Highway infrastructure and safety implications of AV technology in the motor carrier industry

Research in Transportation Economics xxx (xxxx) xxx

Contents lists available at ScienceDirect

Research in Transportation Economics journal homepage: http://www.elsevier.com/locate/retrec

Research paper

Highway infrastructure and safety implications of AV technology in the motor carrier industry Jill M. Bernard Bracy a, Ken Q. Bao b, *, Ray A. Mundy a a b

Center for Transportation Studies, University of Missouri – St. Louis, 1 University Boulevard, St. Louis, MO, 63121, USA University of California Santa Barbara, Dept. of Economics, 2045 North Hall, UCSB, Santa Barbara, CA, 93106, USA

A R T I C L E I N F O

A B S T R A C T

JEL classification: R41 R58 R40 R49 R42 R48

This study explores the infrastructure investments needed to support the adoption of autonomous vehicles and provides potential safety benefits from AV technologies as a possible motivation for such investments. We assume that the motor carrier industry will be (one of) the first adopter(s) of such technologies, and therefore is the focus of this study. Using large truck crash data from 2013 through 2015 obtained from the Missouri State Highway Patrol, Chi-square Automatic Interaction Detection decision trees are estimated to examine the effect of AV technologies on motor carrier crash severity. Results suggest that the greatest contributory predictors of crash severity outcomes are driving too fast for conditions, distracted/inattentive driving, overcorrecting and driving under the influence of alcohol. If these circumstances are altered by AV technologies, it is suggested that between 117 and 193 severe crashes involving large trucks could be prevented annually in Missouri alone. To render such safety benefits, key vehicle needs include autonomously controlling acceleration and steering, monitoring of the environment, and responding to dynamic driving environments without the need for human intervention. Importantly, the safe operations of a system that can perform such AV tasks require readable lane markings to capitalize on potential safety benefits.

Keywords: Autonomous vehicles Connected vehicles AV CV Highway safety CHAID Infrastructure Highway infrastructure Large truck crashes

1. Introduction Safety enhancements provide incentives for mass adoption of Autonomous Vehicle (AV) technologies in the motor carrier industry, and as a result, assumptions are made that motor carriers (defined here as a single-unit truck with two or more axles, truck tractors, and other heavy trucks) will be one of the first adopters of automated driving capabilities that most likely will be first applied to our national inter­ state system. AV technology can replace human drivers with a computation-based decision-making process for a varying array of driving tasks. The Society of Automotive Engineers (SAE) define levels of automation based on the extent to which driving tasks are delegated to the vehicle (SAE International, 2018), as presented in Table 1. Currently, many companies are introducing SAE level 2 and developing SAE level 3 capable vehicles, while others are opting to focus directly on level 4 automation. The purpose of this study is twofold. First, we highlight the infrastructure requirements needed for widespread use of

AV technology, particularly in the motor carrier industry. Within the class of AV technologies, we only focus on technologies classified as SAE level 4 high-automation and level 5 full-automation, in which the automated driving features will not require the human to overtake driving features (SAE International, 2018). Second, we try to motivate such investments by estimating the potential safety gains that are offered by such technologies. Benefits rendered from automated technologies include retaining drivers, improving drivers’ health and wellness, reducing driver fatigue while adhering to the hours of service regulations, mitigating driver distraction, lowering fuel consumption and, most importantly, dimin­ ishing unsafe driving (Bhoopalam, Agatz, & Zuidwijk, 2017; Short & Murry, 2016). Unsafe driving behaviors for motor carrier drivers include speeding, reckless driving, improper lane change, following too closely, improper passing, driving too fast for conditions, failing to yield and distracted driving (Bernard & Mondy, 2016; Short & Murry, 2016). Though unsafe driving will not disappear, it is probable that the number

* Corresponding author. E-mail addresses: [email protected] (J.M. Bernard Bracy), [email protected] (K.Q. Bao), [email protected] (R.A. Mundy). https://doi.org/10.1016/j.retrec.2019.100758 Received 29 January 2019; Received in revised form 17 July 2019; Accepted 12 October 2019 0739-8859/© 2019 Elsevier Ltd. All rights reserved.

Please cite this article as: Jill M. Bernard Bracy, Research in Transportation Economics, https://doi.org/10.1016/j.retrec.2019.100758

J.M. Bernard Bracy et al.

