and even when this information is available, it is often out of date. Signal timing information will not be available from the mobile mapping data. As a result, there will be no way to obtain green or red phase lengths. The only option for obtaining clearance intervals will be from the forward video if the signal head is visible for the entire interval. The team will evaluate 100 RLR incidents when the data are first received to determine whether it is feasible to include clearance interval length. If a success rate of 80% is not achieved, clearance interval will not be included as a covariate. Seconds into the red interval (Rtime) can be determined from the forward video. Rtime will be calculated by measuring the amount of time that elapses from the time the signal head turns red until the vehicle crosses the intersection stop bar. The stop bar can be identified by locating a regular stop bar or pedestrian crosswalk or by observing the edge of the curb loca- tion in the forward video. A protocol will be established to mea- sure Rtime. Seconds into the red interval requires an accuracy of the study vehicle and an opposing vehicle or pedestrian once an RLR violation has occurred is the metric that will be used to evaluate risk. The model will attempt to derive a relation- ship between RLR crashes and time to conflict, if feasible. Rationale for Research Questions and Use of Naturalistic Driving Data Importance of answerIng research QuestIons FHWA (2006) estimates that RLR contributes to more than 100,000 crashes and 1,000 fatalities annually and results in an estimated economic loss of over $14 billion per year in the United States. Retting et al. (1995) found that occupant injuries occurred in 45% of RLR crashes as compared with other urban crashes and that such crashes accounted for 16% to 20% of total crashes at urban signalized intersections. In Iowa, RLR crashes are estimated to account for 35% of fatal and major injury crashes at signalized intersections (Hall- mark and McDonald 2007). ratIonale for use of nDs Data The goal of this project is to evaluate how drivers visually scan signalized intersections and what geometric, opera- tional, and driver factors result in diminished scanning that can potentially lead to RLR. Visual scanning is very impor- tant in understanding how drivers perform under certain situations. Drivers must process information from a number of sources, and most studies agree that visual information plays a significant role in how they perceive and respond to driving situations (Bao and Boyle 2009). Scanning behavior data cannot be obtained from crash data, nor can such data be easily captured in driver-in-the-loop sim- ulation because of the wide variety of factors (e.g., inattention, aggressive driving, insufficient sight distance) that contribute to RLR. Since a large number of factors may contribute to RLR, use of a simulator would require a large number of scenarios to isolate factors such as clearance interval length. In addition, some studies have shown that RLR is related to driver factors that may be hard to replicate in a simulator. For example, being in a hurry (particularly if a driver is late for work) was identi- fied as a factor in why drivers run red lights. However, it would be difficult to replicate being late in a simulator. Moreover, setting up a simulation to test a wide variety of intersection and signal timing configurations would be costly. Scanning behavior must be collected from within the vehi- cle, and only NDS data will provide the amount and types of data necessary for a careful examination of drivers’ scanning behavior relative to RLR. Hypothesis to Be Tested Certain intersection features may contribute to driver over- load that reduces a driver’s ability to perceive and process the

30 night lighted). Sight distance and field of view are decreased for nighttime conditions even when street lighting is pres- ent. Glare from oncoming vehicles, overhead street lighting, traffic signals, and other sources of light pollution may also significantly affect drivers. However, inclusion of light con- ditions would also significantly increase the complexity of the model and data needs; consequently the study will be limited to daytime hours. It should be noted that it is also difficult to account for the effect of glare on driver response and perception–reaction time. The research team assumes that a database that spatially locates all signalized intersections will be available for each study area. Vehicle traces will be requested for all signalized intersections within a study area that meet the geometry requirements and occur during daytime hours. The team expects that environmental conditions will not be reduced, so vehicle traces in which rain, snow, fog, ice, or other low- visibility conditions exist will need to be identified and dis- carded. Data will be requested for 500 feet upstream and 20 feet downstream for each 40+ mph intersection approach; this area will define the vehicle trace for this RLR study. Sev- eral researchers have identified an intersection area of influ- ence that extends beyond the boundary of the intersection itself. Upstream of an intersection, drivers must perceive and react to downstream conditions such as turning traffic, traf- fic signal changes, and queues. Although it is not simple to define the exact extent of the area of influence, studies suggest that an influence area of between 500 and 600 feet exists for a 40-mph approach (Stollof 2008; Stover and Koepke 2002). Certain age groups may be more or less likely to use four- lane arterials than others and will therefore be over- or under- represented. Older drivers, for instance, may be more likely to use lower-speed alternate roadways. Since not all instances of RLR will be included, the team will attempt to sample each age cohort as it appears in the population of drivers selected for the NDS. For example, if 5% of drivers are aged 18 to 25 years, the team will attempt to sample 18- to 25-year-old drivers at the same rate. Even with the constraints on geometry, time of day, weather conditions, and driver age cohorts, it is expected that a large number of vehicle traces will occur in the full NDS data set that meet the above criteria. A review of data that will be collected at the SHRP 2 NDS sites suggests that all but one of the sites is likely to include many signalized intersections. Based on this information, which was available as of January 2010, the team will request data from the following sites: • Tampa, Florida, has 450 units scheduled. The study area will include mostly urban roadways. • Central Indiana will have 150 DAS units. The study area has 380 miles of minor and other principal arterial roadways. at least 0.1 second. The GPS trace will not be sufficiently pre- cise to determine a vehicle’s position to that level of accuracy. Figure 5.8 shows the forward view from an interim version of the DAS. Although the final DAS will have some differences, it is assumed that details related to the signal head and phasing can be determined from the camera views. Driver distraction variables and driver scan behavior will be reduced from driver face video. Driver distraction will be coded using the methodology used by VTTI in the 100-car study (Dingus et al. 2006). Comment: As of January 2010, information was not avail­ able on the quality of driver face video, camera angle in relation­ ship to the driver’s face, field of view, and other parameters that would be necessary to determine how driver scan behavior can be extracted. Face tracking may also be available in the DAS. It will be important for S08 researchers to be familiar with the types of data that will be available to assess driver scan behavior and driver distraction and to provide a methodology for how they will define and extract such behaviors. Static vehicle characteristics (e.g., vehicle length and engine size), as well as dynamic vehicle characteristics (e.g., speed, acceleration, spatial position, and lane position), will be available from the DAS units in the SHRP 2 NDS. Table 6.2 indicates which data elements are necessary and their expected sources. The accuracy necessary for each data element is also provided. At the time this proposal was submitted the final accuracy and resolution of data from the various sources (e.g., Proj- ect S07 NDS and Project S04B) had not been finalized. As a result, the desired accuracies may not be available in the final data sets. If accuracies for certain elements are lower than desired, a determination will have to be made as to whether the accuracy is sufficient to answer the research question. Comment: The expected accuracy should be available for S08 researchers, and they should demonstrate that they understand what will be available and whether it is adequate to answer the specific research question they have posed. Data reQuest The study will focus on intersections on four-lane arterials with posted approach speeds of 40 mph or greater on dry roads during daylight conditions and with approach grades of less than ±4%. Although RLR can be dangerous in any situ- ation, the consequences are greater at higher speeds. Weather conditions can affect intersection crash risk and may contrib- ute to increased likelihood of RLR. Weather will also affect traffic operations. However, although weather may be an important factor, inclusion of environmental factors would increase the complexity of the model and the amount of data required. As a result, the model will be limited to dry roads. Likewise, it is assumed that driver behavior is affected by light conditions (e.g., daytime, dawn, dusk, night unlighted, or

