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.
20 CHAPTER 5 Safer Glances, Driver Inattention, and Crash Risk: An Investigation Using the SHRP 2 Naturalistic Driving Study T. Victor, J. Bärgman, M. Dozza, H. Rootzén, J.D. Lee, C. Ahlström, O. Bagdadi, J. Engström, D. Zholud, and M. Ljung-Aust (SAFER Vehicle and Traffic Safety Centre at Chalmers) Introduction Driver inattention is a long-standing major factor related to fatality in motor vehicle crashes (Evans 2004). It is also a flourishing problem associated with the communication, information, and entertainment technology that is transforming the car and modern portable technology such as the smartphone (NHTSA 2010a). In 2009 distraction was involved in crashes causing 5,474 deaths and leading to 448,000 traffic injuries across the United States (NHTSA 2010b). Consequently, there is a critical need to better understand distraction and the limits of attention while driving. Driver inattention is very high on the national traffic safety agenda (see www.distraction.gov). There is an increasing amount of legislation, regulation, design guidelines, and information campaigns related to driver inattention. The vehicle and electronics industries are moving fast to respond to both enable the use of electronic functionality in a safe manner and to reduce driver inattention through safety systems which are capable of monitoring it. Two main developments have combined in the past few years to create this escalation: (1) there is a growing concern over the driving-compatibility of the ever-increasing functionality available through electronic devices (such as smartphones and intelligent vehicle systems), and (2) research has been showing a much clearer association between driver inattention and crash risk. State of the Art on Dangerous Glance Behavior The specific mechanisms and indicators of the risk of inattention are not definitively quantified. The main source of a clearer association between driver inattention and crash risk has been naturalistic driving studies such as the 100-car study (Dingus et al. 2006). Initial analyses have been able to identify the crash risk associated with distracting activities or tasks, such as texting, dialing, and eating; however this approach does not explain why the tasks are dangerous nor provide the objective inattention performance information needed for guidelines and inattention countermeasures while driving. Subsequent analyses of the 100-car eyeglance data show that high total glance times (e.g., 2 seconds or more in a 6 second period) are associated with increased crash/near-crash risk (Klauer et al. 2006; 2010). Liang, et al (2012) compared 24 different ways to combine various glance characteristics, such as single glance duration, glance history, and glance location. They
21 found that single off-road glance duration was the best crash predictor. Glance history (such as total glance time) and glance location did not improve risk estimation above single glance duration but they were still predictive of crash/near-crash risk. Further analyses of the 100-car data have revealed that risk is pinpointed to the timing of off-road glances in relation to external events. Risk is primarily associated with an inopportune single glance duration (Victor and Dozza 2011; Victor, Dozza, and Lee, forthcoming). The most sensitive measures for risk are those which most precisely quantify an off-road glance that overlaps a change in the state of environment or action that began the sequence leading the crash or near crash called the precipitating event (e.g., a lead vehicle that begins braking). The longer the driver looks away from the road at this specific time, the greater the risk. Naturalistic Driving Data and Validated Inattention Crash Risk Naturalistic driving data is valuable in comparison to driving simulator and field experiments (see NHTSA 2012) because it is able to quantify real crash risk. However, previous naturalistic data have been limited by the number of crashes collected and by the quality of data (e.g., radar data). The SHRP 2 naturalistic data can provide the data that is needed to achieve validated inattention performance measures associated with pre-crash situations. The data are essential to improve the understanding of driver inattention, for guidelines to reduce distraction from electronics devices, for countermeasures that detect and act to reduce distraction while driving, and for regulation and education. Analysis Plan This research seeks to develop a statistically validated set of inattention-risk functions (or relationships) describing how an increase in inattention performance variables in lead-vehicle pre-crash scenarios leads to increased risk. Likewise, these relationships identify which glance behaviors are safer than others. The events needed for this analysis are lead-vehicle crashes, near crashes and associated comparison conditions (baselines). The key data needed for each event are eyeglance variables, variables that describe the visual information that drivers rely on to initiate and control braking (called optical parameters), and measures of event severity, or how serious a crash or near crash is. The analysis plan is devised as an extension of work that has been done previously by the researchers in this project (Liang et al, 2012; Victor and Dozza, 2011; and Jonasson and Rootzén, forthcoming) on the 100-car dataset. These previous analyses provide the requirements for the analysis plan, including a data-driven sample size assessment. The plan is adaptive to accommodate data and sample size limitations. The analysis plan is formulated in five analytic steps. Each analytic step is expected to provide better precision and explore different components of the inattention-risk relationship by providing more detail on inattention-risk relationships under different circumstances â