Research in Transportation Economics xxx (xxxx) xxx

level 4 or higher to operate safely in a mixed vehicle environment in which vehicles range from fully manual to fully autonomous. An SAE level 4 system is one that can autonomously control accel­ eration, steering, monitoring of the environment, and act correctly in a dynamic setting (National Highway Traffic Safety Administration, 2016). The highest level of automation, SAE level 5, is achieved when all the capabilities of an SAE level 4 system can be safely carried over to all driving situations such as merging onto a freeway, traffic jams, con­ struction detours, etc. The ability for a system to correctly execute ac­ celeration decisions, given a determined speed, relies entirely on the attributes of the system and is not dependent on infrastructure quality (e.g., vehicle to infrastructure communications). This technology has existed for some time, and almost all modern passenger cars have cruise control capabilities. However, cruise control settings still require users to input a speed to maintain, and therefore the task is automated as opposed to autonomous. The difference between the two terms is that an automated system is one that can only act under the confines of its pre-specified programs, whereas an autonomous system is characterized by its ability of self-governance and is a higher function than automation (Fitzpatrick et al., 2016). Achieving autonomous rather than automated acceleration requires that a system be able to ascertain speed limit information independently from the driver, which will require vehicle to infrastructure (V2I) communication. For the purposes of autonomous acceleration, the necessary level of V2I capable infrastructure investment will likely be modest. It is possible that a transmitter attachment is the only necessity for most situations; and the attachment would be able to convey speed limit postings to self-driving vehicles as they appear on the road. In theory, there are alternatives that can avoid the need for V2I infrastructure. For instance, a system could retain the speed limit post­ ing for all legs of a certain route or geographical region. However, a problem arises when speed limits change and is not immediately inte­ grated into the operating software. Another possible solution is to develop advanced visual detection systems so that a self-driving system can observe the speed limit posting independently. Yet, any such detection system would likely be less reliable without improved infra­ structure, especially under poor meteorological conditions. Fortunately, it may be possible to combine both methods, which would bypass the flaws of using only one of the two. To implement autonomous steering, a system must be able to reliably detect lane markings and possess independent navigation capabilities (Kockelman et al., 2017). For operational safety, it is likely that lane markings for much of the current roadways need to be improved. SAE level 3 or higher also requires that the automated system perform the task of monitoring the entire driving environment. This goes beyond simply monitoring the driving lane, and includes tasks such as monitoring traffic, traffic signals, and other roadway conditions that may require improved traffic signs and signals. Whether the current traffic signs and signals require improvement depends largely upon the development of sensory technology. Although most of these functions do not require major changes in infrastructure, efficiency gains could result from such improvements. For instance, it would be beneficial to design the driving environment around the autonomous system as opposed to designing the system around the driving environment. This can be achieved through dedicated autonomous lanes for freight and/or pas­ senger vehicles. If the driving environment is simplified and/or more predictable, then relying upon an autonomous system would be much safer and more efficient. A system reaches SAE level 4 if it does not require a human driver for “fall back” operations. Although the definition of a fall back operation is unclear and likely inconsistent among proprietary technologies, the concept itself is unequivocal. If a system is not confident about how to proceed in each situation or if it fails to act completely, it will request driver intervention. Once a system can overcome this dependency, then the infrastructure required to support level 4 vehicles would likely be slim.

Table 1 SAE defined levels (2018). SAE Level

Name

Definition

0

No Automation

1

Driver Assistance

2

Partial Automation

3

Conditional Automation

4

High Automation

5

Full Automation

Full-time performance by the human driver in all aspects of the dynamic driving task, even when enhanced by warning or intervention systems Strictly driver augmenting. Consists of the system performing very specific tasks under specific driving modes with the expectation that the human driver performs all remaining tasks. Same as level 1 except can have multiple driver assistance systems that coordinate with each other (also only under specific driving modes). The Automated Driving System (ADS) can perform all aspects of the dynamic driving task under certain driving modes with the expectation that the human driver will respond when requested to intervene. Same as level 3 but does not require a human intervention. Full time performance by an ADS for all aspects of the dynamic driving task. Steering wheel is optional.

