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.
9C h a p t e r 1 Background of Naturalistic Driving Studies Naturalistic driving experiments have been conducted for several years and include studies of drivers in their own vehi- cles and a series of technology tests to assess the safety conse- quences of advanced in-vehicle technologies. Generally, the 100-car study conducted by VTTI provided naturalistic data to make causal crash assessments, with a focus on the few seconds before and after crashes and events of interest (near crashes and critical incidents). UMTRI conducted an evalua- tion of a roadway departure and curve warning system as part of the U.S. Intelligent Transportation Systems program. While both studies were extensive, they did not focus on analysis methods per se. The S01 project sought to fill the analysis gap by focusing on analysis of existing data with an eye toward developing analysis tools for the 2,500-car study. This report describes examples of analyses conducted in exploring paradigms for naturalistic driving data analysis. The authors build on and begin to test some of the paradigms identified in the Phase 1 report and by Shankar and associates (2008). This report has four parts. The literature review that concludes this first chapter examines the literature concern- ing naturalistic driving studies and hierarchical statistical methods applied to traffic safety. Chapter 2 describes the research methodology and available data, and Chapter 3 dis- cusses the models estimated from the data. Chapter 4 sum- marizes the findings and their implications for SHRP 2 Safety projects and recommends future research. Literature review Naturalistic Driving Studies Stutts et al. (2005) unobtrusively collected video data from 70 volunteer participants driving their own vehicles over a period of 1 week. This study provided some of the first natu- ralistic data on driversâ exposure to potentially distracting events. The data were analyzed by the bootstrap percentile method of Mooney and Duval (1993) and provided some evidence that distractions can negatively affect driving per- formance, as measured by higher percentages of drivers hav- ing no hands on the steering wheel, their eyes directed inside rather than outside the vehicle, and their vehicles wandering in the travel lane or crossing into another travel lane. Teams of researchers at Virginia Tech have collected one of the most extensive sets of naturalistic driving data, includ- ing naturalistic driving of light-duty vehicles (Dingus et al. 2006; Klauer et al. 2006; Lee et al. 2004; Neale et al. 2005). In addition, VTTI has collected data and conducted numerous fatigue and drowsiness studies as part of a series of studies of driver drowsiness systems (Dingus, Neale, et al. 2006; Hanowski et al. 2005; Hanowski, Hickman, Fumero, et al. 2007; Hanowski, Hickman, Wierwille, et al. 2007). The 100-car naturalistic driving study database contains many extreme cases of driving behavior and performance, including severe fatigue, impair- ment, judgment error, risk taking, willingness to engage in secondary tasks, aggressive driving, and traffic violations (Neale et al. 2005). The data set includes approximately 2,000,000 vehicle miles, almost 43,000 hours of data, 241 primary and secondary drivers, 12 to 13 months of data collection for each vehicle, and data from a highly capable instrumenta- tion system that included five channels of video and vehicle kinematic sensors. Driver inattention was analyzed using the driving data set, and risk (ORs) was calculated using both crash and near- crash data, as well as normal baseline driving data, for various sources of inattention (Klauer et al. 2006). The risk percent- ages were also calculated to estimate the percentage of crashes and near crashes occurring in the population that resulted from inattention. Among the research involving naturalistic truck driving data (Dingus, Neale, et al. 2006) were cluster analysis studies of distraction-related incidents (Hanowski et al. 2005). Two pri- mary findings were that single drivers drive significantly more Introduction
10 curves and how they respond when presented with curve speed alert warnings. The study found that lateral accelera- tion was higher during the day than at night and also higher for right turns than left turns (both when evaluating lateral acceleration on a time-based average of the value when it exceeded a specific driverâs 90th percentile value and when the average of the maximum value for individual curves exceeded a driverâs 90th percentile value) (LeBlanc et al. 2006, pp. 8â28). However, no driver had a 90th percentile value of lateral acceleration greater than the 8.2 m/s2 nominal thresh- old for CSW alerts, and only two drivers had 90th percentile values greater than 7.2 m/s2 (LeBlanc et al. 2006, pp. 8â24). A combination of these two evaluations found that the avail- ability of CSW alerts did not have a dramatic effect on a driv- erâs chosen lateral acceleration (mathematically and physically related to longitudinal speed). From a fundamental perspec- tive, drivers may choose speeds at which to traverse horizontal roadway curves based on their comfort level of lateral accel- eration. They may not want to feel excessive force while tra- versing curves and will decrease their longitudinal speed to maintain comfortable conditions. Spacek (2005) examined the