31 Table 6.2. Necessary Data Elements for Intersection Work Plan Data Element Data Stream Minimum Vehicle Factors Vehicle length, center of gravity, acceleration capability, engine size Driver questionnaire NA Latitude, longitude In-vehicle DAS ±6.6 ft Distance to nearest vehicle or pedestrian crossing path In-vehicle DAS, forward video, forward or side radar ±6.6 ft Forward and lateral acceleration and speed In-vehicle DAS ±0.1 ft/s2 and 0.1 ft/s Pitch, roll, yaw In-vehicle DAS NA Distance from intersection stop line DAS, forward video ±1.0 ft Clearance interval time Forward video 0.1 s Rtime DAS and forward video 0.1 s Intersection Factors Lane width Mobile mapping ±0.25 ft Roadway and shoulder surface type, number of lanes, presence of bike lane, turn lane configuration Mobile mapping NA Approach grade Mobile mapping ±0.5% Approach speed limit Mobile mapping NA Presence, type, and condition of crosswalk Mobile mapping or forward video NA Sight distance to signal head Measured from forward video ±6.6 ft Signal head type and configuration Mobile mapping or forward video NA Clearance interval Forward video 0.1 s Exposure Factors Daily entering vehicles State databases NA Time into trip Extracted from DAS NA Traffic density Extracted from DAS forward imagery NA Intersection crash data State databases NA Percentage of time driving on four-lane arterials Extracted from DAS NA Driver Factors Age and gender Driver questionnaire NA Driver distraction Extracted from DAS driver videos NA Alcohol use Inferred from DAS NA Driver fatigue Extracted from driver video NA Driver scan behavior Inferred from driver face tracking NA

32 approximately 15 seconds, and calculation of time into red requires 30 seconds, for a total of about 45 seconds. The above estimates result in 312 + 15 + 30 = 357 seconds (5.95 minutes) to find and reduce intersection variables for a single RLR. Vehicle and other variables available in raw format from DAS (e.g., speed and lateral acceleration) will be retained at the 0.1-second resolution. Driver scan position will be reduced at an interval of 0.5 seconds; it will be assumed to be constant over the 0.5-second interval and can be mapped to 0.1-second intervals. The amount of eyeglance data to be coded depends on the first defining the relevant surrounding of the intersection. Coding data would be extracted approximately 500 feet (inter- section influence area) before the intersection, 48 feet while in the intersection, and an additional 20 feet after the intersection for a total of 568 feet. If the minimum speed is around 40 mph (approximately 59 ft/s), then it would take 9.6 seconds to travel through the designated area: 568 ft ÷ 59 ft/s = 9.6 s. From previous experience, it estimated that manual data reduction could take between five and six times the length of the actual video clip. As a result, driver data reduction will require approximately 9.6 s × 5 = 48 s per clip to code. Summarizing all the above, the total time to identify, char- acterize, and code is 312 s + 45 s + 48 s = 415 s. This includes time to scan traces in which an RLR event did not occur. In addition to this estimate, locating files within the databases and opening files will take more time. Estimates of coding should consider such peripheral logistics that also add time to each clip that is analyzed. Comment: The example is provided as an illustration that could be used to demonstrate that the proposer understands the amount of time and resources necessary for data reduction. Although it may not be necessary to provide this much detail, S08 researchers should provide some basis for why a data reduc­ tion interval was selected and should also make some estimate of how much time will be required to reduce data at the indicated level of segmentation. JustIfIcatIon for segmentatIon approach As time–series analysis was selected, a continuous approach was the only logical choice for data segmentation. It is acknowledged that this method will require a large amount of time for data reduction and will limit the number of samples that can be included in the model. Model Formulation A time–series analysis will be used to examine the propensity for RLR based on driver’s visual scan patterns. The main advan- tage of using time–series analysis for NDS data is that it allows one to model the driver’s scan behavior as the driver progress • Durham, North Carolina, will have 300 units, and the major- ity of the study area appears to be in an urban area. Although they do not describe their study site as either rural or urban, they list 605 miles of primary road without limited access. • Erie County, New York, will have 450 units and is located in a major urban area. Although they do not list by rural or urban, the scheduled study site has 185 miles of primary road without limited access. • Seattle, Washington, will have 450 units with a total of 1,958 miles of urban principal arterial roadways. The central Pennsylvania area is expected to have about 150 units and appears to be predominately rural. As a result, data will not be requested from the Pennsylvania site. Comment: More detailed information about the amount and location of data to be collected and the schedule for data collection will be available once the SHRP 2 NDS data collec­ tion begins and researchers from S04A prioritize collection of roadway data elements. S08 proposers should indicate that they understand where roadway and NDS data will be obtained and what potential data collection schedule constraints may affect their ability to obtain data in a timely manner. Database structure The database will be set up so that it can be shared with other researchers. Shared information will include a description of the data extraction, reduction, and processing methods used, as well as a data dictionary with an operational definition for each term or variable used. Irb reQuIrements Although the final requirements are not available, it is antici- pated that the team can meet the IRB requirements to obtain forward video, driver face video, and GPS data for the sec- tions sampled. Data segmentatIon anD reDuctIon approach Data will be extracted and used at the continuous level (col- lected at the frame rate). This is expected be 10 Hz or 0.1-second intervals. It is acknowledged that reducing data at this level of segmentation will require significant resources. The fol- lowing description provides a rough estimate of the time resources required to reduce a single vehicle trace. A review of a short video clip indicated that an experienced data reductionist would require approximately 10 seconds for each vehicle trace to determine whether a driver ran the red light. Retting et al. (2007) found that approximately 3.2% of drivers run red lights. Using the Retting data as a refer- ence, it will require approximately 1,000 seconds to find three red light running (RLR) vehicle traces, or 312 seconds per single RLR trace. Once an RLR event is identified, reduction of signal head configuration and sight distance requires

