Evaluation of Data Needs, Crash Surrogates, and Analysis Methods to Address Lane Departure Research Questions Using Naturalistic Driving Study Data (2011)

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.

C H A P T E R 5 Defining and Evaluating Lane Departure Crash Surrogate Thresholds Using Naturalistic Driving Study Data Lane departure crashes are the best measure of safety. Natu- ralistic driving studies, however, even the fully deployed SHRP 2 field driving study, will have limited cases of lane departure crashes. The naturalistic driving studies will capture crashes, near crashes, and incidents, as well as normal driving. The fre- quency of incidents and near-crash events is typically greater than the frequency of crashes; incidents and near-crash events may be used as crash surrogates. Using surrogates will also provide an opportunity to study what happens preceding and following an incident or event. The most significant advantage of naturalistic driving stud- ies is that they provide a firsthand record of the events that precede crashes and incidents. Roadway, environmental, vehicle, and human factors can be extracted directly rather than from secondhand information from police records and crash databases to identify relationships among factors that influence lane departure crash risk. This firsthand informa- tion can also be used to determine the factors that lead to a positive outcome. For instance, if a similar number of lane departures occur on roadway sections with and without paved shoulders and the paved shoulders have a higher pro- portion of safe outcomes (vehicles can return safely to the road), the incidents can be used to evaluate the effectiveness of paved shoulders. This chapter discusses potential lane departure surro- gates that can be obtained from naturalistic driving study data. Several data sets were used to evaluate thresholds for lane departure crash surrogates. The data sets are described fully in Chapter 3. For simplicity, naturalistic driving study data from UMTRI’s road departure crash warning (RDCW) field operation test (FOT) is referred to in this chapter as “the UMTRI data set,” and the naturalistic driving study data from VTTI’s 100-car study is referred to as the “VTTI data set.” The naturalistic driving study data from UMTRI and VTTI were used to evaluate which variables may be the most useful in setting triggers to identify lane departure events and to53assess what thresholds may be used. Data were reduced as described in Appendices A and B. The UMTRI data resulted in a number of encroachments but no conflicts or crashes. Only data for rural, paved, two- lane roadways were included. The VTTI data provided near crashes and crashes but no encroachments. Additionally, vari- ables were not consistent between the two data sets. As a result, the two data sets were evaluated separately, as discussed in the following sections. Introduction Frequency and severity of crash data are commonly used to assess whether driver, road, traffic, or environmental fac- tors influence safety and to evaluate whether a counter- measure is effective. However, crash-based safety analyses are plagued by several problems (Songchitruksa and Tarko, 2006). Crashes are rare, and events surrounding a crash are oftentimes random. As a result, safety analyses often depend on small sample sizes. Additionally, crash reporting can be inconsistent, which makes comparisons across sites difficult. Another problem is the timeliness of crash data. Once a countermeasure is implemented, agencies like to evaluate the immediate effectiveness to assess whether more resources should be invested. However, before-and-after crash studies often cannot be completed until several years after treatment installation because a representative sample is not available immediately to assess significant differences with sufficient power. Some researchers have addressed limitations in crash data by using surrogates to measure crash risk. Surrogates may take two forms: safety surrogates and crash surrogates. The differ- ence between the two is related to whether their underlying relationship to safety has been established. The types of surro- gates described in the following two paragraphs are examples of safety surrogates. In most cases, the underlying relationship between crashes and the safety surrogate is assumed. Selected

54safety surrogates are believed to have some relationship to safety, although a demonstrated relationship rarely exists. Additionally, when safety surrogates are used in studies, there is no attempt to define the relationship between the safety sur- rogate and crashes. FHWA (2009) suggests that a reduction in violations is a viable safety surrogate to evaluate the effectiveness of red-light running countermeasures (e.g., camera enforcement). Other common surrogates are traffic conflicts, traffic violations, road user behavior, and speed (Forbes et al., 2003). Change in speed is frequently used to assess the effectiveness of treatments such as traffic calming. It is assumed that if speeds are reduced, crashes will also be reduced. Lane deviation has also been used as a safety surrogate measure for assessing the likelihood of run-off-road (ROR) crashes (LeBlanc et al., 2006) and the likelihood of crashes resulting from distraction (Donmez et al., 2006). Retting et al. (2007) and Bonneson et al. (2004) used reduc- tion in red-light running violations as a safety surrogate to assess the effectiveness of red-light-running cameras. Garber et al. (2005) evaluated reductions in red-light running citations to evaluate the effectiveness of red-light-running cameras. The effectiveness of rumble strips has been evaluated on the basis of lateral placement and speed (Porter et al., 2004), drivers’ lane position with respect to a forced rumble strip encounter (Noyce and Elango, 2004), and vehicle’s lateral position and change in vehicle separation (Pratt et al., 2006). Taylor et al. (2005) observed vehicle placement relative to the edge line using single versus double paint lines to delineate presence of shoulder rumble strips. The second type of surrogate is the crash surrogate. This type of surrogate is expected to have some statistically mea- surable relationship to crashes. Ideally, the relationship between crashes and the surrogate measure is evaluated or known. If so, the crash surrogate can be used as a measure of effective- ness, and the reduction in crashes can be predicted. Shankar et al. (2008) defines a crash surrogate as a marker that is cor- related with a crash, usually based on time, so that as time increases the crash likelihood also increases. The authors also define a crash surrogate as a measure that is as responsive to the same interventions as the related crashes. For instance, edgeline rumble strips are expected to decrease the number of right-side lane departure incidents, as well as the cases of ROR crashes. The main difference between safety surrogates and crash surrogates is that safety surrogates are used by researchers and agencies as a stand-in variable for crash data. They are widely used, but there is no proven relationship between crashes and the variable used. It is assumed that if the safety surrogate changes (e.g., reduction in speeds), crash severity or frequency will improve. Additionally, studies in which safety surrogates are used do not attempt to derive a relationship.Background on Crash Surrogates This section discusses the crash surrogates that have been used in other studies for different types of crashes. Songchitruksa and Tarko (2006) used extreme value theory to model crash risk and frequency for right-angle crashes at intersections. Degree of separation was used as the surrogate variable, and the authors ordered traffic events from safest to most dangerous to assess risk. They defined the boundary between crash and noncrash events using the concept of crash proximity. They evaluated their methodology and concluded that there was a promising relationship between safety estimates and historical crash data. Archer (2001) used the traffic conflict technique, which reg- isters the occurrence of near accidents in real-time traffic to rep- resent accident frequency and outcome. The surrogate measure proposed by Archer (2001) is defined as “time to accident.” Gettman and Head (2003) derived surrogate crash measures from simulation models. The authors indicated that the best surrogate measures for crash risk were time to collision, post encroachment time, deceleration rate, maximum speed, and speed differential. Mounce (1981) evaluated the correlation between stop sign violations and crash rates. Salman and Al-Maita (1995) evaluated traffic conflicts and crashes at 18 T-intersections in Jordan and developed a statistically signifi- cant linear regression model that related annual number of crashes to mean hourly conflicts. Chin et al. (1992) used a time- to-collision method to model freeway merging conflicts. Other researchers have treated crash surrogates as a con- tinuum, with regular traffic incidents at one end and crashes at the other. Thresholds are used to partition incidents. Chin and Quek (1997) describe this as a probability distribution of incidents. Songchitruksa and Tarko (2006) use the notion of ordering traffic incidents from the safest to the most danger- ous (Figure 5.1).Source: Songchitruksa and Tarko, 2004. Figure 5.1. Continuum of traffic incidents.Different researchers have applied different crash surro- gates to define the boundary between incidents. A common

