Attachment 3 - Modeling Capabilities and Data Needs for Assessing the Operational Impacts of CAVs - V1.PDF

Modeling Capabilities and Data Needs for Assessing the Operational Impacts of Connected and Automated Vehicles

Ian Berg, Andrew Berthaume, Rachel James

January 2020

Publication Number

DRAFT COPY

FOR REVIEW.

DO NOT

DISTRIBUTE.

Notice

This document is disseminated under the sponsorship of the U.S. Department of Transportation in the interest of information exchange. The U.S. Government assumes no liability for the use of the information contained in this document.

The U.S. Government does not endorse products or manufacturers. Trademarks or manufacturers’ names appear in this report only because they are considered essential to the objective of the document.

Quality Assurance Statement

The Federal Highway Administration (FHWA) provides high-quality information to serve Government, industry, and the public in a manner that promotes public understanding.

Standards and policies are used to ensure and maximize the quality, objectivity, utility, and integrity of its information. FHWA periodically reviews quality issues and adjusts its programs and processes to ensure continuous quality improvement.

DRAFT COPY

FOR REVIEW.

TECHNICAL REPORT DOCUMENTATION PAGE

1. Report No.

2. Government Accession No.

3. Recipient’s Catalog No.

4. Title and Subtitle Modeling Capabilities and Data Needs for Assessing the Operational Impacts of Connected and Automated Vehicles

5. Report Date January 2020

6. Performing Organization Code

7. Author(s) Ian Berg, Andrew Berthaume, Rachel James

8. Performing Organization Report No.

9. Performing Organization Name and Address 10. 10. Work Unit No. (TRAIS) Volpe National Transportation Systems Center 55 Broadway Cambridge, MA 02142

11. Contract or Grant No.

12. Sponsoring Agency Name and Address U.S. Department of Transportation Federal Highway Administration 1200 New Jersey Avenue, SE Washington, DC 20590

13. Type of Report and Period Covered

14. Sponsoring Agency Code

15. Supplementary Notes [insert name of GTM]

16. Abstract

17. Key Words

18. Distribution Statement No restrictions.

19. Security Classif. (of this report) Unclassified

20. Security Classif. (of this page) Unclassified

21. No. of Pages

22. Price n/a

Form DOT F 1700.7 (8-72) Reproduction of completed page authorized.

DRAFT COPY

FOR REVIEW.

ii

DRAFT COPY

FOR REVIEW.

iii

TABLE OF CONTENTS

CHAPTER 1. INTRODUCTION AND PURPOSE

CHAPTER 2. BACKGROUND AND LITERATURE REVIEW

CHAPTER 3. DESIRED MODELING CAPABILITIES AND DRIVING SCENARIOS

STAKEHOLDER FEEDBACK

DESIRED MODELING CAPABILITIES AND SCENARIOS

List of Desired Capabilities List of Desired Scenarios

CHAPTER 4. DATA REQUIREMENTS FOR DESIRED CAPABILITIES

QUALITY AND QUANTITY

COLLECTION METHODS

STAKEHOLDER FEEDBACK

DATA REQUIREMENTS

Instrumented Vehicle Data Non-IRV Data

APPENDIX. LISTENING SESSION PARTICIPANTS

REFERENCES

iv

LIST OF TABLES

Table 1. Vehicle characteristics data Table 2. Roadway characteristics data Table 3. Roadway conditions data Table 4. Car-following data Table 5. Platooning data Table 6. Lane-changing data Table 7. Driver data Table 8. Non-IRV data v

LIST OF ACRONYMS

ACC adaptive cruise control

AMS analysis, modeling and simulation

CACC cooperative adaptive cruise control

CAN bus Controller Area Network bus

CAV connected and automated vehicle

DLC discretionary lane change

DOT Department of Transportation

FHWA Federal Highway Administration

GPS Global Positioning System

IOO infrastructure owner-operator

IRV instrumented research vehicle

KPI key performance indicator

LiDAR light detection and ranging

MLC mandatory lane change

MPO metropolitan planning organization

NGSIM Next Generation Simulation

ODD operational design domain

SHRP2 NDS Second Strategic Highway Research Program Naturalistic Driving Study

TSMO transportation systems management and operations

V2I vehicle-to-infrastructure

V2V vehicle-to-vehicle

UAV unmanned aerial vehicle

DRAFT COPY

FOR REVIEW.