of unsafe driving events will decline as drivers shift to autonomous driving (Short & Murry, 2016). To gauge the overall safety benefits from reducing unsafe driving events/behaviors requires knowledge about the proportion of all motor carrier accidents that are attributable to human factors. Much research has been conducted to examine the devastating effects caused by colli­ sions with motor carriers and large trucks (Bunn, Slavova, & Robertson, 2013; Islam et al., 2013; Choi, Oh, & Kim, 2014; Islam et al, 2012, 2014; Kin, Wang, & Cutler, 2014; Park, Kim, Kho, & Park, 2017; Zheng, Lu, & Lantz, 2018; Zhu et al., 2011). For example, Zhu et al. (2011) examined the effect of human, vehicle, crash and environmental characteristics using an ordered probit regression models on severity outcomes, given a crash occurs. The authors concluded opposing traffic situations and lack of vehicle familiarity poised the greatest safety risk and recommended an increase in driver training, greater vehicle familiarization and better traffic separation to mitigate the overall severity of large truck crashes (Zhu et al., 2011). Islam and Hernandez (2013) also employed ordered probit models to classify injury severity of large truck crashes and found interaction effects between human-factor and vehicle-factor variables. The authors concluded that further research is needed to improve truck driver safety (Islam & Hernandez, 2013). Additionally, Zheng et al. (2018) analyzed truck crash severity using a gradient boosting data mining model and concluded that the trucking company and driver characteristics as contributing factors have a significant effect on severity, given a commercial truck crash occurs. Thus, the empirical analyses in this paper abstracts away from such issues by focusing only on crashes in which the motor carrier driver contributed to the occurrence of the crash. To the extent that (1) unsafe driving events cause the observed crash, (2) the diffusion of AVs will affect the number of unsafe driving events, and (3) the diffusion of AVs will not introduce new safety problems, transportation policymakers have the responsibility to facilitate the adoption of such technologies. 1.1. Infrastructure needs The vision for the motor carrier industry is characterized by the dominance of highly autonomous trucks with a decentralized platooning capability. Decentralization is used here to refer to the ability of trucks to “match” with one another to form a platoon while enroute as opposed to matching before traveling. For this to manifest into reality, current infrastructure, especially for interstate roadways, must evolve with the demands from such technological advances. This section provides a comprehensive look at the infrastructure needs for AV trucks with SAE 2

J.M. Bernard Bracy et al.

Research in Transportation Economics xxx (xxxx) xxx

Similarly, the advancement from level 4 to 5 requires that all pre­ viously mentioned tasks be applicable in all dynamic driving situations such as poor weather, construction, etc. and implies that no additional changes in infrastructure is necessary. However, since SAE level 5 guarantees that driverless operations are safe, whereas anything below level 5 does not, additional infrastructural concerns arise. In the case of the motor carrier industry where benefits from AV technology are the highest among long haul operations, infrastructure investments are needed to address the refueling demands of self-driving freight trucks. Outside of dense urban areas, however, it is unclear whether it is necessary to develop or update existing gas stations to accommodate autonomous freight trucks. For example, a freight truck with two 150gallon diesel tanks and an average of 7 miles per gallon can drive up to 2100 miles before needing to refuel, then the route would simply need to include a leg with a nearby refueling station. Therefore, it may not be necessary for refueling points outside of cities. However, if the combustible engine becomes obsolete, then recharging stations designed for autonomous vehicles would be required. Alternatively, inductive charging technology could be incorporated onto the road itself or solar charging equipment would eliminate the need for autonomous vehicles to stop and recharge.

explored. After a crash occurs, the crash investigator identifies at least one circumstance from a pre-specified list that is found to have contributed to the crash at the driver level. Table 3 presents the contributing cir­ cumstances considered, and the frequency of occurrence for each contributing circumstance in which the motor carrier driver was found to have contributed to the cause of the crash. The top three contributing circumstances identified given a large truck crash occurred includes improper lane usage/change, driving while distracted/inattentive (including external distractions, passengers, stereo/audio/video equip­ ment, navigation device, communication device, eating/drinking, reading, tobacco use, grooming, computer equipment/electronic games/etc., and other), and driving too fast for conditions (i.e. not maintaining a safe driving speed during adverse road conditions). The contributing circumstances presented in Table 3 are included as binary (0,1) explanatory (predictor) variables in which the investigating officer either found (1) or did not find (0) such circumstance to be a contributory factor in the large truck crash. These explanatory variables are used to examine the crash severity outcomes, which are categorized on a three-class system: Fatality (1), Injury (2) and Property Damage Only (3). The crash severity level is determined from the most severely injured victim of the crash occurrence. Chi-square Automatic Interaction Detection (CHAID) decision tree is used as the methodological approach, and the model results are used to examine the potential safety en­ hancements of AV technology. Statistical analysis is performed using SPSS version 25 to build the CHAID model, which follows the algorithm proposed by Kass (1980), where the algorithm operates using a series of merging, splitting and stopping steps as follows.