different types of paths drivers take while traversing horizontal curves and compared the fre- quency of the different types to the best possible path, which involves following the centerline of the lane perfectly through the entire curve. Using this ideal behavior as a baseline, there were five categories for comparison: normal (slightly cutting into the inside of the curve for a portion of traversal), correct- ing (reaching outside of the curve and overcompensating by turning harder toward the inside of the curve), cutting (strong cutting into the inside of the curve to counteract centripetal accelerationâa conscious process), swinging (starting toward the outside of the lane and finishing closer to the inside of the lane), and drifting (behavior opposite of swinging). The results of the study showed that, excluding undefined paths taken, cutting and normal behavior ranked first and second, respectively, in terms of percentage of track types taken for most curve radii (with a slight reversal of ranking for right- hand curves with radii greater than 65 m). Wilson et al. (2007) provided an in-depth analysis and evaluation of the UMTRI study (LeBlanc et al. 2006). Most of the major findings involving CSW alerts involved accelera- tion characteristics and speed approaching and during curve traversal. Vehicle speed approaching a curve was a major change-in-speed factor for triggering and reacting to alerts. There was a general positive correlation between approach speed and acceleration upon alerts being triggered. CSW alerts on ramps were mostly analyzed for exit ramps, as an alert is much more likely on exit ramps than entrance ramps due to the higher travel speed expected on limited access roadways. The beginning curvature of exit ramps played a role in actually eliciting CSW alerts; if the ramp began with aggressively than do team drivers, and that the frequency of critical incidents and fatigue-related critical incidents varied significantly by the hour of the day. The relationship between sleep quantity and involvement in critical incidents (crashes, near crashes, or crash-relevant conflicts) was studied using detailed sleep and driving data (Hanowski, Hickman, Fumero et al. 2007). Interactions between light-duty and heavy-duty vehicles used the same data for another targeted study (Hanowski, Hickman, Wierwille, et al. 2007). In addition to the previous light-duty vehicle and truck studies, research was conducted at VTTI concerning collision warning systems (McLaughlin et al. 2008). Seventy-three events were collected during actual driving. Data from the host vehicle, such as speed, yaw, acceleration, different con- trol states (e.g., brake pedal, turn signals), and measures of driver attention, were also collected. Other researchers have used more limited naturalistic driv- ing data sets to assess a range of safety and operations ques- tions. One study developed a model of lane-change duration for improving microscopic traffic flow simulation (Tijerina et al. 1999). Another study collected real-world driving data from a small sample of drivers to identify periods of drowsi- ness and inattention and validate drowsy driver detection algorithms. Data on driver exposure to environmental fac- tors and encounters with driving conflicts, near crashes, and actual crashes were used to characterize driver and vehicle performance, as well as the driving environment (Toledo and Zohar 2007). Researchers at UMTRI conducted a series of naturalistic driving studies as part of a series of field operational tests for the U.S. Intelligent Transportation Systems program (NHTSA 2005; Bogard et al. 1998; daSilva and Najm 2006; Ervin et al. 2005; Sayer et al. 2005; Sayer 2006). Systems tested included integrated forward collision warning and adaptive cruise control. Targeted studies of distraction and behavior (Ervin et al. 2005) found that variability of steering angle, mean and variability of lane position, mean and variability of throttle position, and variability of speed were affected by contextual factors such as road type, road curvature, and road condition. Conversing with passengers was the most common secondary behavior (15.3%), followed by grooming (6.5%) and using cellular phones (5.3%). This study found that the use of a cellular phone, eating or drinking, and grooming resulted in increased steering variance but did not affect lane position or speed variance. Another study (Sayer et al. 2005) quantified subjective reliability and performance of an in-vehicle warning system as a function of age, gender, weather conditions, light levels, and roadway classifications. UMTRIâs Road Departure Crash Warning System Field Operational Test: Methodology and Results (LeBlanc et al. 2006) evaluated the effects of CSW alerts on vehicle lateral acceleration in curves by exploring how drivers travel through