55measure that has been used is time to collision (TTC), where at TTC = 0 the subject vehicle and another vehicle/object col- lide, resulting in a crash. Songchitruksa and Tarko (2006) use the concept of degree of separation. When there is consider- able separation between vehicles on the same path that are passing a conflict point, the passage is considered safe. As sep- aration between vehicles decreases, risk increases until the two vehicles collide. This continuum of separation of time is called post-encroachment time (PET). Amount of separation can be partitioned into different levels of crash incidents that correspond to different levels of risk. In Songchitruksa and Tarko (2006), the threshold between the crash and crash-free boundaries is called crash proximity measure. Others refer to this threshold as a crash prevention boundary. Burgett and Gunderson (2001) define a crash prevention boundary as an analytically derived deterministic expression that separates driver performance into successful crash avoidance and unsuc- cessful crash avoidance. For any given set of conditions, there is a subset of driver brake response and level of deceleration that will result in crash avoidance and a subset of values that will result in a crash. Hayward (1972) suggests the use of TTC when modeling sit- uations where both vehicles continue in the same path without changing their speed. TTC is a good crash surrogate for two- vehicle crashes because the distance or the time separating two vehicles can be clearly identified. Crash prevention boundaries have also been used as a crash surrogate. Burgett and Miller (2001) used velocity, separation distance, deceleration, and braking to develop crash preven- tion boundaries for rear-end crashes. Szabo and Wilson (2004) used the amount of acceleration necessary to avoid a crash as a function of the timing or location of a warning to define a crash prevention boundary. Other crash surrogates used include proportion of remain- ing stopping distance (Allen et al., 1978), and deceleration rate (Songchitruksa and Tarko, 2004). Finally, Campbell et al. (2003) suggested that it is preferable to use physical measures of vehicle kinematic motions as the crash margin measure because collisions can be explicitly identified. Summary of Crash Surrogates Used for Lane Departures Several measures have been used as crash surrogates for lane departures, including lane keeping, TTC, and crash preven- tion boundaries. However, little information was available that describes developing statistical relationships between lane departure crashes and lane departure crash surrogates. As a result, it is assumed that these types of relationships will need to be derived from research projects developed as part of the SHRP 2 Safety Project S08, Analysis of the SHRP 2 Nat- uralistic Driving Study Data. Crash surrogates that have beenused for lane departure crashes are discussed in the following sections. Lateral Drift or Lane Keeping Lateral drift or lane keeping is one measure that has been used by several researchers to evaluate lane departures. Several stud- ies conducted at UMTRI, VTTI, and the University of Iowa (UI) have provided insights based on measures of lateral drift. UMTRI uses lane keeping to identify when a vehicle leaves the roadway. UMTRI researchers also define lane offset as the dis- tance between the centerline of the vehicle and the centerline of the lane. Lane position and relative motion within the lane are determined by analyzing the forward-looking monochrome camera data. On tangent sections, the UMTRI lane departure warning system (LDWS, part of the RDCW) shows a “lower cautionary alert” when the vehicle is close to a dashed lane line, which indicates potential movement into an adjacent travel lane with no other evidence of an imminent risk of sideswipe col- lision. Additionally, UMTRI uses the term “lane intrusion,” which suggests that a lane departure is imminent or likely. On curves, the UMTRI definition of a likely or imminent lane departure for the curve speed system (also part of the RDCW) was based on an estimate of most likely path, given vehicle speed, driver braking, turn signal use, assumptions about the unaware driver’s response time, likely deceleration rate, and a threshold lateral acceleration of 0.25 g,which assumes no super- elevation exists on the curve (LeBlanc et al., 2006). Oxley et al. (2004) summarized work by Steyer et al. (2000) and noted that one of the important safety-related features in curve negotiation is vehicle lateral placement. Steyer et al. (2000) argued that the driving path should be considered when investigating crashes on curves. The authors make the distinc- tion between right and left curves. Drivers who make left-side encroachments (across the centerline) may be doing so inten- tionally as they “cut the corner” or “straighten on the curve” if they cannot detect any opposing traffic. Steyer et al. (2000) also indicated that there was lateral placement related to curve radius, curve length, grade, and available sight distance. Time to Lane Departure or Time to Collision Several crash surrogates use time to some critical event as the measure of crash risk. Time to lane departure (TTLD) reflects the time remaining before a vehicle crosses the lane line if the vehicle maintains its current trajectory. Several researchers have used TTLD or TTC as a crash surrogate measure. Szabo and Wilson (2004) used TTC in their lane departure warning sys- tem to determine the point where a vehicle is about to leave the roadway and the driver should be alerted of the danger. Pomerleau et al. (1999) used TTLD to determine when a lane departure warning system should provide a driver alert.

56TTLD was defined as the time until an outer tire edge crosses the lane line. Mammar et al. (2006) developed a method to calculate time to line crossing (TTLC), considering both straight and curved vehicle paths. The authors estimated that a lane departure because of driver drowsiness leads to a slower rate of TTLC than in situations such as loss of vehi- cle control. They also indicated that real-time computation of TTLC is difficult because of limitations in determining vehicle kinematic variables, vehicle trajectory prediction, and lane geometry. As a result, approximate formulas such as the ratio of lateral distance to lateral speed have been used. The authors did develop a linear dynamic mode to predict future vehicle position on the basis of lateral dis- placement, vehicle position, steering angle, relative yaw angle, and roadway geometry. Distance Intruded Distance intruded measures the distance a vehicle crosses into an adjacent lane or shoulder. VTTI uses distance intruded to determine when a lane departure occurs. Its lane tracking sys- tem sets a trigger to define a “lane bust” or “lane abort” inci- dent. A lane bust occurs when a vehicle crosses a solid lane line. An incident is triggered when the vehicle moves a mini- mum of 3 ft outside a lane boundary without completing a lane change while traveling at a speed of 45 mph or higher (Dingus et al., 2006). Crash Prevention Boundaries Szabo and Wilson (2004) used the concept of crash preven- tion boundaries to assess the effectiveness of RDCWs. Two metrics were used, one for curves and one for tangent sections. For curve negotiations, there is a critical point at which driv- ers can receive a road departure warning and respond appro- priately by decelerating as needed, as shown in Figure 5.2. If the alert is provided after this point, the driver will not be able to safely negotiate the curve.Source: Szabo and Wilson, 2004. Figure 5.2. Relationship between curve geometry and warning point.This critical point location is given by where vo = initial forward speed, vs = safe speed for the curve, tr = driver reaction time, xw = distance between the warning location and critical point (CP), and dreq = deceleration necessary to achieve safe speed at CP. d v v x t v req o s w r o = − −( ) 2 2 2 5 1( . )The lateral acceleration limit to determine safe speed enter- ing the curve is where r = curve radius and as = lateral acceleration limit. The authors used a similar concept for tangent sections based on the geometry of a lateral drift into a jersey barrier, as shown in Figure 5.3. v a rs s= ( )0 5 5 2. ( . )The equation to determine the warning location and nec- essary lateral acceleration to avoid a lateral lane departure is where alat = lateral acceleration to avoid departure, θ = departure angle, and yw = distance from warning location and road boundary. Movement beyond the crash prevention boundary rep- resents a situation in which the vehicle is not likely to recover. Burgett and Gunderson (2001) also discussed the concept of a crash prevention boundary for road departure crashes, which is a function of driver brake response time and the level of deceleration needed to avoid a crash. The authors discussed the concept in relation to a driver traveling at a constant speed a v y t v lat o w r o = ( ) −( ) θ θ 2 2 5 3( . )