2. Data and methodology The purpose of this section is to motivate public investment into the necessary infrastructure highlighted in the previous section by exploring potential gross safety benefits of employing AV technology. An investi­ gation into the safety benefits of AV diffusion requires a comparative analysis between the existing system (comprised of no AVs) and the complementary system (comprised of some AVs). However, such an ideal complementary system does not exist and/or is not observable. Instead, the analysis in this paper assumes that AVs do not introduce additional safety issues of their own and can eliminate all human error. In doing so, we can construct the safety outcomes of the complementary system by focusing the analysis on contributing circumstances to crashes and removing the human driver from the equation. We analyze large truck crash data from 2013 through 2015 obtained from the Missouri State Highway Patrol Statewide Traffic Accident Re­ cords System (STARS) database. The three datasets utilized - accident level data, vehicle level data and personal level data - are categorized in the Missouri State Highway Patrol Record Specification form and contain an array of information that is linked together using the accident number and person number. The combined datasets resulted in 1,083,150 unique records. Motor carriers (defined here as a single-unit truck with two or more axles, truck tractors, and other heavy trucks) were involved in 6.8% of the total crashes in Missouri from 2013 to 2015 (28,754 out of the 425,374 crashes), and motor carrier drivers were found to have contributed to 15,338 of these crash occurrences that resulted in 105 fatal and 1596 injury outcomes, as shown in Table 2. Previous studies reported contributing circumstances and driver actions are critical in determining severity outcomes, given a crash oc­ curs (Bernard & Mondy, 2016; Chang & Chien, 2013; Chang & Wang, 2006). Following these studies, the dataset is filtered to only consider crashes in which a motor carrier driver is found to have contributed to a crash occurrence, and the crash contributing circumstances are

Table 3 Frequency of circumstance contributing to a large truck crash.

Table 2 Frequency of crash severity outcomes in Missouri from 2013 to 2015 when a large truck contributed to the crash occurrence. Crash Severity Outcome

Frequency

Percent

Cumulative Percent

Fatal Injury Property Damage Only Total

105 1586 13,647 15,338

.7 10.3 89.0 100.0

.7 11.0 100.0

Contributing Circumstance

Counts

Percentage of Total

Improper Lane Use/Change Distracted/Inattentive Too Fast for Conditions Improper Turn Failed to Yield Following too Close Other Vehicle Defects Improper Backing Vision Obstructed Animal in Roadway Failed to Secure Load Improper Passing Overcorrected Object in Roadway Driver Fatigue/Asleep Violation of Signal/Sign Wrong Side (Not Passing) Alcohol Physical Impairment Improperly Stopped on Roadway Speed - Exceed Limit Improperly Parked Drugs Improper Towing/Pushing Improper Signal Wrong Side (One-Way) Improper Start from Park Failed to Use Lights Improper Riding Failed to Dim Lights Total

2877 1726 1699 1327 1243 1137 1082 758 629 430 414 383 360 326 270 263 256 250 175 132 121 109 78 56 32 30 26 19 6 3 0 16,217a

17.74% 10.64% 10.48% 8.18% 7.66% 7.01% 6.67% 4.67% 3.88% 2.65% 2.55% 2.36% 2.22% 2.01% 1.66% 1.62% 1.58% 1.54% 1.08% 0.81% 0.75% 0.67% 0.48% 0.35% 0.20% 0.18% 0.16% 0.12% 0.04% 0.02% 0.00% 100.00%

a The sum of the frequency of contributing circumstance can exceed the number of cases, since multiple citations of contributing circumstance may be present in a given crash.

3

J.M. Bernard Bracy et al.

Research in Transportation Economics xxx (xxxx) xxx

1. The Pearson measure is used as the chi-square measure to test for independence for categorical targets, and if the p-value is larger than 0.05, the pair is merged into a single category and a new set of cat­ egories is formed. 2. Any category having less than 100 records is merged with the most statistically homogenous category. 3. The adjusted p-value is calculated for the merged categories using Bonferroni adjustments, and the adjusted p-values for each predictor are compared to determine which predictor best splits the node. The predictor with the smallest adjusted p-value is selected. If the adjusted p-value is less than or equal to 0.05 the node is split using this predictor; otherwise, the node becomes terminal. 4. The tree will stop growing and the node will not be split if all cases in a node have identical values of the dependent variable or of each predictor, if the tree reaches a maximum tree depth of 15 branches, if the size of a node is less than 100, or if the node results in a child node whose size is less than 50.

3. Results The results from the CHAID algorithm are visually depicted in Fig. 1, where each node represents an if/then rule. The root node (Node 0) represents the parent node for the model and illustrates the number of occurrences for each severity outcome in the dataset (fatal, 105; injury, 1586; property damage only, 13,647). The initial model split, Too_­ Fast_for_Conditions, suggests that driving too fast and not maintaining a safe speed for adverse road conditions is the greatest contributory pre­ dictor for crash severity, given a crash occurs. If driving Too_­ Fast_for_Conditions is present (Node 1), then the probability of the severity outcome, given a crash occurs, is 0.9%, 21.5% and 77.5% for fatal, injury and property damage outcomes respectively. If driving Too_Fast_for_Conditions is present and Distracted_Inattentive driving is present (Node 4), then the risk of a fatal outcome and an injury outcome, given a crash occurs, increase to 4.5% and 22.4% respectively. In addition to driving too fast for conditions and distracted/inat­ tentive driving, CHAID decision tree results suggest that the greatest contributory predictors of severity outcomes for crashes in which a large truck driver was found to have contributed are overcorrecting and driving under the influence of alcohol. Importantly, model results sug­ gest that driving under the influence of alcohol is the most dangerous behavior examined, and when present the severity of crashes increase to 4.8% and 37.0% for fatal and injury outcomes respectively. Using the results from the CHAID decision tree, historic outcomes are examined to determine upper and lower bounds on the changes in the

The CHAID decision tree approach is selected due to the advantages over other models, which include the data partitioning yields insights into input/output relationships; each path of the decision tree contains an estimated risk factor; nonlinear relationships between variables do not affect performance; missing values are accommodated automati­ cally; and the output is simple to understand and interpret (Bernard Bracy, 2017).