CHAPTER 1. INTRODUCTION AND PURPOSE

The deployment of connected and automated vehicles (CAVs) will impact the ability of infrastructure owners and operators (IOOs) to predict traffic flow and future infrastructure needs, due to differences in driving behaviors and interactions between CAVs and human-driven vehicles. Understanding these effects requires new analysis, modeling, and simulation (AMS) capabilities. Each of these capabilities requires data collection on the behaviors sought to be reproduced, and such data must be sufficient in quality and quantity. This document, the result of literature review and stakeholder engagement, aims to outline the AMS capabilities that practitioners desire to understand the effects of the introduction of CAVs on traffic, as well as provide a description of the data that would be required to develop these capabilities.

This report synthesizes comments from two listening sessions, which targeted practitioners, IOOs, and academic researchers. In these sessions, stakeholders developed a prioritized list of desired microsimulation modeling capabilities for assessing the performance and impacts of CAVs – capabilities which do not yet exist due to the absence of CAV field data. Because these capabilities may initially rely on a certain set of operational assumptions, participants also listed the driving scenarios that are most important to develop for their modeling needs. Additionally, this report includes a comprehensive list of attributes, model inputs, and data elements required to develop these desired capabilities, based on literature review and previous findings. This list also includes information about the data quality (precision, accuracy, fidelity, etc.) and quantity (collection interval, number of runs, etc.) necessary to develop a useful model, to the extent available.

In the immediate term, this report will inform the data collection plan and choice of model to be developed from that data in an ongoing FHWA research project. This effort, like any project, will be constrained by available technology and data collection methods, as well as schedule and funding limitations. Beyond this particular effort, however, this document aims to provide long-term insights for researchers both within FHWA and at other agencies, institutions, and companies about the types of models that will be useful for practitioners in determining the effects of CAVs, as well as the data collection capabilities necessary to construct those models.

Furthermore, this document will ensure that the dataset collected and published as a result of this project is of maximum utility to researchers.

CHAPTER 2. BACKGROUND AND LITERATURE REVIEW

As vehicle automation and connectivity technologies begin to enter real-world traffic, new challenges arise in predicting traffic flow and infrastructure needs. For example, a recent report summarizes potential transportation systems management and operations (TSMO) use cases where CAVs may need to handle special circumstances: for traffic incident management, the ability to yield to first responders; for road weather management, to increase headways on wet pavement; and for work zone management, to navigate lane closures. The report lists the various functional requirements of a system to handle these scenarios, and lists possible performance measures for the systems (Birriel et al. 2020).

As such, being able to predict CAV behavior, and their impact on traffic operations, in a wide variety of scenarios will become more important to practitioners as these technologies start to be deployed. However, these vehicles’ behaviors (and manual drivers’ interactions with them) will not stay constant. As technologies mature and penetration rates increase, CAVs’ impacts on traffic will change over time.

One common tool used in predicting traffic operations is microscopic traffic simulation, which calculates each vehicle in a modeled network’s position based on algorithms for the vehicle’s longitudinal (car-following) and lateral (lane-changing) behavior. Because each individual vehicle is simulated, it is able to capture detailed information about the network’s performance even if there are varying types of vehicle behavior – important if CAVs and human drivers will operate in mixed traffic at varying market penetration levels. Microsimulation outputs can also be useful in creating inputs for other modeling efforts where microsimulation as the sole analysis method would be too resource-intensive or unnecessarily detailed (e.g., on a regional scale).

One major component of any microsimulation tool is the underlying car-following model. In general, a car following model describes how a driver (or, in the case of a CAV, an automated system) responds to environmental stimuli (Toledo 2007). These models can be categorized by the ways that they attempt to replicate this behavior: for example, some of these categories include trajectory translation, safety-distance, social-force, and psychophysical models (James, Hammit, and Boyles 2019). While models differ in complexity, they each have a variety of parameters that can be calibrated. Some researchers have worked on replicating CAV behavior by adjusting parameter values in existing models like Wiedemann 99 (Sukennik 2018; Sukennik, Zeidler, and Van Etten 2018). However, automated driving behaviors may be different enough from manual driving that current models do not have the necessary parameters to replicate CAVs, and new models may be necessary.