22 relationships to timing with optical parameters, glance characteristics, and relationships with respect to different levels of crash severity. In fact, each of the analytic steps provides one or more inattention-risk functions for each step of the analysis. No major changes are planned for Phase 2. The analysis starts with Step 1, which replicates and extends previous research. The results of Step 1 represent an improvement over previous work because they identify a more precise relationship between glance patterns and crash risk for the specific rear-end pre-crash scenarios. Step 2 places these glance patterns and their associated odds ratios in the context of a theoretical explanation of how particular glance patterns increase crash risk: the sweet spot. The (perceptual motor) sweet spot indicates the time when perceptual information is particularly valuable in crash avoidance and when a glance away from the road would be particularly risky. Step 3 integrates the glance measures, such as duration of glances away from the road that overlap the sweet spot, as well as other measures such as the glance history. Step 4 builds on Step 3 by quantifying the injury severity associated with a crash and so creates a more precise assessment of inattention-related risk. Each of the preceding analytic steps represents contributing components, together building a more precise inattention-risk relationship. Step 4 relates glance behavior to injury severity as defined by new severity scales. High severity crashes represent more likely loss of life and disabling injuries: the ultimate risk posed by distraction. Step 5 highlights the importance of going beyond a simple inattention-risk function to consider inattention-risk relationships, represented here as a set of inattention-risk functions. Step 5 provides a family of statistical relationships that indicate crash likelihood or injury severity given contextual characteristics of the crash situation, such as traffic density, road type, and speed. When integrated, the five analytic steps provide more precise knowledge about the relationship between inattention and risk than each of these components separately. This approach will deliver progressively more precise inattention-risk functions (or relationships) at each analytic step. This approach is therefore also adaptive to the outcomes of the analyses, as it can stop when further analysis does not provide âstrongerâ inattention-risk relationships or when the sample size does not permit the analysis. For example, if the sample size is sufficiently large it is possible to analyze the inattention-risk relationship in different conditions of traffic density, road types, speed, and distraction types. Each analytic step provides an important component that will contribute to an increasingly better understanding of the relationship between inattention and risk of crashes and the severity of the associated injuries.
23 Data In total, 13 events were obtained from the NDS database: 4 rear-end crashes, 5 near crashes, and four baseline events. Three crashes were found because they were reported. One crash and all the four near crashes were found using a â1g acceleration trigger. The four baselines were selected randomly from the same trips where the crashes occurred. The baselines were selected to match the criteria that there should be a lead vehicle present and that the driverâs vehicle should be moving a majority of the time. The main data used are glance variables (total glances, overlapping glance, last glances, and total preceding glances), optical parameters (optical size and expansion rate of the lead vehicle, and Tau, which is a combination of the two), and two measures of event severity (calculated from time-to-collision, relative speed, the mass of vehicles involved, start of an evasive maneuver, or when there would be a crash). Results All variables that were needed to be able to conduct the analyses as defined in the analysis plan above were successfully calculated for the 13 events. Data quality difficulties were overcome. For example, an approach to use lead vehicle size coding from video to fill out the gaps in the radar data provided good range and range-rate data for crashes (where the data were often missing close to crash). The main lesson learned was that video has proven to be extremely useful for understanding and validating many measures, but most importantly for filling in radar data gaps. For the replication analysis, aggregated glance data plots show results that are in line with expectations (the 100-car data). They demonstrate a technical realization that improves upon previous methods by including a longer period that includes crash points. For the sweet spot analysis, the optical parameters were successfully objectively quantified from radar and video data (see Figure 5.1). Further case studies, plotting the glances in all events in relation to inverse Tau, show an improvement upon previous methods by realigning data to the crash point and minimum distance. Data also show that there may be a difference in glance behavior between crashes and near crashes.