11 Fitzpatrick et al. (2000) also modeled acceleration in and near horizontal curves based on assumptions used in previ- ous FHWA research: ⢠All acceleration and deceleration occur outside the limits of the horizontal curve. ⢠Acceleration and deceleration rates are constant and equal to 0.85 m/s2. Tangents had to be kept above a certain minimum (244 m), and grades were kept close to 0% (maximum 5% downgrade or upgrade) to maintain conditions that allowed vehicles to reach the maximum possible speed on the tangent approaching the curve without other factors coming into play. Free-flow condi- tions were required (minimum headway of 5 s) as well. They found that speeds did not drop substantially until vehicles reached a point 200 m from the point of curvature of the given curve. They also found that the average acceleration rate in the 200-m zone before the curve was -0.1143 m/s2, which was significantly different from the previously assumed value of -0.85 m/s2. This rate ranged from 0.01 to 0.54 m/s2, which was primarily affected by curve radius. Acceleration within curve limits was found to be -0.0724 m/s2, which was significantly lower than the assumed value of -0.85 m/s2. The maximum observed positive and negative acceleration rates were signifi- cantly different from -0.85 m/s2. Thus, the assumptions from previous research were not appropriate for this particular study. Using these results, models were developed that showed that the drop in speed approaching a curve was generally inversely related to curve radius. Negative acceleration may begin to occur at vari- ous points upstream of curves, which may be due to site-specific differences. Thus, analysis of speed choices and acceleration rates at horizontal roadway curves should be carefully performed and possibly take into account site-specific differences, especially those surrounding curve radii and approach tangent lengths. Different forms of adaptation can occur on curves based on the presence of CSW alerts, personal comfort levels with respect to speed and how it affects the centripetal force the driver feels, comfort, curve perception, and the driverâs ability to maintain a reasonable path while traversing curves. How- ever, most of the analyses of adaptation involve the effects of longitudinal speed choices. Longitudinal speed appears to be the primary factor in assessing the level of danger approaching and traversing horizontal roadway curves, but it can be indi- rectly related to other factors mentioned here. The literature (see Table 1.1) largely confirms the rationale for the SHRP 2 naturalistic driving project (S08): the litera- ture analyzing naturalistic data is limited. Very little attention is paid to identifying paradigms (i.e., frameworks or organized structures) for analysis and the development of methods that evolve from those paradigms. The exploration of analysis paradigms is the focus of the Penn State research. curves with larger radii and progressed to smaller radii, the system would be able to detect over time that it was a ramp and not another roadway classification. However, some false alerts could be triggered if the initial curve on the ramp had a relatively small radius. The Design Quality Assurance Bureau of the New York State Department of Transportation (2003) discussed issues surrounding superelevation and how it relates to speed choices while traversing curves (Bonneson 2000). It is important to note that typical passenger cars will skid before rolling over during a turning movement, especially if the roadway sur- face is wet. Since a passenger car (the Nissan Altima) was used in the UMTRI study, it was assumed that any situation that triggered an alert may have resulted in a skid if necessary response maneuvers were not undertaken. Friction allows deceleration and steering forces to be transmitted from the tires to the roadway surface. The friction factor is used in place of the more common coefficient of friction as a ratio of the lateral forces that the pavement can resist from the vehi- cle. Changes in speed can reduce this friction factor, thus reducing the friction available for cornering, making curve traversal more difficult. This friction factor depends on vehicle speed and weight, tire conditions, and pavement conditions. However, speed is the most important variable in determining the friction factor, as it is the only variable that truly determines if a vehicle can safely traverse a curve under prevailing conditions. This makes speed likely the most important kinematic variable that should be evaluated in this analysis. Interpreting adaptation to alerts through speed changes does not necessarily account for the effects of traffic condi- tions, as certain roadway types have widely varying traffic volumes throughout the day. If higher volumes exist and headways decrease substantially, speed output in the data set will be influenced and may affect model results. Fitzpatrick et al. (2000) discussed speed prediction by recording speeds on two-lane rural highways; they only included vehicles that were at free-flow speeds (headway >5 s). Regression models were run to relate speed to several geometric variables, in- cluding horizontal curve radius (some models did not include horizontal curve radius as a predictor). Models that included radius as a predictor (always in the form of 1/R, where R = radius) had adjusted R2 values above 0.5. Regard- less of vertical geometry, 1/R had a strong correlation with speed (85th percentile speed) and was always significant, sometimes being the only significant predictor. Sufficiently large horizontal radii were considered by the researchers to be a condition that drivers would deem insufficiently severe to require speed reductions. If vertical alignment was consid- ered an important factor, it was considered the controlling factor in speed decisions for radii greater than 800 m. Sharp drops in speeds occurred for radii less than 250 m.