57Source: Szabo and Wilson, 2004. Figure 5.3. Relationship between lane geometry and warning point.on a tangent roadway section. The point at which the vehicle crosses the lane edge is defined as t = 0, and the crash preven- tion is defined by the driver’s steering maneuver and level of lateral acceleration created by the steering maneuver, as depicted in Figure 5.4. The authors use the geometric relation- ship between speed, side acceleration, departure angle, steer- ing angle, reaction time, and radius of curve for curve sections to develop crash prevention boundaries. An example is shown in Figure 5.5 for a 1,000-ft radius curve with a vehicle speed of 50 mph, a shoulder width of 10 ft, and an initial vehicle offset of 2 ft from the edge of the road.Source: Burgett and Gunderson, 2001. Figure 5.4. Relationship of driver steering maneuver and prevention of a road departure.Source: Burgett and Gunderson, 2001. Figure 5.5. Crash prevention boundary for a curve road departure.Selection of Lane Departure Surrogates One of the major benefits of naturalistic driving studies is that they can capture all levels of incidents, including those related to lane departures. One of the research questions addressed in this project involved assessing existing naturalistic driving study data and determining the most appropriate crash surro- gates to use in lane departure analyses. Potential crash surro- gates were also evaluated to determine what vehicle kinematic triggers could be used to flag potential lane departures in the full-scale study. Since a large amount of data will result from the full-scale NDS, it will be necessary to have some automated method to flag events of interest. The following section organizes information related to lane departure crash surrogates and outlines a process that could be used to develop lane departure crash surrogates in the full-scale study. Lane departure crashes provide a more complex situation for developing crash surrogates than other crash types. Most

58crash types, such as broadside or rear-end, can be defined by a time or distance metric (time or distance collision) because the hazard, collision with another vehicle, is clear. In a lane depar- ture crash, the main hazard is not always clearly identified, and in some cases multiple hazards may be present. For instance, when a vehicle departs the edge of the roadway, multiple haz- ards may be present and multiple outcomes (sequences of events) may be possible. Potential hazards include encounter- ing pavement edge drop-off, which could lead to loss of con- trol; encountering differential friction between the roadway surface and shoulder, which could also lead to loss of control; having the vehicle rollover; or striking a fixed object. The events accompanying each hazard can also result in a number of different outcomes, each of which can lead to the vehicle encountering different hazards. An initial sequence of events could result in several different outcomes based on dif- ferent hazards, driver responses, and roadway conditions. Fig- ure 5.6 shows three possible outcomes for the same initial ROR event. Each outcome may have a different type of crash surro- gate to describe it. The first sequence of events for all three sce- narios includes the vehicle running off the roadway to the right, encountering loose shoulder material, overcorrecting, and then crossing the centerline. In the first scenario, the vehicle then runs off the road to the left and strikes a tree. In the second scenario, the vehicle runs off the road to the left and overturns before striking the tree. In the third scenario, the vehicle crosses the centerline and strikes another vehicle head-on. The initial sequence of events was the same, but three different outcomes were possible with three different hazards. Each of the first events (e.g., ROR, encountering loose shoulder material) could have led to a different subsequent series of events. For instance,Figure 5.6. Three possible roadway departure outcomes from same initial sequence of events.the vehicle could have left the roadway to the right, encoun- tered loose shoulder material, and rolled over. Each stage of a lane departure can result in a number of out- comes, and each outcome may need to be described by differ- ent crash surrogates. As a result, lane departures in the present study were divided into categories where hazards would be consistent. Five categories were selected, which include normal driving, lateral drift within the travel lane, right- or left-side lane departure where the vehicle stays within the traveled way (lane encroachment), right- or left-side lane departure where the vehicle leaves the traveled way (shoulder encroachment), and lane departure crash. A crash surrogate or surrogates were identified for each cat- egory, except for normal driving because normal driving by definition is absence of conflict. A crash surrogate is the metric used to set boundaries between events and assess crash risk (e.g., time/distance to encroachment on the lane edge line). The crash surrogates and thresholds for each category were deter- mined after reviewing the available literature on crash surrogate measures, evaluating existing naturalistic driving study data, and assessing what is likely to be available with the full-scale SHRP 2 naturalistic driving study. The crash surrogates and threshold parameters for each category are described below, along with the rationale for the selection. Table 5.1 summarizes the lane departure categories and associated crash surrogates.Normal Driving • Description: This category represents the range of behav- ior remaining when crash surrogates and crash activity are removed.