Fig. 1. CHAID Decision Tree Results for Motor Carrier Drivers who contributed to a Crash. 4

J.M. Bernard Bracy et al.

Research in Transportation Economics xxx (xxxx) xxx

number of drivers involved in fatal, injury or property damage only crashes if selected contributory circumstances are individually elimi­ nated because of AV technology. Bounds for changes in the annual number of drivers involved in each level of severity outcomes are calculated by 1) removing the contributing circumstance for each driver and assuming the crash still occurs with severity outcome probabilities now determined by the outcome probabilities of the complementary node (a lower bound) and 2) removing the contributing circumstance and assuming that the driver is not involved in a crash at all (an upper bound) (Bernard Bracy, 2017). The analysis provides a context for un­ derstanding the relative reduction in risk associated with reducing the frequency of circumstances likely to contribute to different crash severity outcomes, which can provide justification for the necessary AV infrastructural modifications. Table 4 presents the estimated reductions in the number of drivers involved in each severity outcome if the considered contributing circumstance is eliminated. The lower bound is estimated by removing the contributing circumstance and assuming the crash still occurs with severity outcome probabilities determined by the outcome probabilities of the complementary node (N1fatal – (N1Total * N2probablity), where N1¼Node 1, N2¼Node 2) For example, if the contributing circumstance of driving too fast for conditions was eliminated as a result of intro­ ducing automation that slowed the vehicle during adverse weather, yet the crash still occurred, the 1699 crash outcomes with driving too fast for conditions present (Node 1) would take on the probabilities of the adjacent node where driving too fast for conditions was not a present contributing circumstance (Node 2), and the lower bound of fatal crashes would be calculated as 15 – (1699*.007) ¼ 3.107. The upper bound is calculated by removing the contributing circumstance and assuming that the driver is not involved in a crash at all. For example, if driver assistance or AV technologies kept vehicles from drifting out of the lane and eliminated the need for drivers to overcorrect, which eliminated the crash completely, then 8 fatal outcomes could be avoided over this period (calculated by adding Node 6 and Node 8 fatal out­ comes, six and two respectively). If the significant contributing cir­ cumstances of driving too fast for conditions, distracted/inattentive driving, overcorrecting and alcohol use are all altered by AV technology, it is suggested that between 117 and 193 severe crashes (fatal and injury outcomes) involving large trucks could be prevented annually in Mis­ souri alone. For AV technologies to mitigate the safety consequences of over­ correcting, improper lane usage, driving too fast for conditions, distracted/inattentive driving and alcohol use, said technologies must be able to autonomously control acceleration and steering, monitor the environment, and respond to dynamic driving environments without the need for human intervention, as presented in Table 5. Table 5 also shows which contributing circumstances we expect would be eliminated at each SAE level of automation. For instance, if all truck crashes associ­ ated with overcorrecting and improper lane usage were equipped with SAE level 1 technologies, then we expect those contributing circum­ stances to be eliminated. This table also reflects our own conjectures about the implications of implementing higher SAE levels on safety benefits. Thus, using Table 5 in conjunction with Table 4 produces

