Initial Analyses from the SHRP 2 Naturalistic Driving Study: Addressing Driver Performance and Behavior in Traffic Safety (2013)

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.

15 CHAPTER 4 Car Following, Driver Distraction, and Capacity-Reducing Crashes on Congested Freeways: An Investigation Using the SHRP 2 Naturalistic Driving Study Gary A. Davis, John Hourdos, and Indrajit Chatterjee (University of Minnesota) Introduction It has been estimated that about 50% of the congestion experienced on urban roadways is due to nonrecurring causes, with capacity-reducing incidents being responsible for about 25% (FHWA 2012). One prominent type of incident is the freeway rear-ending crash. Although these crashes result in serious injuries on occasion, they also generate substantial social costs through delays imposed on other freeway users by reducing available freeway capacity at critical times. Reducing the frequency of these crashes should help achieve better use of an increasingly strained urban freeway system. At present, however, there is a shortage of countermeasures that have been proven effective in reducing freeway rear-ending crashes, although a range of potential countermeasures are currently being considered. Empirical evidence indicates that freeway rear-ending crashes tend to occur when stopping waves form on freeways, while a theoretical analysis suggests that the occurrence of “critical events,” in which following drivers brake at rates substantially higher than their leaders, should play an important role in determining whether stopping waves result in crashes. This is because harder braking by a following driver tends to reduce the stopping distance available to the follower’s followers, making avoidance of a crash successively more difficult (Brill 1972). An important insight from Brill’s model is the relationship among reaction time, following headway, and safe braking rate. Drivers whose reaction times exceed their following headways need to brake harder than their leaders, while drivers whose following headways exceed their reaction times can brake more gently. Proposed Analysis The SHRP 2 NDS offers a unique and potentially powerful source of data relevant to understanding driver behavior in the field, and the goal of this project is to advance the understanding of how drivers behave when confronted with a freeway stopping wave. This will be done by extracting from the NDS a sample of relevant events and then reducing the data for these events to produce estimates that characterize the situations confronting drivers as well as their reactions to these situations. This reduced dataset will then be used to address three general questions: 1. What is the role of drivers’ allocation of attention—especially of potentially distracting activities—in determining whether or not a braking event is critical?

16 2. What is the potential for an idealized driver-assist system to reduce the likelihood of critical events? 3. What impact on driver behavior would an idealized roadway-based system need to achieve to realize a target reduction in the likelihood of critical events? Phase 1 Objectives During Phase 2 the research team proposes to construct, from the NDS, a dataset of approximately 250 freeway trips containing brake-to-stop events. Using NDS radar, speedometer, forward video, and GPS data, the team will then compute estimates of features such the initial speeds of the NDS vehicle and its leader, the braking decelerations used by the two drivers, and the follower’s reaction time and following distance. The work during Phase 1 focused on four questions related to the feasibility of this plan: 1. Can freeway brake-to-stop events be identified efficiently from the NDS database? 2. Can the NDS in-vehicle video be used to characterize driver actions and identify intervals of possible distraction? 3. Can quantitative features of brake-to-stop events be estimated from the NDS time-series data? 4. What types of analyses can be conducted using the resulting dataset? Phase 1 Accomplishments To address question 1, after identifying locations and times where the Seattle freeway system was likely to be congested, the team obtained from VTTI 214 useable trip segments satisfying the team’s time and space constraints. Manual study of the forward-facing video and time-series data revealed that 14 of these trips contained events in which the instrumented vehicle braked to a full or near stop. Using this as a learning sample, the team then explored different screening procedures aimed at automatic identification of brake-to-stop events. A simple screening based on a speed threshold was able to eliminate all but 33 of the non-braking trips, while retaining the 14 braking trips. A more sophisticated algorithm was able to eliminate all non-braking trips while retaining 13 of the 14 braking events.

17 Figure 4.1. Speeds of a leading and following vehicle versus time, showing the time point at which the leader began braking and the time interval during which the follower looked down to adjust a radio. To address question 2, characterization of driver behavior before and during braking events, one of the team members spent two days at VTTI’s secure enclave, where he viewed the video for 20 events, including the 14 braking events identified earlier. It was not only possible to classify potentially distracting activities but also, via the video frame time stamps, to determine when and how long the distracting activities occurred and to relate these intervals to the NDS time-series data. Figure 4.1 illustrates the integration of different NDS data types. Figure 4.1 plots the speeds of the leading and following vehicles as constructed from NDS time-series data, the point at which the leading vehicle’s brake lights went on as observed in the NDS forward-camera video, and a period of apparent distraction by the following driver obtained by observing the NDS face video. Question 3 is concerned with computing quantitative estimates of features such as reaction time, braking rates, and speeds. The research team developed a two-level approach to reducing and analyzing the NDS data. In the first and simpler level, each braking event was described using six elements: the initial speeds of the leading and following drivers, the follower’s headway and reaction time, and the average decelerations used by the leader and the follower. The team developed a spreadsheet tool to compute the first-level estimates of the six elements from inputs obtained from the NDS forward video and time-series data. The second level of reduction and analysis involved replacing the average deceleration rates of the first level -10 0 10 20 30 40 50 0 2 4 6 8 10 12 14 16 18 20 Speed feet/sec Time in secs Follower Leader Distraction period for follower