59Category Hazard Surrogate Metric Normal driving Lateral drift Lane encroachment Shoulder encroachment Lane departure crash Table 5.1. Summary of Lane Departure Categories None Crossing lane line Sideswipe with adjacent or opposing vehicle Head-on collision Crossing lane line onto shoulder (left or right side) Shoulder (loss of control) Rollover Fixed object collision Crossing shoulder edge Vehicle, rollover, fixed object Lane deviation Distance to lane departure (DTLD) or time to lane departure (TTLD) Time to collision (TTC) or distance to collision (DTC) TTC or DTC Same as for shoulder encroachment Time to lane edge (TTLE) or distance to lane edge (DTLE) Change in steering angle or yaw rate Rollover potential TTC or DTC Time to shoulder edge (TTSE) or distance to shoulder edge (DTSE) NA ft s or ft s or ft s or ft s or ft Degree/s Lateral acceleration (g) s or ft s or ft NA• Crash surrogates: No crash surrogates are used for normal driving. However, amount of lane deviation is the metric used to distinguish between normal driving and a lateral drift. • Lower boundary: This category has no lower boundary. • Upper boundary: The boundary or threshold level between normal driving and lateral drift will need to be determined in the full-scale study. Lane keeping for individual drivers will vary, and drivers will maintain lane position differently based on a variety of factors, such as different roadway con- ditions (e.g., two-lane versus four-lane, presence and type of curve), weather conditions, time of day, and length of time driving. In order to set this threshold, it will be necessary to develop a range of normal vehicle activity under different situations and then determine what constitutes normal driv- ing for a given scenario. Depending on the resources available, normal driving can be established for individual drivers for situations of interest(daytime versus nighttime driving) or can be determined for a cohort of drivers. Some evaluation of what might define nor- mal driving was conducted using the UMTRI data set and is discussed in the section “Identifying Lane Departure Incidents Using Existing Data Sets” (p. 65). Lateral Drift • Description: This category includes incidents in which a vehicle’s deviation within its lane or its direction of travel to the right or left will result in a lane departure unless the driver changes course. • Crash surrogates: A single crash surrogate can be used to measure lateral drift because the only hazard is leaving the vehicle lane. The crash surrogate selected to assess lat- eral drift is distance to lane departure (DTLD), as shown in Figure 5.7. Distance is used rather than time because the lane tracking system used in the full-scale naturalistic study is expected to provide distance measurements.Figure 5.7. Distance to lane departure.

60TTLD can also be calculated and used as the surrogate measure. • Lower boundary: Lower boundary is the point where nor- mal driving ends, which will need to be determined in the full-scale study. This category will need to be defined after examination of data in the full-scale NDS. • Upper boundary: The upper boundary for lateral drift and a lane departure is where the vehicle’s outside tires come within a certain tolerance distance (Xtol) of the lane line or lane boundary. A tolerance distance is necessary because lane tracking systems can only locate a vehicle within its lane to a certain level of accuracy. As a result, the tolerance distance reflects uncertainty in locating the vehicle. UMTRI used a tolerance distance of 0.1 m in its study. The distance that will be used in the SHRP 2 naturalistic study will depend on the accuracy of the lane tracking system used. • Data needs: DTLD requires vehicle position relative to its lane. The GPS system that will be available with the full-scale system will not be accurate enough to locate a vehicle pre- cisely within its lane. The only way to obtain vehicle lane position will be to use the lane tracking system that will be available with the vehicle instrumentation package. The planned lane position tracking system for the full-scale studyhas a stated accuracy of ±0.656 ft (0.2 m). This is within the range of a normal tire width and is likely to be adequate. • Limitations: It will be difficult to calculate TTLC from any of the available SHRP 2 study variables, with the exception of the lane tracking system. The system is not expected to perform on gravel roads or in situations where either lane lines or some other lane delineation are not present (e.g., snow-covered roadway). The vehicle trace from the UMTRI data, depicted in Fig- ure 5.8, shows an example of a lateral drift, where the vehicle clearly drifted to the right but did not leave its lane.Source: UMTRI RDCW data set. Figure 5.8. Vehicle trace of nondeparture lateral drift.Lane Encroachment • Description: This category includes incidents where a vehicle departs its original lane of travel and encroaches into an adjacent travel lane. This adjacent lane may have other vehicles traveling in the same or opposite directions. An encroachment is defined as one or more tires encroach- ing or crossing the edge of the lane line. • Crash surrogates: The main hazard for an encroachment into an adjacent lane is a sideswipe or rear-end collision with

61another vehicle. The main hazard for an encroachment into an oncoming lane is a head-on or opposing-direction side- swipe crash. The crash surrogate for a head-on, sideswipe, or rear-end collision with another vehicle is TTC. DTC can also be calculated. Level of risk can be defined by a threshold value as the point at which an evasive steering maneuver is required to avoid a collision. If the subject vehicle does not encounter another vehicle after encroaching into one or more lanes of travel, the next proximate hazard is leaving the roadway. The crash surrogate in this case would be time to lane edge (TTLE) or distance to lane edge (DTLE). • Lower boundary: This is the threshold for lateral drift, as described above. • Upper boundary: The threshold between a lane encroach- ment and a shoulder encroachment is the point at which the vehicle’s outside tires come within a certain tolerance dis- tance (Xtol) of the lane line or lane boundary separating the lane and adjacent shoulder, similar to what was described for a lateral drift. The threshold between a lane encroach- ment and lane departure crash is the point at which the sub- ject vehicle strikes another vehicle. • Data needs: Calculation of time to collision requires vehi- cle speed, distance to adjacent or oncoming vehicle, coef- ficient of friction, grade, deceleration rate, and vehicle braking characteristics. Distance to adjacent vehicle will need to be determined using forward or side radar. Thus, time or distance to collision can only be calculated when a vehicle is within the tolerance of the radar systems. • Limitations: The ability to determine time or distance to collision or lane edge depends on the accuracy of the lane tracking system and the accuracy with which and distance at which the instrumented vehicle radar system can track objects in its path. Shoulder Encroachment • Description: This category includes incidents where a vehicle departs the traveled roadway surface onto a paved or unpaved shoulder. This is often referred to as a road departure or ROR incident. • Crash surrogates: There can be several hazards once a vehicle leaves the traveled portion of the roadway. The first hazard encountered when leaving the traveled way is the shoulder itself. Specific shoulder hazards that might be encountered include differential friction between the road- way and unpaved shoulder and other shoulder irregulari- ties (e.g., loose material, muddy shoulders) that may lead to loss of control or overturning of the vehicle.Hazards are different for a paved shoulder and unpaved shoulder. The team met with several lane departure experts at the Iowa Department of Transportation (Iowa DOT) and FHWA, and it was decided that encroachments onto the shoulder under different circumstances present different levels of crash risk. An encroachment onto a paved shoul- der introduces a lower level of risk, in the absence of haz- ards, than an encroachment onto an unpaved shoulder, even if TTC or time to shoulder edge (TTSE) is the same. The friction differential between the unpaved shoulder and paved roadway poses a risk for loss of control anytime the vehicle partially or fully leaves the paved roadway surface. This may be addressed by categorizing the encroachment into different levels of risk or considering time to paved shoulder edge as one crash surrogate and time to unpaved shoulder edge as another. It is difficult to determine a crash surrogate for loss of control on the shoulder because it does not fit within any of the typical metrics used in crash surrogates. Changes of a certain magnitude in steering angle or yaw rate may be used to identify loss of control, but they are not crash sur- rogates per se. It may be necessary to define the next most likely sequence of events (overturn, return to travel lane, cross centerline) and then use the corresponding crash sur- rogate for that event. Rollover potential is the crash surrogate when rollover is a possibility. Rollover potential is described in the section “Determining Rollover Potential” (p. 72). The next hazard encountered when leaving the traveled way is collision with a fixed object. This can occur on the shoulder or when the vehicle leaves the shoulder. When a fixed object (e.g., tree, guardrail, mailbox, utility pole, bridge abutment) presents the most immediate hazard, the pro- posed crash surrogate is TTC or DTC. Another hazard when leaving the roadway is that, once a vehicle leaves the shoulder, it may encounter an adverse slope, which may result in overturning. When the primary hazard is leaving the shoulder, the proposed crash surro- gate is TTSE. Rollover risk may also be used. Level of risk for most of the crash surrogates listed above can be defined by the actions that need to be taken to avoid a crash. The point between a lower risk and a higher risk event may be defined as the point at which a severe evasive action is required to avoid a crash. Once a vehicle leaves the roadway onto to the shoulder, the recovery options are braking to a stop before leaving the shoulder or striking an object, or steering back onto the original travel lane. Evasive actions occur when the vehicle undergoes a steering or braking maneuver that exceeds normal steering or braking. AASHTO uses a deceleration rate of 11.2 ft/s2 (0.35 g) for stopping distance because this deceleration is within the capability of most drivers to stay within their