In contrast to car-following, lane-change models – another key component in microsimulation – focus on the vehicle’s lateral movements. These models typically divide lane change movements into mandatory lane changes (MLC) and discretionary lane changes (DLC). Once drivers desire to change lanes, they must then determine whether there is enough of a gap in traffic for them to merge into the target lane (Ben-Akiva, Choudhury, and Toledo 2009). However, a driver’s decision to seek an opportunity to change lanes is not easily observed – in particular, it is difficult to tell which gaps the driver rejected before changing lanes (Toledo 2007). However, DRAFT COPY

FOR REVIEW.

some field research has been conducted on more observable gap-acceptance decisions; for example, left-turn gap acceptance at a signalized intersection (Davis and Swenson 2004).

Existing microsimulation research studying the operational effects of autonomous vehicles has had with varying results. For example, recent simulation studies of ACC and CACC performance have estimated changes to road capacity and throughput from about a 25% decrease to 100% increase, due to differing researcher assumptions including (among others) the car-following model used, time gap/following distance settings, market penetration, and sensor delay (Eilbert, Berg, and Smith 2019). Another review notes that none of the simulations they identified used field data to calibrate their models, and almost all studies focus exclusively on the longitudinal behavior of CAV technologies – they generally do not model the traffic impacts of CAV lane-changing behavior, or making no adjustments to the default lateral model (Do, Rouhani, and Miranda-Moreno 2019).

Other FHWA projects have worked to identify gaps in existing tools’ abilities to evaluate CAV operational performance. For example, one report found that today’s tools do not represent (Mahmassani et al. 2018, 93–100):

• The effect of wireless telecommunication networks and information flow.

• Sensor performance and reliability aspects that influence vehicle performance.

• Differences in human driving behaviors for conventional- and connected-vehicle drivers.

• Vehicle-following and lane-changing for diverse AV systems.

• Interactions of CAVs with vulnerable road users.

That report also noted the following data sources used in published CAV simulation studies (Mahmassani et al. 2018, 65–68):

• Hypothetical and real road networks.

• Traffic data, both macroscopic and trajectory-level.

• Data collected in the field from instrumented ACC/CACC vehicles.

Hidas and Wagner (2004) note that the ideal datasets for the development and calibration of microsimulation tools are full vehicle trajectories, which can be collected through video recording or vehicle instrumentation – for example, with a differential Global Positioning System (GPS) receiver or an accelerometer. However, they note that such datasets are difficult to collect, and thus there is limited data available (Hidas and Wagner 2004). These limits are important to note because the same driver can behave differently under different road types and conditions (Berthaume, James, et al. 2018; Berthaume, Jackson, et al. 2018). Recent research also shows that human drivers follow automation-driven vehicles differently than human-driven vehicles, and this can affect string stability (Rahmati et al. 2019; Stern et al. 2018).

Beyond models of vehicle behavior, additional data is needed to apply the underlying behavioral models to a specific study (Wunderlich et al. 2019). For example, to develop a network on which to run microsimulations, road geometries and traffic controls (signals and signs) must be modeled. Then, the demand on the network must be determined, as well as other travel conditions like weather or incidents. Parameters for vehicle and driver characteristics should also

DRAFT COPY

FOR REVIEW.

be configured and validated. If these factors are not calibrated correctly, the resulting analyses will be of limited utility.

Each of these datasets have varying required levels of effort to collect – for example, origin/destination demand and routing behavior are difficult to observe directly (Hellinga 1998).

Even as technical limitations decrease, and new options like cellphone data make collecting this data more technically feasible, challenges remain in making such data available to researchers.

Both privacy (for travelers) and intellectual property concerns (from manufacturers) are obstacles to research dataset availability (Excell et al. 2019). CAV modeling, particularly of connected vehicles, may also require the development of methods to include parameters in the simulation that do not exist in a model of manual driving, such as wireless communication and signal loss or delay (Excell et al. 2019).

Despite these challenges, successful microscopic data collection efforts that have provided value to researchers can provide insight on best practices. For example, the Next Generation Simulation (NGSIM) program collected fifteen minute stretches of vehicle trajectories from overhead cameras, as well as data from instrumented vehicles and detectors (Alexiadis et al.