33 Documentation of Results Project outcomes will be presented in a report that will include a description of the data, data reduction, model formulation, analysis, results, and conclusions. The team will generate a white paper on the algorithm developed to flag RLR. The team will also make several presentations at national conferences. Expected Impact or Outcome of Research on Countermeasures or Policy Implications The main outcome will be information about which inter- section geometric, operational, and driver characteristics result in increased RLR and crash risk. This information will be useful for cities and other transportation agencies to make intersection improvements. For example, if the study demonstrated that yellow-phase length is correlated to RLR, traffic engineers could make recommendations for better signal-timing practices. This research has implications for intersection design, intersection signal timing and coordination, and for appli- cation of countermeasures such as all-red phasing or use of RLR camera enforcement. example Work plan 3: Driver Distraction and Crash Likelihood Overview of Research Topic Driver distraction has recently emerged as a high-profile driv- ing safety concern. The increasing popularity and complexity of electronic devices that are either built-in or carried into cars makes distraction an increasing threat to driving safety. Design- ers and legislatures work to balance convenience and access to information with driving safety, but much critical research regarding the risks posed by various types of distractions is still missing. In addition, sophisticated sensor systems may enable future vehicles to track drivers’ eye movements, identify dis- tracted drivers, and potentially warn drivers before mishaps occur. Naturalistic driving data can help identify the distrac- tions associated with different activities and provide the data necessary for the development of eye-tracking sensor systems. Specific Research Question What pattern of glances away from the road and steering wheel movements predicts breakdowns in lateral and longi- tudinal control? Crash Type(s) Addressed Rear-end crashes are most commonly associated with dis- traction, which was found to have contributed to 93% of through the intersection and to examine how changes in one time period affect the scan patterns in the next time period. Drivers’ scan patterns can be examined in two time peri- ods: before entering the intersection and while maneuvering through the intersection. The expectation is that the sequence of events in the first time period will provide insights into the likelihood of running a red light in the next time period. An autoregressive moving average approach may be appropriate to model and predict various types of driver behavior based on various scenarios. Analytical Approach A time–series analysis will be used to examine drivers as they encounter and go through a yellow phase as exhibited by speed and acceleration patterns. Speed and acceleration patterns in the vicinity of a signalized intersection that dif- fer significantly, indicating RLR, can then be related to inter- section geometry and operational factors. The model will be developed for dry roads during daylight hours to control for environmental factors. Data will be modeled for vehicles traveling through the intersection (not turning left or right). Crash risk will be modeled by time to conflict. Comment: The section above on issues related to time­dependent variables discusses issues related to time­dependent methods. As a result, model formulation is not further expanded on here. A time–series approach was selected because it was deter- mined to be the best method to account for dependencies in driver behaviors from previous time intervals. The main advantage of time–series analysis for naturalistic driving data is that it allows relationships between variables across time to be incorporated into the model. As a result, relationships such as driver distraction in previous time periods and probability of an RLR crash in a subsequent time period can be established. Comment: The authors acknowledge that this is certainly not the only valid approach for the research question and may not even be the best approach. Model Validation Approximately three-fourths of the data will be used to develop a time–series model. The remaining one-fourth of the data will be input to the model to determine how well it performs. Pitfalls or Limitations That May Be Encountered and How to Address Them It is expected that the request for vehicle activity in the vicin- ity of signalized intersections along major arterials will result in a significant amount of data to process. The time–series method may also be more difficult to present to lay persons at transportation agencies who are the most likely stakeholders to benefit from the results.