12 tion with respect to the vehicle operator. This allows one to estimate, separately for males and females, the effect of predic- tor variables on the dependent variable (crash, near crash, or critical incident). The hierarchy of the model allows one to structure the analysis to reflect more closely what actually hap- pens on the road. If one considers the crash event itself, a natu- ral hierarchy is to consider event and context variables at one level. These represent details of the situation at the immediate Hierarchical Modeling Methods Applied to Road Safety There are a number of advantages to applying hierarchical methods to naturalistic data. First, safety data frequently have natural hierarchies. For example, it is well known that males and females have important differences in crash etiology and outcome. One natural hierarchy is thus a gender differentia- Table 1.1. Related Reports from Literature Review Reference Research Objective Analysis Method Stutts et al. 2005 Study nature of driver distractions Frequency of distractions; bootstrap analyses of lane wanderings, lane encroachments, and sudden braking associated with distractions Hanowski et al. 2005 Study driver distraction in commercial vehicle operations Cluster analysis; cross-classification analysis Dingus, Neale, et al. 2006 Collect and analyze naturalistic data applied to driving fatigue Hazard analysis combined with analysis of variance Hanowski, Hickman, Fumero, et al. 2007 Identify and analyze light vehicleâheavy vehicle interactions Descriptive comparisons of percentages of critical incidents by category Hanowski, Hickman, Wierwille, et al. 2007 Quantify and analyze sleep of commercial vehicle drivers and associations with crashes Matched paired t-test McLaughlin et al. 2008 Analyze collision avoidance systems using naturalistic data Method constructed to test collision avoidance systems based on driver reaction and vehicle kinematics Bogard et al. 1998 Characterize safety and comfort issues of driver interactions with adaptive cruise control Histogram; descriptive statistics analysis Tijerina et al. 1999 Identify periods of driver drowsiness and inattention and validate drowsy driver detection algorithms Drowsy detection algorithm developed by Wierwille Dingus et al. 2006 Collect large-scale naturalistic driving data; define a near crash using quantitative measure Characterize driver behavior (e.g., driver inattention) and roadway environment as they relate to incidents, near crashes, and crashes Characterize changes in driver behavior over time with consideration of rear-end conflict and lane change as contributing factors Range/range rate approach applied to quantify a near crash (Kiefer et al. 2003); risk ratio applied to driver behavior change over time (Greensberg et al. 1993); estimation of Poisson rate per million vehicle miles traveled in relation to Heinrich triangle by scenarios (Heinrich et al. 1980) Lee et al. 2004 Characterize and analyze nature and severity of lane changes Analysis of variance (ANOVA); chi-square analyses Klauer et al. 2006 Characterize driver inattention using driving data collected in the 100-car naturalistic driving study Odds ratios Ervin et al. 2005 Analyze impact of integrated forward collision warning (FCW) and adaptive cruise control (ACC) systems on driver safety and acceptance Paired t-test; ANOVA Sayer et al. 2005 Determine frequency and conditions under which driv- ers engage in secondary behaviors; explore relation- ship between behaviors and driving performance using UMTRI RDCW field operational test data Mixed-model analysis of variance; autoregressive integrated moving average model (ARIMA) LeBlanc et al. 2006 Analyze suitability of RDCW system, which combines LDW and CSW functions Descriptive statistics analysis Sullivan et al. 2007 Examine how driver behavior is influenced by the reli- ability of an in-vehicle warning system using data derived from UMTRI RDCW field operational test Mixed-model analysis of variance
13 performed using a GLIMMIX macro in SAS software. The GLIMMIX macro employs a pseudolikelihood (Kim et al. 2007; Wolfinger and OâConnell 1993). This study suggested that incorporating naturalistic driving data (personal charac- teristics such as driver attentiveness, reaction times, vision, and aggressiveness and vehicle data such as braking character- istics, mass, steering characteristics, and tire condition) into the models may improve prediction accuracy. A hierarchical binomial logistic model has been developed (Wolfinger and OâConnell 1993) with a two-level specification: the response variable is dichotomous for high (fatal or severe) and low (slight or no) injury severity at the individual level (Level 1), and the crash level (Level 2) includes various crash features such as street lighting and road surface conditions. Multilevel negative binomial (NB) models have been used to capture the spatial variation of the effect of