2004). Another example is the Second Strategic Highway Research Program’s (SHRP2) Naturalistic Driving Study (NDS), which instrumented over 2,000 study participants’ vehicles and collected their driving data for up to two years (Campbell 2012).

However, measurement errors and accuracy limitations can necessitate extensive preprocessing of these types of data before they become useful. For example, some academics have found discrepancies in the Next Generation Simulation (NGSIM) dataset – finding that 13% of the calculated trajectories collide with the vehicle ahead in a 45-minute study period (Coifman and Li 2017). In particular, research has focused on cleaning and processing the acceleration values in the NGSIM dataset – a parameter which is key in characterizing car-following behavior (Montanino and Punzo 2013; Coifman and Li 2017).

Additionally, existing datasets are missing a key component to building models of CAV behavior

– they were collected before CAV technologies were ready to be deployed. A lack of naturalistic CAV driving data is a significant gap hindering the development of robust models of CAV performance in the traffic stream. Very few datasets exist for modeling the performance of a single CAV or a string of CAVs in naturalistic driving conditions; moreover, no datasets currently exist documenting how human driven vehicles react to automated vehicles in naturalistic driving conditions. Modelers use field data for model development, verification, calibration, and validation, and without information about how real vehicles behave, it is impossible to tell whether a model’s predictions of CAV traffic impacts is realistic.

On a longer time frame, CAVs may cause a range of changes in land use. Some suggest that CAVs could lead to narrower travel lanes, increasing the right-of-way dedicated to other road users (Stead and Vaddadi 2019). On the other hand, if CAVs make commuting easier, people may be willing to commute longer distances, and adjust residential land use patterns accordingly (Soteropoulos, Berger, and Ciari 2019). While these changes may not directly affect CAVs’ driving behavior, they could still have an impact on the utility of microsimulation results by affecting the network layouts or input volumes used in simulation.

DRAFT COPY

FOR REVIEW.

CHAPTER 3. DESIRED MODELING CAPABILITIES AND DRIVING SCENARIOS

This chapter summarizes the results of the listening sessions and the prioritized list of AMS capabilities that stakeholders desire to obtain useful information on CAV driving behaviors and interactions. The modeling capabilities listed in this section were identified through both a literature review and panelist feedback.

Our review and stakeholder engagement primarily focused on the operational impacts of CAVs;

however, CAVs and their data will have applications to other areas of research and practice. For example, experts expect changes to the supply as well as demand for various travel options with the deployment of CAVs (Mahmassani et al. 2018). Travel forecasting models currently rely on tools like household surveys for their underlying data; CAV data may allow better understanding of household travel, as well as of long-distance and commercial trips (Walker et al. 2015). More broadly, and over a longer time scale, CAV deployment may have socioeconomic impacts, or result in changes to areas like public health or land use (Smith et al. 2018).

While each model identified in this list could provide useful results for practitioners on its own, the synthesis of multiple capabilities may lead to additional insights that would not be available from a single model. For example, safety performance (e.g., number of crashes) might be derived from a combination of models for vehicle behavior and system reliability.

STAKEHOLDER FEEDBACK

Previous FHWA projects have engaged with stakeholders on similar questions. For example, a meeting in February 2018 had stakeholders rank potential CAV applications by their interest and need for modeling tools about those applications. This meeting found that those stakeholders were particularly interested in vehicle-to-vehicle (V2V) merging, vehicle-to-infrastructure (V2I) speed harmonization, and CACC on both light-duty and heavy vehicles on freeways; and eco- CACC and eco-traffic signal controls on arterials (Leidos 2018a). However, some outreach efforts were scoped specifically for technologies at Levels 1 and 2 (SAE International 2018), and so may not consider the range of scenarios or desired capabilities that the feedback collected for this report encompasses.

Those projects have also previously identified modeling capabilities that might be useful to practitioners. For example, (Leidos 2018b) organized some of these projects by application.

Related to CACC applications, for example, they identified:

• Car-following behavior of CACC vehicles.

• Models of the formation, operation, and splitting of CACC strings.

• Models of the transition between manual and CACC.

• Lane-changing behavior (and desired behavior) for CACC vehicle drivers.

• Sensor and communication system performance.