predictions of safety benefits afforded by the implementation of different SAE levels. Importantly, safe operations of a system that can perform these tasks autonomously require readable lane markings and traffic signals and signs. To focus specifically on improper lane usage and overcorrecting to reduce both crash frequency and severity, the vehicle contributing cir­ cumstances of ‘Ran off Roadway’ in which the vehicle exits the roadway to the left, right or other, and ‘Crossed Center of Road’ in which the vehicle crosses the center of the road into the lane of coming traffic are explored. Table 6 provides the number of ran-off-roadway and crossed center lane crash types by level of crash severity. When including these variables with the aforementioned contrib­ uting circumstances into the decision tree model, the aggregated ran-offroadway crashes is the most important variable examined to explain crash severity outcomes and other variables previously found statisti­ cally significant are excluded from the model, as illustrated in Fig. 2. When the motor carrier runs off the roadway to the left, right or other, the probability of a fatal outcome is 1.8%, given a crash occurs (as presented in Fig. 2, Node 2). Additionally, the circumstances of runs off the roadway and crosses the center of the road (which results from overcorrecting) are present, the probability of a fatal outcome increases to 3.5% (as presented in Fig. 2, Node 10). Table 7 uses the results from the CHAID decision tree that includes ran-off-roadway and crossed center of road crash types (Fig. 2), historic outcomes are re-examined to determine upper and lower bounds on the changes in the number of drivers involved in fatal, injury or property damage only crashes if these crash circumstances are individually eliminated because of AV technology. It is found that, by removing the crash circumstance for each driver and assuming the crash still occurs with severity outcome probabilities now determined by the outcome probabilities of the complementary node, 727 of severe crashes will be reduced and, by removing the contributing circumstance and assuming that the driver is not involved in a crash at all, 1079 of severe crashes will be reduce. However, to mitigate the effects of these crash type variables on crash outcomes, a system must be able to reliably detect lane markings and possess independent navigation capabilities in order to implement autonomous steering. Focusing on Missouri highways, defined as either rural interstate, rural other expressways/highways, rural other principal arterials, urban interstate, urban other expressways/highways, or urban other principal arterials, we calculate initial estimates for the cost of readable lanes and paving the way for early adopters of this new technology. First, the number of highway miles that require new lane markings using Federal Highway Administration (FHA) provided statistics that contains data on highways conditions for each state are estimated. Pavement condition classifications are based on the international roughness indicator (IRI) and if a certain segment sample exceeds a value of 170 on the IRI then that portion of the road is classified as “poor”. For the purposes of this introductory examination, we assume that if a mile of highway is in poor condition, then it contains lane markers that are not readable under some conditions rendering it dangerous for AV operations. All highways with portions classified as

Table 4 Estimated Reductions in the Number of Drivers Involved in Each Severity Outcome if a Contributing Circumstance is Eliminated. Contributing Circumstance

Too Fast for Conditions Overcorrecting Distracted/Inattentive Driving Alcohol Use a b

Fatal

Injury

Na

Property Damage Only

Est Lower Bound

Est Upper Bound

Est Lower Bound

Est Upper Bound

3 4 2 6

15 8 3 7

215 77 1 42

366 111 15 54

Est Lower Boundb 218 82 3 48

Est Upper Bound 1318 202 49 85

N ¼ Number of estimated cases for the three-year period and equal to the sum of the estimated upper bounds. A negative value for property damage only outcome represents an increase for the least severe outcome, given the assumption that the crash still occurs. 5

1699 321 67 146

J.M. Bernard Bracy et al.

Research in Transportation Economics xxx (xxxx) xxx

Table 5 Comparison of SAE defined levels (2016), AV/CV technologies, and crash contributing circumstances. SAE Level

Name

Execution of Steering and Acceleration

Monitoring of Driving Environment

Fallback Performance of Dynamic Driving Task

System Capability (Driving Modes)

Contributing Circumstances

1

Driver Assistance

Human and System

Human

Human

Some driving modes

3

Conditional Automation High Automation

System

System

Human

Some driving modes

System

System

System

Some driving modes

Overcorrecting And Improper Lane Usage Too fast for Conditions and Distracted Driving Alcohol Use

4

replacement time of three years. Therefore, cost figures presented here are based on the cost of thermoplastic lane markings. According to a study done by the FHA in 2005, the cost per linear foot of markings was $2.3476 (Gillespie, 2005). Adjusting for inflation, the cost per linear foot in 2018 dollars is roughly $3.01. Table 8 illustrates the information and the infrastructure investment cost needed (Table 9). In Missouri, the most dangerous roadways by crash frequency are Highway 55, 44 and 70 – resulting in 6%, 11% and 13% of all crashes, respectively. Table 9 depicts the total frequency of crash severity when running off the roadway or crossing the center of the road contributed to the crash occurrence. Based on MoDOT’s Transportation Assessment Management Plan, all sampled portions of the above three highways were all classified as either good or fair and none were poor in quality. Based on the method from the previous analysis, this indicates that no new high-quality lane markers are needed for these highways and that

Table 6 Frequency of crash type by crash severity. Crash Type Ran off Roadway – Left Ran off Roadway – Right Ran off Roadway – Other Crossed Center of Road Total

Crash Severity

Total

Fatal

Injury

Property Damage Only

32 37 1 24 94

328 601 21 215 1165

1088 1823 70 688 3669

1448 2461 92 927 4928

poor are then considered to need new lane markings and derive a cost estimate for such an infrastructure investment. Although many types of materials are available for lane markings, the most durable are thermoplastics and epoxy resin with average

Fig. 2. CHAID Decision Tree Results for Motor Carrier Drivers who contributed to a Crash with Ran off Roadway and Crossed Center of Road Crash Type included. 6

J.M. Bernard Bracy et al.