34 2009). While a great diversity of sources both within and out- side the vehicle account for these deaths (e.g., eating, groom- ing, other motorists, and billboards), the soaring popularity of in-vehicle information and entertainment systems (e.g., navigation systems, MP3 players, and cell phones) and the potentially distracting tasks introduced by these systems could increase the influence of distraction on driving safety. Although some tasks associated with these systems are driv- ing related, such as in the case of navigation systems, these tasks may interfere with safe driving and thus are still con- sidered distracting (Lee et al. 2008). Developing distraction countermeasures and reducing distraction-related crashes require an understanding of how the broad variety of second- ary tasks associated with new information technology affects driver behavior. ratIonale for use of nDs Data Over the last decade, hundreds of studies have investigated how distraction can undermine driver performance and safety. The vast majority of these studies involve experiments conducted in driving simulators. These studies carefully con- trol for confounding variables and provide a precise indicator of how distraction affects drivers’ performance in control- ling the vehicle. Generalizing the effect of distractions on performance in the simulator to driving safety on the road represents an important challenge. Drivers may adapt their driving and engagement in distracting activities on the road in a way that they do not in the simulator. Other studies have used epidemiological analysis of crash data to estimate the contribution of distraction to crashes. Understanding the effect of distraction on driving safety from such analysis is problematic because crash records may fail to identify the presence of a distraction. Moreover, crash data do not pro- vide a detailed description of the role of the distraction as a contributor to the crash. Naturalistic driving data help to fill the gap between simulator and crash data by providing a detailed account of the driver’s engagement in the distracting activity in the driver’s natural environment. Hypothesis to Be Tested Drivers’ glance patterns and steering behavior can indicate increased crash risk associated with breakdowns of both lat- eral control (e.g., lane-departure crashes) and longitudinal control (e.g., rear-end collisions). Both lateral and longitu- dinal events will be examined to assess whether distraction indicators from steering control can predict breakdown in longitudinal control. This relationship will be robust to dif- ferences in road type, distraction type, and driver age and gender. An extended interaction with a distraction will magnify crash risk defined by glance patterns and steering behavior. rear-end crashes in a recent NDS (Klauer et al. 2006). Less commonly, distraction also contributes to ROR crashes and head-on crashes. Although less frequent, ROR and head-on crashes are disproportionally responsible for fatalities and serious injuries. Recent crash data show that distraction con- tributed to 16% of fatal crashes and 21% of injury crashes (NHTSA 2009). Similarly, in a study of NDS data, Klauer et al. (2006) report that distraction contributed to 25% of all crash and near-crash events and approximately 65% of rear- end crashes and near crashes. Proposed Surrogate Measures Two general approaches to selecting crash surrogates will be employed to assess the sensitivity of the results to the choice of surrogate. The first approach considers driver response. In the case of rear-end crashes, a crash surrogate based on driver response would be a severe braking event. Severe could be operationalized as an absolute value, such as 0.5 g. Because maximum deceleration value is a function of initial speed, vehicle braking system characteristics, and driver response, a normalized value such as 99% for a given speed range would likely provide a more precise indicator of severe braking events. Dividing the speeds into ranges based on a maximum entropy function would guarantee a well-distributed set of ranges (Tan n.d.; Tan and Taniar 2007), but this choice would be vulnerable to any unequal distribution of rear-end crashes across speeds. To address this problem, a second approach uses information regarding close proximity to the vehicle ahead as a crash surrogate. Close could be operationalized using algorithms developed to trigger forward-collision warn- ing alerts. A simple forward-collision alert algorithm could be triggered by situations that cross a time-to-collision threshold of 2 seconds. A more complex algorithm might include the distance, speed, and acceleration of the two vehicles. Most likely these crash surrogates will co-occur—if a driver gets very close to the vehicle ahead, the driver is more likely to need to brake severely. Therefore, three analyses will be performed: one using the instances in which both surro- gates agree, one in which only severe braking occurs, and one in which the driver only gets dangerously close to the vehicle ahead. Assessing the type and degree of distraction associated with these three crash surrogate combinations might suggest different crash types and crash severity associated with differ- ent types of distraction. Rationale for Research Question and Use of Naturalistic Driving Data Importance of answerIng research QuestIon Distraction represents a clear threat to driving safety, account- ing for 5,870 deaths and 515,000 injuries in 2008 (NHTSA

35 of the driver and vehicle sampled for each 0.1 second of the epoch. Driver and epoch number will key a second database to this database. The second database will summarize the first by aggregating the data to the level of the epoch. Each row will represent a single epoch, and the eye gaze and steering data will be combined in several possible algorithms that can represent crash risk associated with different patterns of eye gaze and steering movements. Analytical Approach The statistical modeling involves two phases: the first com- bines gaze and steering data over the 180-second epoch pre- ceding the event to arrive at an index of expected risk. Long glances away from the road, short glances back to the road, a long interaction with a distraction, and lapses in steering followed by abrupt corrections might contribute to a higher degree of expected risk of a mishap. This index of expected risk can be derived from previous research or through data mining methods that identify the combination of variables and their weighting that best reflect the likelihood of a crash or near-crash event. The second phase evaluates the ability of this index to differentiate between crash and near-crash events. A conditional logistic regression model will calculate the odds ratios associated with the various risk indices. Risk indices associated with high odds ratios are those that accu- rately integrate gaze and steering data to predict distraction- related crash and near-crash events. Model Validation The statistical model will be validated with a sensitivity analysis that will examine the extent to which model predic- tions depend on the parameter values. The model will also be validated by assessing its performance by using a subset of the data that is withheld from the data used for model estimation. Pitfalls or Limitations That May Be Encountered and How to Address Them Video review and coding could be prohibitively expensive. Machine vision approaches to automatic gaze tracking are at the research prototype stage, and their output would require validation. Coding of distraction types would require man- ual coding. Currently available surrogate measures reflect abrupt responses of the driver in the form of braking and steering wheel reversals and do not capture lapses related to near-crash events such as failing to stop for a red light. The surrogate defined by driver response and by the lack of driver response outlined in this proposal begins to address this problem, but only for rear-end crashes. Data Analysis Plan Comment: The following sections provide a brief example of how a researcher might populate the work plan to address this research question. Data Sampling, Segmentation, and Aggregation An event-based sampling approach will be used to describe the 180-second period preceding the event and the 5 sec- onds following the event. The relatively long period pre- ceding the event will be used to assess the broad driving context leading up to the event and the contribution of the duration of the distraction to the likelihood of a crash or near-crash event. A case–control method will match the event-triggered sample as closely as possible with another sample selected at the same time of day on the same day of the week for the same driver in the week preceding that in which the event occurred. The continuous data will be aggregated using algorithms that maximize the ability of a sequence of eye movements to differentiate between dis- tracted and attentive drivers. The data variables necessary to answer the research question include road type (e.g., residential urban arterial, freeway), driver factors (e.g., age, gender), and driver behavior (e.g., steering wheel move- ment, speed, frequency and duration of off-road glances, distraction type). Driver distraction variables will be reduced from the driver’s video data. Ideally, on-road and in-vehicle gaze location information will be extracted with a machine vision algorithm. It is assumed that a database that links GPS location to road will be available either from existing data sets or from mobile mapping data sets. Data reQuest Epochs consisting of the 180 seconds preceding and the 5 seconds following each crash and near-crash event will be requested, along with three matching epochs. The epochs will be matched by driver, road type, time of day, and type of day (weekday versus weekend) rather than on a random selection to minimize extraneous variation and to identify the increased risk of distractions associated with the behavior of an individual rather than the overall safety profile of that individual. Database structure The database will be set up so that it can be shared with other researchers. Descriptions of data extraction, reduction, and processing methods, as well as a data dictionary with an operational definition for each term or variable used, will be provided. One database will include the continuous data for each epoch, including steering behavior, speed, and eye gaze location. Each row of this database will represent the state