In addition to work within the United States, the European Union-United States-Japan Trilateral Impact Assessment Subgroup has also sought international practitioner input on desired information about CAV impacts through a survey on recommended key performance indicators (KPIs) for connected and automated driving (Innamaa and Kuisma 2018). For example, the KPIs for vehicle operations that generated the most interest from survey respondents included number of human takeovers, and the duration of transition between manual and automated modes, and for network efficiency, those surveyed indicated an interest in throughput, capacity, and peak period travel time.

To determine what AMS capabilities that practitioners would find useful in managing the transition to CAV technologies, we conducted two listening sessions with stakeholders to determine their modeling needs, and to seek their input on the data necessary to develop the tools that would meet those needs. These sessions were held in late November and early December 2019. Participants were invited based on their previous participation in CAV AMS or Work Zone Driver Model stakeholder meetings.

During the listening sessions, participants were presented with draft lists of modeling capabilities, data needs, and operational scenarios, and asked to provide feedback about additional items to include in each list, as well as to discuss whether certain fields were more important than others. Forty-six unique attendees either were present at the sessions or provided written comments, listed in the Appendix. Attendees were from a wide variety of backgrounds, including academic institutions, state and local DOTs, metropolitan planning organizations (MPOs), consultancies, and software development firms.

On the topic of modeling needs, one common thread among participants was an expressed desire to collect data on CAVs beyond the direct network performance impacts of CAV behavior and human-CAV interactions, extending into the realms of safety, travel demand, and land use.

While operational data may not directly provide information about these topics, some connections may be possible between them. For example, on the topic of safety, the number of conflicts output from a microsimulation may be related to the number of crashes that occur (Dijkstra et al. 2010).

With respect to travel demand, participants expressed a desire for information about the potential shifts in value-of-travel-time associated with CAVs (for example, collected by stated-preference surveys). As noted by one stakeholder, if trip generation is incorrect, simulation will not provide accurate predictions no matter how well driving behaviors are modeled. The introduction of CAVs into traffic also introduces behaviors that are either new or will become more common – for example, passenger pick-up/drop-off. Some participants expressed interest in integrating microsimulation with travel demand models, leading to specific behavior attributes associated with individual agents based on demographic data – either driver behaviors associated with individualized agent characteristics, or CAVs which adapt their behavior based on user input (e.g., smoother acceleration for higher comfort).

Human-CAV interactions at intersections, as well as data collection in highly urban environments and capturing interactions with vulnerable road users, were other common topics

DRAFT COPY

FOR REVIEW.

of discussion as a topic area where more research would be useful, although the data collection effort at hand will likely focus on freeway applications.

The participants also provided feedback on which scenarios should be prioritized for data collection and/or model development. Several participants cautioned that any scenarios tested will have to fall within the operational design domain (ODD) of the CACC system. They also noted that many conditions listed cannot be described in simple “present” or “not present” terms;

for example, current CAV technologies may be able to perform well in a light rain but not in a downpour.

Participants also wanted to ensure that the driving conditions and scenarios during which testing is conduct are structured and classified in a well-defined and specific way, since definitions of terms like “congested conditions” vary. They stressed that the underlying data used to establish scenario classification should be included so that other data users can make their own determinations under their own definitions. However, despite these cautions, they discussed many scenarios where CAV data collection and/or models of CAV impacts would be useful.

DESIRED MODELING CAPABILITIES AND SCENARIOS

Each modeling capability will have a baseline set of assumptions that guide the initial development of a model. For example, a baseline scenario for a car-following model might be a straight, uncongested single-lane freeway with clear weather. Ideally, each model could accurately represent driving behavior or interactions under differing weather conditions, network geometries, market penetration rates of various technologies, traffic states, the presence of events like work zones, and combinations of these. However, a model’s performance under these scenarios would need to be calibrated and validated based on data collection under those conditions, where possible. A list of such scenarios is also included.

List of Desired Capabilities

The literature and stakeholders show interest in determining the impacts on traffic flow of the following behaviors:

• Car-following behavior of CAVs, including the differences between types of CAV technologies (with/without connectivity, different manufacturers, etc.).

• CAV interactions with connected infrastructure and other connected vehicles, including platoon formation/splitting and interactions with non-platooning vehicles.