Research in Transportation Economics xxx (xxxx) xxx

Table 7 Estimated Reductions in the Number of Drivers Involved in Each Severity Outcome if a Contributing Circumstance is Eliminated. Crash Circumstance

Fatal

Ran off Roadway Crossed Center of Road a b

Injury

Na

Property Damage Only

Est Lower Bound

Est Upper Bound

Est Lower Bound

Est Upper Bound

53 18

64 19

593 63

812 184

Est Lower Boundb 623 74

Est Upper Bound 2740 341

3616 544

N ¼ Number of estimated cases for the three-year period and equal to the sum of the estimated upper bounds. A negative value for property damage only outcome represents an increase for the least severe outcome, given the assumption that the crash still occurs.

Table 8 Missouri highway data and lane marking cost estimates (U.S. Federal Highway Administration, 2016). Functional Classification

Lane Miles

% Poor 2015

Poor Lane Miles

Poor Lane Feet (rounded)

Lane Marking Cost (rounded)

Rural Interstate Rural Other Expressways/Highway Rural Other Principal Arterials Rural Minor Arterials Urban Interstate Urban Other Expressways/Highway Urban Other principal Arterials Total Investment Cost

3426 4097 4666 62 3113 2178 3381 $ 29,510,209.63

0.70% 0.68% 1.99% 6.62% 4.41% 3.90% 16.48%

23.87 27.86 92.92 4.10 137.43 84.90 557.32

126,051 147,106 490,629 21,657 725,645 448,283 2,942,657

$ 758,827 $ 885,57 $ 2,953,587 $ 130,376 $ 4,368,385 $ 2,698,664 $ 17,714,796

markings to capitalize on potential safety benefits. Other infrastructure investments needed for efficiency gains are the various V2I communi­ cation upgrades on existing infrastructure to facilitate acceleration/ speed control and dedicated lanes for AVs. Dedicated lanes could help mitigate the uncertain nature of safety issues that arise due to just AV technologies. Limitations of this study do exist. First, it is difficult to predict when AV technology will be widely adopted, and two separate timelines must be considered: the development stages of the technology and the adoption stages of the developed technology. Elon Musk, CEO of Tesla anticipates a fully autonomous Tesla car to be on market by 2020, and many other motor companies have similar aspirations (Ohnsman, 2017). Using Texas survey data and a multinomial logit model to predict market adoption rates of level 4 autonomous cars, Kockelman et al. (2017) predicts that 3.4%–38.5% percent of households with at least one vehicle that will adopt AV technology by 2045. Second, this study only considers data from Missouri, while the implementation of such tech­ nologies would be nationwide. Finally, assumptions were made due to the lack of observation of safety outcomes under both an established system with AVs and the current system, and infrastructure re­ quirements and safety concerns regarding the implementation of such technologies in the motor carrier industry require additional research efforts.

Table 9 Frequency of running off roadway and crossing center of road by crash severity for roadways in Missouri with greatest percentage of crashes. Highway 44 55 70 Total

Crash Severity

Total

Fatal

Injury

Property Damage Only

25 5 13 43

939 25 1070 2034

3593 144 2600 6337

4557 174 3683 8414

the only requirement would be maintenance and potentially saving 14 lives annually. The above suggests that implementing high quality lane markings for all highways within Missouri in anticipation of AV adoption would cost approximately $30 million. Moreover, if the distribution of poor pave­ ment is roughly constant every year, then the estimated amount would need to be incurred annually. Fortunately, the benefits of this invest­ ment can be partially realized immediately in terms of safety enhance­ ment in addition to meeting future needs induced by AV technology. However, the magnitude of this immediate safety benefit is uncertain considering that highways 44, 55, and 70 all have readable lane mark­ ings, but are the most dangerous highways in Missouri by crash frequency.

Acknowledgements

4. Conclusion

We thank the U.S. Dept. of Transportation for their funding and support through their grant to Iowa State and the Midwest Trans­ portation Consortium (grant #: DTRT13-G-UTC37).