36 to relate to crash likelihood in drowsy drivers. It is likely that crashes will be rare events in the naturalistic driving data, and ROR crashes may be even rarer. Hence, lane-departure and lateral drift events and standard deviation of lane posi- tion are considered as ROR crash surrogates for this specific research question. Other possible crash surrogates include lateral acceleration and speeding, which have been shown to relate to the likelihood of ROR events and safe negotiation on curves and through intersections (Reymond et al. 2001; Classen et al. 2007; Fildes et al. 2005). Each crash surrogate will be evaluated, and a determination of the best crash sur- rogate will be made after examining the available naturalistic driving data. Rationale for Research Question and Use of Naturalistic Driving Data Importance of answerIng research QuestIon Driver fatigue is a major contributor to motor vehicle crashes and is responsible for approximately 40,000 injuries and 1,500 deaths each year in the United States alone (Knipling and Wang 1995; Laube et al. 1998; Lyznicki et al. 1998). Royal (2003) estimated that 1.35 million drivers were involved in a fatigued driving–related crash over a 5-year period. A NHTSA study revealed that there are six million crashes annually resulting in an economic impact of over $230 bil- lion (Blincoe et al. 2002). Thus, over 4% of these costs are probably attributed to fatigue, and even this estimate may be low. In a separate study conducted by McCartt et al. (1996) approximately 55% of 1,000 drivers surveyed indicated they had driven while drowsy, and 23% had fallen asleep at the wheel. This confirms other findings that fatigue may play a role in crashes that are erroneously attributed to other causes (Connor et al. 2001). ratIonale for use of nDs Data Naturalistic driving data can provide insights into how fre- quently drivers exhibit safety-relevant errors while fatigued. Crash data are generally poor at identifying behavioral causes, and driving simulators cannot tap into how frequently or at what time of day or night such events occur. Crash data do not provide enough details to answer this specific research question since there is no preimpact information. Obviously crash data do not describe how frequently a fatigue event has occurred without a negative outcome (e.g., crash). There is also no way to observe speed, acceleration, and lane offset as the driver progresses on his or her trip or gets sleepier. Driving simulator studies can control for environmen- tal situations and can capture the performance of fatigued drivers by having them traverse over monotonous drives (Boyle et al. 2008; Reyner and Horne 1998). However, sleepi- ness can occur even during complex driving situations that Documentation of Results Project outcomes will be presented in a report that will include a description of the data, data reduction, model formulation, analysis, results, and conclusions. Beyond the standard report- ing of the overall results, a detailed description of the algorithms used for data segmentation and aggregation will be produced so that the process can be exactly duplicated. This description of the algorithms will be accompanied by intermediate data sets. Expected Impact or Outcome of Research on Countermeasures or Policy Implications The main outcome will be an algorithm that relates patterns of eyeglances and steering wheel reversals to crash risk. If such an algorithm predicts rather than coincides with crash risk, then it might be used as the basis for an in-vehicle counter- measure to mitigate driver distraction. To be useful, the algo- rithm must also be robust across different types of drivers and roadway situations. This research could identify effective algorithms for detect- ing distraction and thereby support interventions to prevent or mitigate distraction. The role of extended interactions with distractions could provide justification for greater legal sanctions associated with such behavior, such as those being adopted in Great Britain. example Work plan 4: Driver Fatigue and Crash Likelihood Overview of Research Topic Several issues associated with driver fatigue were observed in the S01 and S05 reports. The original questions related to the role of driver fatigue in various crash types (i.e., rear-end, head-on, backing, and lane change) and in crashes involving other vehicles, pedestrians, and other objects. Issues related to the influence of fatigue or drowsiness on driver behavior were also of concern. Specific Research Question How do episodes of fatigue affect drivers’ lane-keeping ability? Crash Type Addressed This study focuses on ROR crashes. Drivers’ lane-keeping ability is influenced by driver fatigue and sleepiness and hence may affect ROR crashes. Proposed Surrogate Measures This specific research question is based on examining crash surrogate measures (i.e., lane keeping) that have been shown