• Gap acceptance of CAVs at intersections, including varying intersection control types and whether or not the intersection has V2I equipment (e.g., signal phase and timing broadcasts or signal priority).

• Transitions between manual and CAV driving (driver takeover).

• Lane-changing behavior of CAVs, both mandatory and discretionary.

• CAV behavior when interacting with other road users, including manual drivers, bicycles, and pedestrians.

DRAFT COPY

FOR REVIEW.

• Manual drivers’ behavior (lane-changing, car-following, etc.) when interacting with CAVs: do CAVs trigger changes in surrounding driver behavior, and are assumptions in existing driver models still valid in mixed traffic?

In addition to these microsimulation capabilities, participants also noted that data collection could advance the study of kinematic or shock waves at a macroscopic level.

List of Desired Scenarios

Stakeholders expressed interest in having models valid for (and, implicitly, data collected in) varying conditions of the following types:

• Levels of congestion.

• Road weather (visibility, precipitation, etc.).

• Road geometries (e.g., interchanges, freeway merge/diverge/weaves, lane drops).

• Work zone presence (with or without lane drop, taper zones, advanced warning zones).

• Vehicle mix (CAV penetration rate, heavy- vs. light-duty vehicles).

• Daytime vs. nighttime driving.

• Regions of the country (to capture regional differences in driving behavior).

• Roadway types (freeway, arterial).

Each capability and scenario identified in this chapter requires data collection to develop these models. The following chapter investigates data collection methods, best practices, and the specific data fields that advance these capabilities.

CHAPTER 4. DATA REQUIREMENTS FOR DESIRED CAPABILITIES

Given the AMS capabilities listed in the previous chapter, this chapter outlines the attributes, model inputs, and data elements necessary to build these capabilities.

QUALITY AND QUANTITY

Defining data quality requires knowledge of the data’s intended purpose. Data collectors can measure the quality of their data, and report on it in metadata when transferring or releasing data (Turner 2004). In developing descriptive measures of traffic data, Turner (2004) reviews data quality metrics from other disciplines, and finds six “fundamental” measures that apply universally:

• Accuracy: freedom from error.

• Completeness: whether data is missing and how to handle missing data.

• Validity: whether data passes quality control checks.

• Timeliness: how long ago the data was collected.

• Coverage: how much data is collected and in which situations.

• Accessibility (usability): whether researchers can use the data, and the ease of using it.

There are additional ways to characterize data quality, such as its reliability, level of detail, or consistency (Arentze et al. 2000). But aside from quality, there is also a question of quantity– how much data is necessary to develop a useful tool? The converse question can also exist for both quality and quantity – how much is sufficient for researchers? Data that is too granular, or just too voluminous, may lead to privacy concerns or unnecessary extra processing.

Accuracy can be defined as an assurance that the level of uncertainty in the data does not exceed a tolerance threshold (Kovvali et al. 2004). As part of the NGSIM project’s data format plan, accuracy standards were developed for trajectory, instrumented vehicle, and detector data, as well as road network information (Kovvali et al. 2004, 45–52). While certain technologies – for example, GPS precision – may have improved in the last fifteen years, these can be taken as a baseline for quality standards for certain types of data.

It is not only important to collect the data necessary to estimate or calibrate a model; if possible, the appropriate data for model validation should also be collected. However, there is no agreed-upon standard for how much data is enough to properly calibrate or validate a car-following model, although statistical tests may be able to provide some guidance on necessary quantities (Yang and Cooke 2018).

As with any data collection, any data collected to develop these capabilities should also be appropriately managed and contain properly-recorded metadata. This metadata should include specific descriptions of each data field, information about the collection process, and provide information about precision and units, as well as adhere to common standard formats when available (Smithsonian Libraries 2018). For example, the SHRP2 NDS training dataset includes

DRAFT COPY

FOR REVIEW.

a detailed description of the variables collected, routes traveled, quality control measures taken, and the format in which the data is available, with references to related metadata (Li, Hankey, and Perez 2016).