Although the infrastructure challenges from the employment of AV motor carrier technologies are not insurmountable, they remain largely unexplored and inadequately investigated. As a result, this study con­ siders key infrastructure needs for AV technology to be widely and safely used in the motor carrier industry and motivates such infrastructure investments with estimated safety benefits realized from the imple­ mentation of such technologies. Model results suggest that the greatest contributory predictors of crash severity, given a crash occurs, are driving too fast for conditions, distracted/inattentive driving, over­ correcting and driving under the influence of alcohol. To render such safety benefits, key vehicle needs include autonomously controlling acceleration and steering, monitoring of the environment, and responding to dynamic driving environments without the need for human intervention. Importantly, the safe operations of a system that can perform such AV tasks require, at a minimum, readable lane

References Bernard Bracy, J. M. (2017). An application of decision tree models to examine motor vehicle crash severity outcomes. Journal of the Transportation Research Forum, 56(2), 73–92. Bernard, J. M., & Mondy, C. M. (2016). Correlation of driver gender with injury severity in large truck crashes in Missouri. Transportation Research Record: Journal of the Transportation Research Board, 2585, 49–58. Bhoopalam, A. K., Agatz, N., & Zuidwijk, R. (2017). Planning of truck platoons: A literature review and directions for future research. Transportation Research Part B: Methodological. Bunn, T., Slavova, S., & Robertson, M. (2013). Motor vehicle injuries among semi truck drivers and sleeper berth passengers. Journal of Safety Research, 44, 51–55. Chang, L. Y., & Chien, J. T. (2013). Analysis of driver injury severity in truck-involved accidents using a non-parametric classification tree model. Safety Science, 51(1), 17–22.

7

J.M. Bernard Bracy et al.

Research in Transportation Economics xxx (xxxx) xxx FHWA/TX-17/0-6847-1. University of Texas at Austin. Center for Transportation Research. National Highway Traffic Safety Administration. (2016). Federal automated vehicles policy: Accelerating the next evolution of roadway safety. U. S. Department of Transportation. Ohnsman, A. (2017). Musk wants to begin shifting teslas to ’full’ self driving capability within 6 months. Forbes. January 24. Accessed https://www.forbes. com/sites/alanoh nsman/2017/01/24/elon-musk-targets-full-self-driving-capability-for-teslas-with in-6-months/#2938aa5579ee. (Accessed 12 June 2017). Park, H. C., Kim, D. K., Kho, S. Y., & Park, P. Y. (2017). Cross-classified multilevel models for severity of commercial motor vehicle crashes considering heterogeneity among companies and regions. Accident Analysis & Prevention, 106, 305–314. SAE International. (2018). Taxonomy and definitions for terms related to driving automation systems for on-road motor vehicles. Warrendale: SAE International. Short, J., & Murry, D. (2016). Identifying autonomous vehicle technology impacts. Arlington: American Transportation Research Institute. U.S. Federal Highway Administration. (2016). Office of highway policy information. Highway statistics. U.S. Department of Transportation https://www.fhwa.dot. gov/policyinformation/statistics/2016/. Zheng, Z., Lu, P., & Lantz, B. (2018). Commercial truck crash injury severity analysis using gradient boosting data mining model. Journal of Safety Research, 65, 115–124. Zhu, X., & Srinivasan, S. (2011). A comprehensive analysis of factors influencing the injury severity of large-truck crashes. Accident Analysis & Prevention, 43(1), 49–57.

Chang, L. Y., & Wang, H. W. (2006). Analysis of traffic injury severity: An application of non-parametric classification tree techniques. Accident Analysis & Prevention, 3(5), 1019–1027. Choi, S., Oh, C., & Kim, M. (2014). Risk factors related to fatal truck crashes on Korean freeways. Traffic Injury Prevention, 15(1), 73–80. Fitzpatrick, D., Cordahi, G., O’Rourke, L., Ross, C., Kumar, A., & Bevly, D. (2016). Challenges to CV and AV applications in truck freight operations. No. NCHRP Project, 20–102(03). Gillespie, J. S. (2005). Enhanced night visibility series, volume XI: Phase II—cost-Benefit analysis. Office of Safety Research and Development. Enhanced Night Visibility Series Volume XI: (Phase II—Cost-Benefit Analysis). Islam, M., & Hernandez, S. (2012). Multi-vehicle Collisions involving Large Trucks on Highways: An Exploratory Discrete Outcome Analysis, 99. CUTR Faculty and Staff Publications. https://scholarcommons.usf.edu/cutr_facpub/99Attachment. Islam, M., & Hernandez, S. (2013). Large truck–involved crashes: Exploratory injury severity analysis. Journal of Transportation Engineering, 139(6), 596–604. Islam, S., Jones, S. L., & Dye, D. (2014). Comprehensive analysis of single- and multivehicle large truck at-fault crashes on rural and urban roadways in Alabama. Accident Analysis & Prevention, 67, 148–158. Kass, G. V. (1980). An exploratory technique for investigating large quantities of categorical data. Applied Statistics, 119–127. Kin, X., Wang, K., & Cutler, C. (2014). Logistic regression models of the safety of large trucks. Transportation Research Record: Journal of the Transportation Research Board, 1–10. Kockelman, K., Sharon, G., Simoni, M., Albert, M., Fritz, H., Hutchinson, R., et al. (2017). An assessment of autonomous vehicles: Traffic impacts and infrastructure needs. No.

8