62lane and control their vehicle when braking on wet surfaces (AASHTO, 2004). A value of 14.8 ft/s2 (0.46 g) is used for emergency braking. VTTI used a lateral acceleration of ≥0.7 g as a trigger that a lane departure incident had occurred (Dingus et al., 2006). Thus, between 0.35 g and 0.7 g is a good starting point for setting the threshold decel- eration between an encroachment and a lane departure conflict. It will be necessary to examine a number of lane departure incidents and subjectively assess what consti- tutes an encroachment versus a lane departure conflict and then determine the boundary between normal and signif- icant braking. Time to collision (tTTC) is a function of initial vehicle velocity, angle of departure (θ), coefficient of friction ( f ) between the tires and shoulder, braking capabilities of the vehicle, driver reaction time, driver response, distance to the object (dobj), grade, and deceleration rate (a). The time to collision (tcritical) that requires a critical deceleration rate (acrit- ical) in order for the vehicle to stop safely is the threshold between an encroachment and a lane departure conflict, as shown in Figure 5.9. The time to collision (tnorm) where a vehicle can stop safely with normal deceleration rates (anorm), as shown in Figure 5.10, is given by the following: t t tnorm TTC critical= − ( . )5 4Figure 5.9. Threshold requiring evasive deceleration.Figure 5.10. Threshold requiring normal deceleration.Similarly, TTSE or distance to shoulder edge (DTSE) is a function of initial vehicle velocity, angle of departure (θ), coefficient of friction (f) between the tires and shoulder, braking capabilities of the vehicle, driver reaction time, driver response, distance to the shoulder edge (dshld), shoul- der width (wshld), grade, and deceleration rate (a). The dis- tance the vehicle travels before crossing the edge of the shoulder (Figure 5.11) is given by the following: d w shld shld = ( )sin ( . )θ 5 5Figure 5.11. Distance to edge of shoulder.Less information was available about what constitutes a normal range of steering angles than was available for nor- mal passenger vehicle deceleration rates. As a result, it mayalso be necessary to examine a number of lane departure incidents and subjectively assess what constitutes a signif- icant evasive action. The threshold defining a significant evasive action is the point at which a driver must employ excessive steering maneuvers in order to avoid the object or shoulder edge, as shown in Figures 5.12 and 5.13.• Lower boundary: This is the threshold for lane encroach- ments as described above. • Upper boundary: The threshold between a shoulder encroachment and a lane departure crash is when the vehicle physically strikes an object or physically rolls over (crash), which may defined as TTC or TTSE = 0. • Data needs: Calculation of time or distance to collision or time or distance to shoulder edge requires vehicle speed, deceleration, angle of departure, shoulder width, distance to fixed object, coefficient of friction, grade, and decelera- tion rate. • Limitations: In the full-scale NDS, friction will not be avail- able. The expected spatial accuracy of roadside features is ±3.0 ft (0.914 m). The accuracy of the GPS, used to deter- mine the vehicle’s spatial position, is unknown, but the GPS will not have differential correction capabilities. Accuracy for a nondifferentially corrected GPS can be as low as ±15 m (49.2 ft). This would significantly affect the ability to cal- culate TTC. The vehicle instrumentation packages are expected to be able to determine distance and heading to objects using the forward or side radar. The vehicle instru- mentation system will have a forward radar capable of tracking and storing information for the five objects closest to the vehicle. Objects can be identified and tracked for up to 200 m (656.2 ft) in front of the vehicle within ±0.324

63Figure 5.12. Threshold for normal steering. Figure 5.13. Threshold for evasive steering.radians (18°) of the horizontal field of view, given that the field of view is centered on the test vehicle heading. Figure 5.14 shows a typical vehicle trace for a nonconflict encroachment (UMTRI data). As shown, the vehicle leaves the roadway for some distance and then safely returns.Source: UMTRI RDCW data set. Figure 5.14. Vehicle trace of nonconflict run-off-road incident.Lane Departure Crash • Description: A crash is defined as an incident where a vehi- cle strikes another vehicle or object (TTC or TTSE = 0). A vehicle overturning one or more times is also considered to be a crash. A crash may also be defined as a situation wherethe vehicle leaves the roadway and is forced to an unplanned stop. For instance, sliding off a roadway during a winter weather event and then sliding to a stop in the median may be considered a crash. This category includes all lane depar- ture crashes, whether or not they are reported in a police document, that are observed in the naturalistic study. A reported collision is one where a crash report is filed. An unreported collision is one observed in the naturalistic study but for which a police accident report has not been filed; consequently, the crash would not show up in a crash data- base. For example, a driver leaves the roadway and strikes a mailbox but proceeds after recovering. In some states, there is no requirement to report property-damage-only crashes