COLLECTION METHODS

Traffic data collection methods either involve measurements taken at a fixed location, or taken from the point-of-view of a “floating car” moving through the traffic stream (Hall 2001). There are a variety of stationary data collection methods: devices built into or placed on the roadway (e.g., pneumatic tubes, magnetic loops), roadside sensors (infrared, radar, cameras, etc.), as well as manual counting (Leduc 2008). Each technology has different capabilities – for example, some count axles while other count vehicles, and some may only effectively handle up to a certain traffic volume; resources like FHWA’s Traffic Monitoring Guide provide more specifics about the variety of sensors available (Federal Highway Administration 2016).

There is less variety in collecting floating car data; typically, an instrumented research vehicle (IRV) collects information as it travels through traffic – although the specific instrumentation varies depending on the amount of information required for the desired use of the data. For example, if one wanted to collect data about the IRV’s own following behavior, that could be collected via a front radar, but if instead the researcher desired information about the vehicles following the IRV, a rear radar would be used instead. As unmanned aerial vehicles (UAVs) become more advanced, researchers have begun exploring how they can be used to collect traffic data; however, their use has not become commonplace by practitioners (Kanistras et al. 2015).

CAVs offer new opportunities for data collection via onboard communications technologies. If CAVs communicate with roadside infrastructure, floating-car information may be collectible by these roadside units. Already, some researchers have begun to use information transmitted by drivers’ cellphones to gain some insights into traffic behavior; however, both accuracy and user privacy are concerns with these technologies (Horn et al. 2014).

Given that CAVs are not currently in wide deployment, it would not be feasible to conduct naturalistic studies on CAVs themselves, which may limit how the data can be interpreted.

However, of as much interest as the CAVs’ behaviors are the interactions that human drivers have with CAVs, and, if a CAV collects data on a public road, then data about those manually driven vehicles’ interactions are naturalistic.

STAKEHOLDER FEEDBACK

In addition to providing feedback about the modeling capabilities and scenarios that they would like to see developed, participants also shared comments about data collection methods, considerations to take into account when deciding how to collect data, and how the methods used affect the capabilities that are developed with that data. For example, participants identified a risk that the current state of maturity of CAV technologies may not represent with full accuracy how manual drivers would interact with a CAV that has matured enough to become commercially available. Relatedly, if each manufacturer develops their own control algorithms, then “CAV behavior”, in the general sense, may not exist – each manufacturer’s models may

DRAFT COPY

FOR REVIEW.

behave differently – commenters pointed to commercial ACC systems as an example. More broadly, CAV maturity may lead to new vehicle designs or concepts that do not currently exist, and could change how road space is used. Overall, the uncertainty associated with CAV technologies is a major limiting factor to collecting useful data on their potential impacts.

Another concern raised was the sentiment that if CAVs have no obvious physical differences (such as a mounted LiDAR unit) from any other vehicle on the road, then a driver will have no way to distinguish it from a human-operated vehicle, and will interact with it just the same as they would with any other. On the other hand, drivers may interact with instrumented vehicles differently than they would other vehicles – a concern for data collection, although many data fields can now be collected via CAN bus or other less visible means (Fugiglando et al. 2019).

However, one stakeholder indicated that he already notices cars equipped with commercial ACC technologies, because of the large headways that these systems leave, and treats them differently while driving.

Participants also provided feedback about the level of detail at which data should be collected, overall taking a “the more data, the better” approach – even suggesting, for example, that all data collected by the CAN bus should be recorded in case they end up being useful at a later date, even if there is no immediate plan to use or process that data. Their experience with data collection also showed that GPS data were not accurate or precise enough to track lane-changes or acceleration in a useful manner for microsimulation.

Several stakeholders felt that NGSIM-style trajectories of many vehicles over a stretch of road would be useful in developing modeling capabilities, and noted that unmanned aerial vehicles (UAVs) provide a new opportunity for collecting this type of data, which didn’t exist at the time of NGSIM.

DATA REQUIREMENTS

This section identifies specific data fields that are necessary or desirable for the development of the capabilities identified in the previous chapter. Many of these data fields are standard measures from traffic flow theory and the microsimulation literature, useful in any operational modeling context – not just for CAVs (Hall 2001; Toledo 2007). Others are more specific to automation and/or connectivity, or for certain driving scenarios.