18 with more detailed estimates of the acceleration/deceleration sequences used by the leader and the follower. This was done through detailed modeling of the leader and follower trajectories, using the NDS forward radar, speedometer, and GPS speed time-series data. Table 4.1. Estimates of Braking Sequences for the Leading and Following Vehicles Depicted in Figure 4.2 Leader Follower Time (seconds) Acceleration (ft/sec2) Time (seconds) Acceleration (ft/sec2) 0 1.8 (a21) 0 1.4 (a11) 5.4 (t21) -3.2 (a22) 7.4 (t11) -2.7 (a12) 8.3 (t22) -5.8 (a23) 10.8 (t12) -10.8 (a13) 12.4 (t13) -3.15 (a14) Figure 4.2. Graphic display of leader and follower speeds, along with braking sequences from Table 2. Table 4.1 gives the estimated sequence of braking rates for the leader and follower shown in Figure 4.1, while Figure 4.2 is a graphic display of these results. The leader was initially accelerating at about 1.8 ft/sec2; at about time 5.4 seconds, the leader began decelerating at about -3.2 ft/sec2; and at time 8.3 seconds, the leader increased deceleration to about -5.8 ft/sec2. The follower was initially accelerating at about 1.4 ft/sec2; at time 7.4 seconds, the follower began decelerating at -2.7 ft/sec2; and at time 10.8 seconds, the follower increased deceleration to about -10.8 ft/sec2. The follower’s period of apparent distraction shown in Figure 4.1 immediately preceded the leader’s increase in deceleration, and the time between the leader’s increase in deceleration and the response by the follower was a relatively long 2.5 seconds. Finally, to address question 4, the team prepared a small (24 event) example representing the dataset the team plans to construct. This example dataset was used to illustrate how one could determine if potentially distracting activities on the part of a following driver were risk factors 0 2 4 6 8 10 12 14 16 18 -5 0 5 10 15 20 25 30 35 40 45 Time in secs S pe ed in f/ s Predicted Speed of leader Predicted 95% Upper 95% Lower Measured 0 2 4 6 8 10 12 14 16 18 -5 0 5 10 15 20 25 30 35 40 45 Time in secs S pe ed in f/ s Predicted speed (follower) Measured predicted 95% Upper 95% Lower a11 a12 a13 a14 t11 t12 t13 a21 a22 t21 t22 a23

19 for that driver causing a critical event, and to illustrate how cluster analysis can be used to classify the elements characterizing braking events. Regarding possible countermeasures, the example dataset was used to predict the effect a vehicle-based countermeasure might have on the incidence of critical events, and to estimate specifications or constraints that a road-based countermeasure would need to satisfy to achieve a target change in the likelihood of critical events. Conclusion Rear-ending crashes on congested freeways are responsible for a significant fraction of the nonrecurring delays that are becoming regular features of urban travel, and reducing the frequency of these crashes should lead to better use of our increasingly strained freeway system. At present, there is a shortage of countermeasures that have been proven effective in reducing freeway rear-ending crashes, but it is becoming clear that the key is to understand how drivers respond to the occurrence of stopping waves. An important class of events, which the team calls critical events, are those in which a following driver brakes at a significantly higher rate that his or her leader. Even when the following driver is not involved in a crash, this sort of event can establish conditions that make a crash later in the stopping wave more likely. The study’s goal is to compile, from the NDS, a dataset of brake-to-stop events on freeways and use this dataset to investigate relations between the conditions that make rear-ending crashes more likely and how drivers distribute attention between driving and non-driving tasks. The bulk of the effort during Phase 1 was devoted to demonstrating the feasibility of this plan by (1) developing and testing triggers for identifying brake-to-stop events from NDS time-series data, (2) developing and testing a method for characterizing drivers’ allocation of attention from the NDS face video, and (3) developing and testing methods for estimating quantitative features of brake-to-stop events. The Phase 1 work then illustrated how the proposed dataset could be used to identify distraction as a risk factor for critical events, how cluster analysis could be used to classify brake-to-stop events, and how the proposed dataset could be used to do first-cut estimates of the effectiveness of a vehicle-based driver-assist system and a road-based driver warning system.