64unless the damage exceeds some value. The VTTI driving study found that out of 82 minor nonproperty-damage con- tact collisions, only 15 were reported to the police (Dingus et al., 2006). • Crash surrogates: NA. • Lower boundary: This is the threshold for lane or shoul- der encroachments as previously described. • Upper boundary: NA. • Data needs: Identification of lane departure crashes will require setting triggers for vehicle kinematics that provide indications that a crash has occurred (e.g., sudden deceler- ation). It is expected that crashes will be identified as part of the data quality assurance (SHRP 2 Safety Project S06, Technical Coordination and Quality Control). • Limitations: It is expected that VTTI will identify crashes during the full-scale instrumented vehicle data collection and that most crashes will be identified. However, the iden- tification of crashes will be highly dependent on the thresh- olds used, and, as such, some crashes (e.g., vehicle sliding off the roadway during winter weather) may not be included. Evaluating Incident Outcome with UMTRI Data A number of lane departure incidents in the UMTRI data set were assessed so that they could be divided into categories and crash surrogates in order to test the system for its ability to cat- egorize lane departures, as described in the previous sections. For each lane departure, the hazard that the vehicle was most likely to encounter after departing its lane based on fac- tors such as angle of departure, surrounding hazards, and vehicle speeds was determined. Hazards were those that pre- sented the most imminent threat. The object (hazard) most likely to be struck was determined for each situation by esti- mating the anticipated vehicle path and by a visual inspection of the forward and aerial imagery. A hazard could also include an oncoming vehicle or the shoulder if no specific objects were in the vehicle’s likely path. Figure 5.15 shows an example of how potential hazards were determined for one vehicle. The subject’s vehicle exited the roadway and encroached 2.1 ft onto a paved shoulder. If the vehicle were to continue along its path, the intersecting roadway provided the first hazard that the subject vehicle would encounter. If the vehicle returned to the roadway and overcorrected, the first hazard the vehicle would encounter was an oncoming vehicle, as shown in Figure 5.16. Each vehi- cle’s position was determined for various points in time.Figure 5.15. Schematic of a vehicle departing its lane and its likely path if the driver does not correct lane departure.Figure 5.16. Schematic of vehicle departing its lane and overcorrecting.Figures 5.17 and 5.18 show other examples of how potential hazards can be determined. In this case, the vehicle departed its lane to the left and crossed the centerline. The forward imagery and aerial imagery for the location were examined for poten-tial hazards (Figure 5.18). The most imminent hazard for the scenario if the vehicle were to continue on its current path (to the left) or overcorrect to the right was determined using vehi- cle speed, vehicle position, and location of potential hazards. At the vehicle’s current speed and trajectory, if the vehicle did not correct its path, the most imminent hazard was the left guardrail. The shoulder is paved and poses a low hazard. If the vehicle overcorrected, an estimation of the potential paths off the right side of the roadway indicated that the most likely haz- ard was a mailbox. No oncoming vehicles were present, so col- lision with another vehicle was not a potential threat.

65Source: UMTRI RDCW data set. Figure 5.17. Forward image for where vehicle departed the roadway (cross centerline).Figure 5.18. Schematic of potential outcomes and hazards encountered.Identifying Lane Departure Incidents Using Existing Data Sets The full-scale SHRP 2 naturalistic driving study will result in a tremendous amount of data. This amount will necessitate an automated method to identify lane departures. An auto- mated method would entail selecting variables within the nat- uralistic driving data sets that are most likely to experience a significant change during a lane departure, and establishing a threshold value for these variables so they can be used as flags for potential lane departures. Triggers Used in Other Naturalistic Driving Studies VTTI selected relevant variables and triggers to set thresholds between valid and invalid critical events (Dingus et al., 2006). The study used a sensitivity analysis to evaluate placement of triggers at various levels. If a trigger is set too low (Type I error, or lower sensitivity), a larger percentage of actual incidents is selected, as well as a larger number of nonincidents (false alarms). This results in longer and less useful data reductiontime. Alternatively, if the trigger is set too high (Type II error), nonincidents are less likely to be selected, but a larger number of actual incidents may be missed as well. VTTI used an iterative process to select triggers for valid incidents. Triggers were set to a lower sensitivity, and data reduction was used to evaluate resulting incidents. VTTI researchers used a normal distribution to depict how Type I and Type II errors could be minimized based on signal detection theory, as shown in Figure 5.19. The final triggers for variables related to lane departure incidents include the following: • Lateral acceleration ≥0.7 g; • Longitudinal acceleration ≥0.6 g; • Longitudinal acceleration ≥0.5 and forward TTC ≤4 s; • Longitudinal deceleration 0.4 g to 0.5 g, forward TTC ≤4 s, and distance to collision <100 ft; and • Yaw rate ≥⎟ 4°⎟ change in heading within a 3-s window of time.McLaughlin et al. (2009) evaluated ROR crashes and near crashes using the VTTI 100-car study. These researchers iden- tified ROR maneuvers by evaluating steering wheel position, yaw rate, and braking, as shown in Figure 5.20.The University of Iowa teen driver study used a trigger of 0.5 g for lateral acceleration to indicate when a potential inci- dent had occurred (McGehee et al., 2007). Evaluation of Lane Departure Thresholds Using UMTRI Data The naturalistic driving study data from UMTRI and VTTI were used to evaluate which variables may be the most useful

66Source: Dingus et al., 2006. Figure 5.19. Graphical depiction of setting trigger criteria using the distribution of valid events.Source: McLaughlin et al., 2009. Figure 5.20. Steering wheel angle relative to ROR maneuver.in setting triggers to identify lane departure events and to assess what thresholds may be used. Data were reduced as described in Appendix A. The UMTRI data indicated a number of encroachments, but no conflicts or crashes. Only data for rural, paved, two-lane roadways were included. The VTTI data provided near crashes and crashes, but no encroachments. Additionally, variables were not consistent between the two data sets. The two data sets therefore were evaluated separately. This section describes the evaluation using the UMTRI data, and the following sec- tion describes the evaluation using the VTTI data. The first section below examines differences in kinematic variables between normal driving data and left- and right-side lane departures. The second section compares normal driv- ing to assess variables that could be used to partition normal driving data. The third section discusses sample size issues. It should be noted that left-side lane departure in curves is sometimes intentional and, rather than being due to an unin- tentional lane departure, is due to the driver intentionally “cutting the curve.” While the researchers did not account forthis specifically, this should be considered when evaluating lane departures. Examining Kinematic Variables The UMTRI data reduction resulted in 22 right-side and 51 left-side lane departure events for two-lane rural roads. An incident was a situation where the vehicle departed its lane by 0.1 m or more at some point. All incidents were considered to be encroachments because the vehicle departed its lane in each case but was not forced to take some evasive maneuver and did not lose control on the shoulder. The continuous data surrounding each incident was extracted. Data for which no incident had occurred was termed “normal” driving data. Several variables were exam- ined to determine whether they could be used to set thresh- olds between normal driving and lane departure incidents. The maximum positive and negative value for each incident was extracted for the following vehicle kinematic variables: lateral speed, lateral acceleration, yaw rate, forward accelera- tion, roll rate, and pitch rate. The maximum negative and positive values for various vehi- cle kinematic variables for right-side and left-side events were compared with approximately 105,400 records (in 0.1-s inter- vals) of nonincident (normal) driving. Data for each kinematic variable (lateral speed, yaw rate, side acceleration, forward acceleration, roll rate, and pitch rate) for the normal data were graphed against the data for the lane departure events. Figure 5.21 shows the distribution of data for the kinematic variable “lateral or side speed (in m/s)” for left- and right-side lane departures that are all approximately normal. The distri- bution for normal driving is the center distribution, shown in maroon. The distribution of maximum positive side acceler- ation for left-side lane departures is shown to the right in green, and the distribution of maximum negative side accel- eration is shown to the left in blue. Data for left- and right-side