alcohol enforce- ment intensification (Yannis et al. 2008). The response vari- able is the number of road accidents with casualties, and the explanatory variables include the alcohol controls in Level 1 and socioeconomic parameters such as population in Level 2 (different regions). The parameter estimates were also obtained through the iterative process of restrictive iterative general- ized least squares and quasi-likelihood implemented using MLwiN software (Rasbash et al. 2000). Summary In summary, a series of naturalistic driving studies has been conducted for both light-duty and heavy-duty vehicles. Some modeling has been conducted with multiple predictor vari- ables, but no hierarchical models have been applied to the data. Existing hierarchical applications include several studies of injury severity and two studies, one focusing on enforcement and the other on crash type, but no applications to naturalistic data were found. Shankar and associates (2008) argue that naturalistic data analysis would benefit greatly from the appli- cation of hierarchical methods because, among other reasons ⢠The functional form for models is not well documented, ⢠Sample size limitations may hinder frequentist approaches, and ⢠Driver, event, and context variables are known in the data, but their interrelationship in crash modeling is largely untested. Chapters 2 and 3 focus on empirical data analysis: Chapter 2 describes the data and analysis approaches used with each data set, and Chapter 3 summarizes the results of the model- ing. Chapter 4 provides an overall summary of the study linked specifically to the five Penn State research questions and their implications for the SHRP 2 Safety program. The last portion of Chapter 4 summarizes lessons learned along with sugges- tions for future research. time surrounding the crash event. A second level of consider- ation could be driver attributes, such as years with a driverâs license (i.e., driving experience), which operates over a longer time period. These examples illustrate the value of hierarchy: to better represent the reality one is seeking to model. Yet another advantage of hierarchical models is that they allow flexibility in model structure; this feature is particularly appealing when initially exploring new data sets for which model structure is unknown or not well defined. The Phase 1 report noted that in this context, hierarchical methods allow better exploration of variable effects with alternative struc- tures. This is a common theme touched on in Chapter 3. Computational advances have facilitated the use of specific hierarchical methods, but from the perspective of this report the principal advantages are flexibility and facilitated data exploration. There have been several applications of hierarchical mod- els in the analysis of crash frequency and severity level (Gold- stein 1995; Rasbash et al. 2002; Sullivan et al. 2007; Wolfinger and OâConnell 1993), and there is an additional application to law enforcement (Yannis et al. 2008). Iterative generalized least squares (Jones and Jorgensen 2003) was employed to fit a binary logistic regression model in which the response vari- able indicated whether each casualty survived with serious injuries (response = 0) or died (response = 1). When normal- ity is assumed, however, the full Bayesian estimation and empirical Bayes, treating the prior distribution as known, are the same as iterative generalized least squares. Data were ana- lyzed using a hierarchy of casualties (Level 1), within acci- dents (Level 2), and within municipalities (Level 3). Another study (Goldstein 1995) compared the efficiency of multilevel logistic models (MLMs) by using maximum likelihood, generalized estimating equation models, and logistic models using simulated French road crash data between 1996 and 2000. The hierarchical structure was modeled as the prob- ability that an occupant (Level 1) in a car (Level 2) during a crash (Level 3) died; the response variable, severity, is treated as binary. The MLM was the most efficient model, while both generalized estimating equation models and logistic models underestimated parameters and confidence intervals (CIs). MLM estimates were obtained through the iterative process of restrictive iterative generalized least squares and penalized quasi-likelihood imple- mented using MLwiN software (Rasbash et al. 2000). Hierarchical binomial logistic models (Rasbash et al. 2000) were used to solve the suspected heterogeneity in underlying causal mechanisms associated with different crash types. This study built models for angle, rear-end, and sideswipe crashes with the response variable of crash probabilities and data con- sisting of two levels: Level 1 consisted of crash-level character- istics, and Level 2 consisted of intersection-level characteristics from 91 two-lane rural intersections in the State of Georgia. The estimation of multilevel binomial logistic models was