37 assumption of data independence would not be appropri- ate. Thus, a repeated measures analysis of variation and con- ditional logistic regression will be used depending on the nature of the dependent variables. Fatigued episodes will be identified from driver face video. Screening criteria for fatigue can include driver’s eye move- ments (e.g., eyelid closures for more than 2 seconds, multiple blinks), number of head-nodding events (Heitmann et al. 2001), and yawning. However, the research team notes that an examination of eye movements is highly dependent on the driver’s eye and eyelid geometries and whether sunglasses conceal the eyes. samplIng approaches The continuous data will be examined during the events when drowsy episodes are observed. Continuous data are necessary because they provide the only means for observing fine eye movements. A comparative sampling set will be needed for two nondrowsy episodes (or epochs) for the same driver. A 5-second sample of the vehicle kinematics will be used for the case and crossover events. This sample-based approach will help reduce the fluctuation and noise that are typically observed when examining raw data. Hence, calcula- tions of mean and standard deviation for vehicle kinematic information (e.g., speed, acceleration, lane position) will be potentially smoothed at this level. Data for all roadway types will be requested from the NDS. Video and driver (e.g., driver eye position, distraction type), vehicle (e.g., speed, acceleration, lane position), roadway (e.g., lane width, curve radius), and environmental (e.g., time, weather, lighting) factors will be used for the baseline and case episodes. The influence of sleepiness will be time dependent and will most likely degrade driver performance over the length of a trip. Since the effect of fatigue is continuous and may have an extended duration within a trip, the baseline episodes will come from separate trips from the same driver. Sampling both urban and rural roadways will allow the research team to compare the effect of fatigue under differ- ent road type and traffic conditions; it is expected that the driving environment will be more critical (i.e., there will be more traffic and the distance between vehicles will be closer) in urban areas. Database structure The database will be a flat file indexed by driver number, event index, and state of the driver. Each line will represent a summary of 5-second interval data, including numerical vehicle kinematic data (e.g., speed, acceleration, lane posi- tion), environment information (e.g., weather, roadway con- ditions, time of day), and reduced driver behavior data for that time interval (e.g., normal driving, fatigue). can increase the already high workload encountered by the sleepy driver. These complex and varied situations can only be observed in a naturalistic environment. Hypothesis to Be Tested It is hypothesized that drivers’ lateral control ability is affected by fatigue. By capturing the influence of fatigue on lateral control, insight can be gained on ROR crashes. Data Analysis Plan Comment: The following sections provide a brief example of how a researcher might approach populating the work plan for this research question. Data Sampling, Segmentation, and Aggregation The specific research question will be answered with data gathered from random and event epochs. The specific data will include roadway features such as road type, curve radius, and lane width; time of day; and weather and lighting. The majority of these elements would need to be provided from mobile mapping or state databases, as well as in-vehicle data collection. Weather and lighting information may need to be reduced from forward images. Vehicle kinematic data such as speed, acceleration, curve speed information, and lane position will be obtained from the in-vehicle DAS. Lane and roadway departures will be detected by automated lane position data from DAS. Driver face video and face and eye tracking would also be used and captured from DAS. If possible, evidence of eyelid closure (e.g., PERCLOS) would be determined. Model Formulation A case-crossover design will be used to compare cases (drivers during fatigued episodes) with control or baseline situations (drivers during nonfatigued episodes). In this analysis, each participant will serve as his or her own control, thus minimiz- ing confounding effects of age, gender, driving records, and other fixed characteristics (Maclure 1991). Thus, data from multiple baseline drives and events will be needed for each driver. If there are not enough fatigued episodes, the study can be set up to oversample the control condition such that a 2:1 matched approach can be used with one case (fatigue) episode matched to two control (nonfatigue) episodes, with all episodes based on the same roadway condition (urban or rural) for the same driver. It is important to note that because driving performance tends to be similar for each driver (i.e., within-driver data are highly correlated), using analysis measures based on an

38 research team will gain a better understanding of the role that fatigue plays in safety incidents and ROR crashes. Such data will also help determine what types of countermeasure techno logies are most effective. This research will have implications for the development of driving assistance systems, such as drowsy driver detection and alerting systems, as well as for driver education policies. example Work plan 5: Influence of Driver Impairment Caused by alcohol on Crash Likelihood Comment: This work plan demonstrates the issues and implica­ tions related to the use of the NDS alcohol sensor data. The alco­ hol sensor provides continuous sampling of cabin air and may have some utility in identifying driving segments where alcohol may be present. Several confounding factors include whether a window is open or the HVAC system of the vehicle is on refresh (i.e., if the cabin air is recirculated) and the possibility that the alcohol being detected stems from mouthwash or even perfume. Hence, the following example is focused on the challenges that can be encountered if a proposal includes alcohol sensor data. Data from other sensors can also produce ambiguous and noisy estimates (e.g., precision of lane position depends on the clarity of lane markings and other road textures), and each S08 pro­ poser needs to account for and manage these limitations. Specific Research Questions Do speed and variation in steering and speed differ when alcohol is detected in a vehicle compared with situations in which alcohol is not detected? If differences exist, how do they affect the likelihood of ROR crashes? Crash Type(s) Addressed All crash types related to speeding are relevant, most par- ticularly ROR crashes on both straight and curved roadway segments. Most ROR crashes occur on curves and are more typical in rural roadway environments (Liu and Subrama- nian 2009). Hence, this specific research question has high relevance to SHRP 2 safety improvement objectives. Proposed Surrogate Measures A key finding of the UMTRI S01 study was that yaw rate error could be a good surrogate for roadway departures on curves. As described by UMTRI, the yaw rate error generates a smooth, continuous, and unique data series, even when a lane boundary crossing occurs and appears to be a better pre- dictor than TTEC. Both these measures will be used as initial surrogates for ROR probability. However, their performance The database will be set up so that it can be shared with other researchers. Shared information will include a descrip- tion of data extraction, reduction, and processing methods, as well as a data dictionary with an operational definition for each term or variable that will be used. statIstIcal analysIs Repeated measures analysis of variance (ANOVA) will be used to analyze the speed and standard deviation of lane position. Conditional logistic regression will be used to analyze the like- lihood of lane departure and lateral drift incidents. Depend- ing on data availability, time of day, weather, lighting, and roadway features will be controlled to investigate the influ- ence of fatigue under different environmental conditions. Model Validation The statistical models will be validated with a sensitivity analy- sis that will examine the extent to which model predictions depend on the parameter values. The model will also be vali- dated by assessing its performance by using a subset of the data that will be withheld from the data used for model estimation. Pitfalls or Limitations That May Be Encountered and How to Address Them The raw video data will require extensive data reduction to capture driver fatigue. This process can be very time consum- ing and will add to the cost of this study. It is also important to note that distinctions between driver sleepiness, fatigue, and drowsiness will not be made because video data can only provide the analyst information on whether the driver appears to be tired. In addition, distinguishing between a driver’s fatigued and normal appearance using naturalistic driving data might be difficult and depends on individual eye and eyelid geometries (e.g., a driver with droopy eyelids may always look tired). However, the kinematic variables in combination with the video data will provide insights on the propensity of drivers to drive in an unsafe manner given the indicators of sleepiness. Documentation of Results Project outcomes will be presented in a report that will include a description of the data, data reduction, model for- mulation, analysis, results, and conclusions. Expected Impact or Outcome of Research on Countermeasures or Policy Implications The results of this research will provide insights in how fatigue could affect driving safety. By analyzing lateral control, the