24 Figure 5.1. Optical variables for one crash event (2880551) on left, and one near-crash event (2880553) on right. A hypothetical spatio-temporal sweet spot is illustrated with heat maps. The glance characteristics analysis focused on an analysis of whether glances could be automatically calculated by using the head tracker sensor data. The analysis showed clearly that it is not recommended. The head tracking is often lost, and when data is available, it is often erroneous or misleading. Manually reduced eyeglances are on the other hand highly reliable, feasible, and efficient.
25 Note: Crashes = O; near crashes = X. Figure 5.2. The three types of severity measures M-SEC, MSDeltaV, and the SHRP 2 event severity variable. Two new scalar severity measures, M-SEC and MSDeltaV, were implemented as an alternative to the traditional severity-categorical approach (Figure 5.2). Plots of the scales reveal examples of how near crashes can be more severe than crashes, demonstrate the implementation feasibility, and how the scales can be better suited for statistical analyses. As the last step, an expression of the set of inattention-risk functions was developed: Pi (Severity | Context) = Fi (Sweet Spot, Glance Characteristics) where i denotes different sets and combinations of Context parameters such as traffic density, road type, SV speed and distraction type. Different contexts will provide different probability functions. Further, two approaches (Extreme Value Theory and Context Dependent Risk) were developed to make it possible to extrapolate from behavior in near crashes to crashes, and from less severe crashes to more severe ones. Extreme Value Theory is also used to detect and correct for the selection biases in the data. Expected Results and Implications for Countermeasures The main scientific result of this research is a quantitative relationship between inattention and risk. A set of inattention-risk functions, or family of statistical relationships, will provide more
26 precise knowledge as a whole than individually. This research will identify a more precise relationship between glance patterns and their associated risk around a sweet spot, a time when perceptual information is particularly valuable in crash avoidance. Further, it will relate glance behavior to injury severity as defined by new severity scales. This set of functions will indicate crash likelihood and/or injury severity for certain contextual characteristics of the lead-vehicle crash scenario, such as traffic density, road type, and speed. These relationships can be used to show more precisely which glance behaviors are safer than others. For example, this research can be used to show how much the risk of a serious injury can be reduced when tuning a radio by changing the series of single glances to be shorter, and can relate this net benefit to a potential cost of increasing the number of glances needed to finish tuning the radio. It can determine how this risk varies with different types of contexts (e.g., stop- and-go versus free-flowing traffic), can determine the point in time where the eyes are needed most for the control of braking, and can be used to understand the type of glance behavior that leads to crashes as opposed to near crashes. Safer glance strategies for interacting with electronics and the traffic environment can be encouraged in a number of ways including design guidelines, education, and in-vehicle feedback. Likewise, the most dangerous glances can be pinpointed and associated with improvements to appropriate countermeasures such as distraction guideline performance criteria and active safety system technology. These results can address current limitations in scientific knowledge regarding driver distraction guidelines for in-vehicle electronic devices. More evidence is needed to provide valid assessment and compliance criteria for performance testing. Project results can be used to support evidence-based distraction policy and regulations, and can be used to teach safe glance behaviors. Systems that detect inattention while driving are highly prioritized for intelligent vehicle safety systems (NHTSA 2010a). This research will improve safety systems such as Forward Collision Warning Systems to make them inattention-adaptive. It will reduce nuisance warnings and warn more exactly when the risk is greatest. Further, this research will greatly improve how distraction and inattention is detected because the inattention-risk functions directly describe what a system should be looking for. Distraction feedback and warnings can more appropriately be given and driver coaching feedback is improved.