67(a) (b) Figure 5.21. Distribution of lateral speed (m/s) for normal driving compared with maximum positive and negative values from (a) left-side and (b) right-side lane departures.lane departures were evaluated separately because they have different kinematic signatures. The analysis showed differ- ences in acceleration among the left- and right-side departures compared with normal driving (p < 0.05 for all comparisons). Figure 5.22 shows the distribution of data for the kinematic variable “forward acceleration for left- and right-side lane departures in g’s.” Figures 5.23 to 5.26 show the same infor- mation for lateral or side acceleration in g’s, roll rate in degrees per second, pitch rate in degrees per second, and yaw rate in degrees per second, respectively.(a) (b) Figure 5.22. Distributions of forward acceleration (g) for normal driving compared with maximum positive and negative values from (a) left-side and (b) right-side lane departures.Because the values for the left side and right side were dis- crete, a ranking test was used. The Wilcoxon Rank Sum Testwas used to determine whether the normal driving data were statistically different from events data. The test determines whether two independent samples of observations are from the same distribution. It evaluates the sign and magnitude of the rank of differences between pairs, and assesses whether two independent samples have similar rankings. In all cases except maximum low yaw rate for right-side lane departures, the test showed that the data were statistically different. Although the distributions for most variables were deter- mined to be different at the 95% level of significance, a signif- icant amount of overlap exists, as shown in Figures 5.21 to 5.26. Thus, setting a higher threshold to ensure that a larger

(a) (b) Figure 5.23. Distributions of lateral acceleration (g) for normal driving compared with maximum positive and negative values from (a) left-side and (b) right-side lane departures.(a) (b) Figure 5.24. Distributions of roll rate (degrees/s) for normal driving compared with maximum positive and negative values from (a) left-side and (b) right-side lane departures.(a) (b) Figure 5.25. Distributions of pitch rate (degrees/s) for normal driving compared with maximum positive and negative values from (a) left-side and (b) right-side lane departures.

69(a) (b) Figure 5.26. Distributions of yaw rate (degrees/s) for normal driving compared with maximum positive and negative values from (a) left-side and (b) right-side lane departures.number of normal driving conditions are not included may result in a threshold that will likely miss a large number of events. The alternative is also true: setting a lower threshold to include most events may also result in the inclusion of a large number of nonevents, which will result in more unnec- essary data reduction. This is similar to what VTTI found as its study evaluated methods to set appropriate triggers, as dis- cussed in the section “Triggers Used in Other Naturalistic Driving Studies” (p. 65). To summarize the data, differences exist in vehicle kine- matic values for lane departure events and for normal driving.However, although the data can be shown to be different and statistically significant, a considerable overlap still exists. This indicates the difficulty in setting thresholds low enough to include all incidents but still high enough so that a large amount of nonincident data does not have to be evaluated. Tables 5.2 and 5.3 show the range of maximum and mini- mum values for each kinematic variable. The values for lateral speed in Table 5.2 can be interpreted to mean that the contin- uous data for each left-lane departure included at least one value between −1.04 m/s and −0.03 m/s and at least one value with a lateral speed of 0.20 m/s or higher. If these values representedLateral Speed (m/s) Yaw Rate (deg/s) Side Acceleration (g) Max Negative Max Positive Max Negative Max Positive Max Negative Max Positive −1.04 to −0.03 0.20 to 1.56 −13.3 to −0.2 0.10 to 6.15 −0.23 to −0.01 0.02 to 0.42 Forward Acceleration (m/s2) Roll Rate (deg/s) Pitch Rate (deg/s) −0.13 to −0.03 0.03 to 0.13 −8.70 to −0.06 1.45 to 8.54 −5.64 to −0.13 0.48 to 10.19 Table 5.2. Range of Maximum Negative and Positive Values for Left-Lane Departure Events for UMTRI DataLateral Speed (m/s) Yaw Rate (deg/s) Side Acceleration (m/s2) Max Negative Max Positive Max Negative Max Positive Max Negative Max Positive −1.28 to −0.08 0.22 to 2.40 −1.90 to −0.05 0.15 to 12.20 −0.26 to −0.01 0.01 to 0.1 Forward Acceleration (m/s2) Roll Rate (deg/s) Pitch Rate (deg/s) −0.14 to −0.02 0.01 to 0.07 −7.43 to −0.76 0.69 to 8.29 −5.86 to −0.81 1.05 to 6.26 Table 5.3. Range of Maximum Negative and Positive Values for Right-Lane Departure Events for UMTRI Data

70a large number of lane departure incidents, they could be used as a starting point to set threshold values for flagging lane depar- tures. However, as indicated, there were not enough samples of incidents to determine what threshold values should be set. It will be necessary to set thresholds after examination of a much larger number of incidents in the full-scale study. How- ever, initial results suggest that for left-side lane departures, roll rate, yaw rate, side acceleration, and side speed are likely to be good candidates to identify events. Results suggest that for right-side lane departures, yaw rate, side acceleration, and lateral speed are good candidates to identify encroachments. Ayers et al. (2004), for instance, indicated that yaw rate is the first indication that a potential vehicle movement may be occurring. Selecting Parameters to Partition Driving Environments The analysis in the previous sections included all incidents and normal driving data that were extracted for two-lane rural roads. In reality, vehicle kinematic variables will differ among different driver, roadway, and environmental factors; in the large-scale study, differences in vehicle operation under different roadways and environments should be considered and statistically controlled. The Wilcoxon Rank Sum Test was used to compare whether distributions of lateral offset were different under several driving scenarios. Although only a limited amount of data was available, an exploratory analysis of differences in normal data for different situations was made to get a sense of how data might be parti-tioned. The lane offset variable, which indicates the offset of the vehicle’s center from the center of the lane, was used for com- parison. Normal driving data on a tangent section was com- pared with normal driving on a right-hand and left-hand curve (orientation of curve from the perspective of the driver—e.g., a right-hand curve to the right). As shown in Figure 5.27, lat- eral offset differs from a tangent section to a right- or left-hand curve, and lateral offset on a right-hand curve also differs from lateral offset on a left-hand curve. Data were not sufficient to compare lateral offset for different curve radii, but this vari- able is expected to have a large impact. Results indicated that differences were statistically significant at the 95% level of significance.Left Right Tangent -3 -2 -1 0 1 LA N E O F F S E T Figure 5.27. Vehicle offset for tangent section versus right and left curve.Differences in lateral offset were also compared for day- time versus nighttime driving. As indicated in Figure 5.28, the mean lateral offsets in the two situations were similar, but the results of the Wilcoxon Rank Sum Test indicate that the dis- tributions were different at the 95% level of significance. Data were compared for tangent sections only.Differences between drivers were also compared. Lateral off- set by driver for normal driving is shown in Figures 5.29 and 5.30. Data were compared for tangent sections only. As shown, lateral offset among drivers varies significantly. Distribution of lateral offset was compared among all drivers and differences were statistically significant for all driver pairs except for Driv- ers 14 and 48, Drivers 14 and 60, and Drivers 64 and 85. This is the expected result because different drivers have different driving styles.Lateral offset was compared for several situations, as described in the previous paragraphs. Differences were noted

71NightDay -1.5 -1.0 -0.5 0.0 0.5 1.0 La ne O ff se t Figure 5.28. Vehicle offset for daytime versus nighttime driving on tangent sections.Dr12 Dr14 Dr16 Dr17 Dr18 Dr24 Dr6 Dr8 -1.5 -1.0 -0.5 0.0 0.5 1.0 La ne O ff se t Figure 5.29. Vehicle offset for Drivers 6, 8, 12, 14, 16, 17, 18, and 24 for tangent sections.