39 Data Analysis Plan The data used for this analysis will be restricted to single- occupant vehicles (vehicles without passengers) to avoid any potential confounders from alcohol use by other occupants of the vehicle. Data Sampling and Aggregation Included in the data collected in the NDS are GPS location, the outcome of the alcohol trigger (alcohol presence), yaw rate, and steering and speed variability. Drivers tend to stay on familiar routes when traveling between home, work, and other routine destinations. Cap- turing driving during segments when no alcohol is detected on routine roadway links could be considered baseline (i.e., nonimpaired) driving. Aggregated baseline profiles for a par- ticular roadway link would be compared with instances of positive alcohol sensing on the same link. When a driver is impaired, the profile would change based on differences in speed, steering and/or yaw rate. Positive alcohol triggers on curves and straight road segments will be matched within each route to segments with no alcohol detection. Roadway, regional, and environmental characteristics (e.g., weather, day or night) will also be examined for each road segment and included in the model as covariates as appro- priate. Information on roadway characteristics (e.g., road type, lane width, cross slope, and type and width of shoul- der) will be provided by mobile mapping or will be available from existing state databases. These databases should be cross referenced with the vehicle GPS coordinates to identify fea- ture-based epochs of data upstream of the point of curvature (Campbell et al. 2008). Regions can be classified as rural and urban according to the vehicle GPS coordinates and state or national standards. Weather and other environmental condi- tions will most likely be extracted from the forward video. Comment: Other data sets may also be needed depending on available information (e.g., aerial images), and the proposer will need to identify where such information will be obtained. Model Formulation samplIng approach Epochs for each event will be collected for straight and curved segments with and without the detection of alcohol. Data for this specific research question will be reduced to 30-second blocks and aggregated across a road segment, depending on the length of the segment. Database structure The database will be set up so that it can be shared with other researchers. Shared information will include a description of data extraction, reduction, and processing methods, as well is based on a field operational test that considered only one vehicle type and whose primary focus was on evaluating a roadway departure system. Therefore, other variables will also be considered for crash surrogates in SHRP 2, and the outcomes will be compared to yaw rate error. Rationale for Research Questions and Use of Naturalistic Driving Data Importance of answerIng research QuestIons In 2006, there were 35,588 fatal crashes in the United States (NHTSA 2007). Twenty-eight percent of drivers (and motor- cycle operators) involved in fatal crashes failed to keep proper lane position or ran off the road (NHTSA 2007). ROR crashes are related to both alcohol and speeding (Liu and Subrama- nian 2009). In addition, alcohol and speeding are significant behavioral risk factors for other crash types (ETSC 1995; NHTSA 2001). Therefore, this research is relevant to the SHRP 2 traffic safety goals. The influence of alcohol is a behavioral crash factor of great concern given its impact on crash fatalities (NHTSA 2001). The research questions also identify the targeted form of driver error in terms of lane-keeping and selection of appropriate speeds (ETSC 1995). ratIonale for use of nDs Data Current research on the crash risk associated with alcohol use is based on epidemiological studies or derived from driving simu- lator or test track experiments. However, the data and conclu- sions from these studies are limited. Simulator and test track studies can identify the intervening behaviors that result in increased risk, but such studies involve artificial environments that lack the natural motivation factors inherent in the real world that might affect behavior such as speed choice. Driv- ers are less likely to misbehave in a driving simulator given the demand characteristics of a controlled experiment in which the driver knows he or she is being observed. Further, it is not ethi- cally possible to test alcohol involvement at the high blood alco- hol concentration (BAC) levels encountered in fatal crashes. Hypothesis to Be Tested Crash and near-crash events that occur when alcohol is detected are preceded by systematically different patterns of steering behavior and speed adjustment compared with those events when alcohol is not detected. Several shifts in driv- ing behavior can occur as a result of alcohol-related driver impairment: (1) increased, but less effective, steering behav- ior; (2) diminished steering input punctuated by lapses; or (3) increased speed combined with less effective steering. The analysis will assess the prominence of these potential alcohol- induced shifts in steering behavior.

40 Pitfalls or Limitations That May Be Encountered and How to Address Them Perfumes and other substances could falsely be identified as alcohol by the current NDS alcohol-detection system. Additional data coding resources may be needed to sepa- rate single- and multiple-occupant situations. Even though the proposed scenario is constrained to one occupant (i.e., the driver), it would still be difficult to determine whether the driver is under the influence of alcohol. Thus, video data will also be examined to assess whether the driver appears to have been drinking in the moments leading up to the episode being examined or whether he or she was using some other alcohol- based substance that was not ingested. Eliminating situations that might cause inaccurate indica- tions of alcohol presence might bias the data (few drivers will have their windows open in northern states in the winter), consume substantial effort, and diminish the sample size. However, according to current reports in the literature, there will be many instances of drivers drinking alcohol (NHTSA 2001); thus, a large-enough sample size to achieve adequate statistical power may be possible to mitigate these biases. Documentation of Results Project outcomes will be presented in a final report that will include a description of the data, data reduction, model for- mulation, analysis, results, and conclusions. Expected Impact or Outcome of Research on Countermeasures or Policy Implications The research is important because a large number of fatal alcohol-related crashes occur, with the highest rate of alcohol use and fatal crashes occurring in rural areas. This research will provide results that could further aid agencies in under- standing the relationship between alcohol consumption and driving safety. However, without BAC data, providing addi- tional support for alcohol policies that relate to the BAC level considered impaired and the potential methods for detecting impaired drivers from their driving performance measures is not possible. The profiles of behavior that significantly correlate with driver impairment can provide valuable information (1) to support methods for officers to detect impaired drivers from observed actions of vehicles and (2) to develop vehicle-based systems to monitor vehicle control and diagnose inferred impairment states. For example, real-time measures of vehicle speed in relation to posted speed limits can be used with intel- ligent speed adaptation systems that can automatically warn drivers and control speed for alcohol-impaired drivers. This study could, therefore, demonstrate the value of controlling the speed of cars that are being driven by alcohol-impaired drivers. as a data dictionary with an operational definition for each term or variable used. Based on the sampling approach and specified data, a processed database will be derived from the integration of the vehicle, roadway, and environmental data elements (see above). This processed database will have col- umns to represent the data shown below each epoch: • Trip information: date, trip number, and segment number; • Driver information: driver ID, driver face video, and eye point of gaze; • Vehicle information: 44 Alcohol presence or absence, 44 Speed variation on road link when alcohol is and is not detected, 44 Speed and steering variation on road link when alcohol is and is not detected, 44 Yaw rate when alcohol is and is not detected, 44 TTEC when alcohol is and is not detected, 44 Vehicle speed at curve averaged between 5 meters before and 5 meters after the point of curvature, and 44 Vehicle speed on roadway averaged between a distance 60 and 50 meters before the curve entrance; • Roadway information: 44 Road type, 44 Lanes (number, marking type, width), 44 Speed limit (posted and roadway), 44 Road curvature (horizontal, vertical), 44 Indicator for roadway speed >9 mph above the posted speed limit, 44 Indicator for curve entrance speed >9 mph above the posted speed limit; and • Regional information: urban or rural based on GPS. statIstIcal analysIs The central focus of this analysis is on the quantification of steering behavior and speed associated with situations in which alcohol is detected. These steering patterns can be defined in various ways. One approach is to describe the steering behavior in the time domain with traditional measures of steering perfor- mance, such as standard deviation or the frequency of steering reversals greater than some threshold. Another approach is to use a frequency domain description that describes the steering behavior in terms of a power spectrum using Fourier analysis or techniques such as wavelet analysis that are more robust to the characteristics of naturalistic driving data. Parameters extracted from the time and frequency domain analysis could then be evaluated with a cluster analysis to identify distinct types of steering behavior for alcohol and no-alcohol cases. Risk ratios applied to event counts (e.g., alcohol detected) and other forms of event-based analysis such as logistic regression are an appro- priate statistical method to test the association of alcohol with cluster membership (types of steering behavior) and cluster membership with crash and near-crash ROR events.

Next: Chapter 7 - Recommendations for Project S08 »
Integration of Analysis Methods and Development of Analysis Plan Get This Book
×
MyNAP members save 10% online.
Login or Register to save!
Download Free PDF

TRB’s second Strategic Highway Research Program (SHRP 2). Report S2-S02-RW-1:Integration of Analysis Methods and Development of Analysis Plan provides an analysis plan for the SHRP 2 Naturalistic Driving Study (NDS) to help guide the development of Project S08, Analysis of In-Vehicle Field Study Data and Countermeasure Implications, and to help assist researchers planning to use the SHRP 2 NDS data.

This publication is only available in electronic format.

  1. ×

    Welcome to OpenBook!

    You're looking at OpenBook, NAP.edu's online reading room since 1999. Based on feedback from you, our users, we've made some improvements that make it easier than ever to read thousands of publications on our website.

    Do you want to take a quick tour of the OpenBook's features?

    No Thanks Take a Tour »
  2. ×

    Show this book's table of contents, where you can jump to any chapter by name.

    « Back Next »
  3. ×

    ...or use these buttons to go back to the previous chapter or skip to the next one.

    « Back Next »
  4. ×

    Jump up to the previous page or down to the next one. Also, you can type in a page number and press Enter to go directly to that page in the book.

    « Back Next »
  5. ×

    To search the entire text of this book, type in your search term here and press Enter.

    « Back Next »
  6. ×

    Share a link to this book page on your preferred social network or via email.

    « Back Next »
  7. ×

    View our suggested citation for this chapter.

    « Back Next »
  8. ×

    Ready to take your reading offline? Click here to buy this book in print or download it as a free PDF, if available.

    « Back Next »
Stay Connected!