72Dr28 Dr35 Dr48 Dr51 Dr59 Dr60 Dr64 Dr85 -1.5 -1.0 -0.5 0.0 0.5 1.0 La ne O ff se t2 Figure 5.30. Vehicle offset for Drivers 28, 35, 48, 51, 59, 60, 64, and 85 for tangent sections.between driving on a tangent and on a left- or right-hand curve, between nighttime and daytime driving, and between individual drivers. As indicated, differences are expected as to what constitutes normal driving behavior. Normal driving can be stratified by a large number of variables. Assuming the focus of lane departures will be on rural roadways, the minimum roadway and environmental characteristics should include the following: • Roadway type (e.g., two lane, four-lane undivided, four- lane divided); • Tangent versus curve; • Radius of curve (may be aggregated to ranges of curve radii); • Paved versus unpaved shoulders; • Narrow versus wide shoulders; • Dry versus wet versus snow- or ice-covered roadways; • Nighttime versus daytime driving, including presence of overhead street lighting; • Presence of rumble strips; • Posted speed limit; and • Roadway surface (paved versus dirt or gravel). Evaluation of Lane Departure Thresholds Using VTTI Data The VTTI and UMTRI data were evaluated separately because different data were available from each. Additionally, the UMTRI data had only data for encroachments, while the VTTI data had only data for crashes and near crashes. The disadvan-tage of the VTTI data was that it included all roadway types and only had a limited sample size (n = 29). Additionally, no expo- sure (normal) driving data were available to the research team, so event thresholds could not be compared against normal driving conditions. Thresholds for the crash and near-crash incidents available from the VTTI data were evaluated. Continuous data were available for 29 lane departure incidents. The crash and near- crash events showed distinct changes in forward and side accel- eration, which were the only two variables that could be used to evaluate the data. Several examples are shown in Figures 5.31 and 5.32.Table 5.4 shows the range of maximum and minimum val- ues for each kinematic variable for each lane departure event. The values for side acceleration in Table 5.2 indicate that all left-lane departures had at least one negative value that was −0.01 g and at least one positive value that was greater than 0.02 g. As shown in Table 5.3, each right-side lane departure had at least one negative value for side acceleration that was −0.01 g or lower and one positive that was 0.01 g or higher. These values could therefore be used as starting points in identifying lane departure events.Determining Rollover Potential Background On average, approximately 274,000 light vehicles were involved in rollover crashes annually between 1999 and 2003 (NHTSA,

73Source: VTTI data set. Figure 5.31. Forward and side acceleration trace for near crash on two-lane roadway.Source: VTTI data set. Figure 5.32. Forward and side acceleration trace for near crash on two-lane roadway.Side Acceleration (g) Forward Acceleration (g) Max Negative Max Positive Max Negative Max Positive Left-lane departure −0.68 to −0.05 0.03 to 0.23 −0.80 to −0.03 0.03 to 0.74 Right-lane departure −2.69 to −0.01 0.08 to 0.29 −1.90 to −0.11 0.05 to 0.95 Table 5.4. Range of Maximum Negative and Positive Values for VTTI Crash and Near-Crash Events2005b). Although rollover crashes made up only 2% of crashes, they accounted for almost one-third of light vehicle occupant fatalities (including 59% of sport-utility vehicle fatalities) in 2003. Rollover crashes accounted for 10,182 fatalities for pas- senger vehicle occupants in 2007 (Insurance Institute for High- way Safety, 2009). Most rollovers occur when a driver loses control of a vehicle and the vehicle begins to slide sideways. At this point, a curb, guardrail, tree stump, or soft or uneven ground on the side of the roadway can “trip” the vehicle and cause it to roll over. Rollovers are also caused by a driver turning too aggressively either at high velocity or with a sharp turning radius, causingthe vehicle to tip up and then roll over. Rollovers can also take place after a collision with another vehicle. Estimating Rollover Propensity The risk of a vehicle being involved in a rollover depends on a number of factors, including the vehicle’s center of gravity, vehicle design, friction between surface and tires, steering input, roadway geometry, and vehicle speed. As a result, deter- mining the likelihood that a vehicle will roll over can be fairly complicated. The simplified methods that follow, however, have been used to estimate roll propensity.

74Source: AASHTO, 2001. © American Association of State Highway and Transportation Officials (AASHTO). All rights reserved. Figure 5.33. Maximum side friction factors.

75One of the main methods to assess rollover risk is to use the static stability factor (SSF) (NHTSA, 2005a). SSF is given by Equation 5.6: where t = vehicle track width and h = height to center of gravity. Gillespie (1992) expanded the concept to develop the rela- tionship between forces acting on the vehicle and a mea- sure of vehicle stability against rollover. The relationship is given by where athreshold = maximum side acceleration sustained before a vehicle engages in rollover and ϕ = cross slope (for flat roads) or superelevation (for curves). In designing horizontal curves, the radius is calculated (AASHTO, 2001) using the formula and ϕ = −v Rg f 2 127 5 9( . ) R v e f g = +( ) 2 127 5 8( . ) a g t h threshold = + 2 5 7ϕ ( . ) SSF = t h2 5 6( . )where v = advisory speed or design speed limit in m/s, f = safe side friction coefficient, g = 9.81 m/s2, and R = radius of the curve expressed in m. Substituting Equation 5.9 into Equation 5.7 yields the following: Friction (f) can be obtained using Figure 5.33. Rollover threshold (RT) = = + = a g t h e t threshold 2 2 127 5 10 2 h v Rg f+ − ( . )Summary The value of the naturalistic driving study is the ability to gain insight on crashes that may not be observed using other data collection approaches (e.g., crash databases, test tracks, driving simulators). Thus, even though the number of crashes may not be as representative, naturalistic studies do capture many use- ful safety and crash surrogates that may not be observed in police-reported crashes but can provide more insight into what can precipitate a crash—before a crash actually occurs. This report has outlined information that may be used to develop crash surrogates for lane departures. Selection of crash surrogates for lane departures is not an easy task because mul- tiple hazards can be present for each lane departure, and differ- ent surrogates may need to be specified depending on the most likely hazard. Lane departures are partitioned into categories and the surrogates are defined for the hazards most likely to be encountered. Additionally, the team evaluated lane departures in the UMTRI and VTTI NDSs and identified some starting points for setting triggers for the full-scale study.