It makes sense to divide these data fields by the nature of data collection. First, there is microscopic data that is best collected by vehicle instrumentation, continuously during driving – position, speed, and acceleration, as well as information about car-following and lane-changing, sensor/communications reliability, and so on. These can either be directly collected, or calculated based off of instrumented vehicle sensors. In addition, information should be collected about each “run” during which the vehicle collects data. For example, this might include information about the operating conditions (for example, road weather), or, for data collection involving manual driving behavior, demographic information about the vehicle’s driver.

In contrast, data that is generally macroscopic, describing the traffic flow and link status, would be better collected by other means. These data may be historical or aggregated (e.g., average

DRAFT COPY

FOR REVIEW.

traffic characteristics), collected by detectors, roadside instruments, or estimated from data collected from multiple vehicle (for example, travel speeds reported by traffic apps). For both categories, there should also be general descriptions of the data collection effort; such as vehicle equipment and the route to be traveled during collection.

In addition to the stakeholder sessions, sources for the data fields include:

• Literature on traffic flow theory (Hall 2001).

• Other data collection efforts, such as NGSIM and SHRP2 NDS (Hranac, Margiotta, and

Alexiadis 2004; Kovvali et al. 2004; Hankey, Perez, and McClafferty 2016).

• Reports from recent FHWA projects (Mahmassani et al. 2018; Birriel et al. 2020; Leidos

2018b).

Instrumented Vehicle Data

Instrumented vehicles can collect a variety of data fields, but information about the vehicle used to collect those data should also be recorded. Table 1 lists some of the information that should be reported about each data collection vehicle, although some of this information may be easier to record in a narrative format rather than in a table or data dictionary.

Table 1. Vehicle characteristics data.

Data Field Description

Make / model / year Vehicle identifying information AV technologies equipped A description of the types of automated technologies being used on the vehicle and their ODD/maturity CV technologies equipped A description on the connected technologies used on the vehicle Sensor detection / communications range

How far ahead or behind does the vehicle have information about? Is this information just recorded (as part of research instrumentation), or used by CAV technologies?

Physical appearance Does the vehicle look different from commercially available technologies? An instrumented vehicle (even if human-driven) that looks very different may provoke different responses from surrounding drivers

System error logs Any errors recorded by the CAV systems or the IRV systems while collecting data.

In addition to information about the vehicle used to collect data, data about the roadway where data is being collected should also be recorded, since roadway condition factors affect driver behavior. The data in Table 2 describe information about the roadway that should be collected once per route that data is collected on. In addition to Table 2’s information (which is relatively static), there are also roadway conditions which would vary from run-to-run, and thus would need to be collected during each IRV run to capture their effect on vehicle/driver behavior (for example, lighting conditions or road weather). These are recorded in Table 3.

Table 2. Roadway characteristics data Data Field Units or classification, if applicable

Description

Location (long/lat) Latitude-longitude

Identifier for each point along the roadway where data were recorded.

Location (roadway) Road name, city, state

Name of the current street that the subject vehicle is traveling on. For roads that cross city or state lines, also note the current city/town and state, and mile marker information, if available.

Grade % (For vertical curves) What is the current grade?

Turning radius Degrees (For horizontal curves) What is the current radius of the curve?

Lane width Feet Width of the current lane Lane markings* -- Information regarding the type, width, condition of lane markings on the road (Federal Highway Administration 2019, chap. 3).

Number of lanes -- Number of travel lanes present in the current direction Travel lane -- (When there are multiple travel lanes in the same direction) A designation for the current travel lane of the subject vehicle.

Roadway type Functional classification

Type of road being driven on: collector, arterial, freeway, on-ramp, off-ramp, etc. (Federal Highway Administration 2013).

Work zone info AW, TZ, WZ1, WZ2

Note if the subject vehicle is currently traversing a freeway work zone or a segment of freeway where travel behaviors might be influenced by the presence of a work zone. This includes: the advanced warning area (AW), taper zone or transition area (TZ), a work zone with a lane drop (WZ1), a work zone without lane drop (WZ2) (Federal Highway Administration 2019, figs.

6C-1).

Rural/urban* R, U or population density

Calculated from geographic information.

Recommended guidelines for "urban" (U) vs. "rural"

(R) roadways are available from (Federal Highway Administration 2013, sec. 3).

Alternatively, this could be reported as population density so researchers can designate rural vs. urban using their own parameters.

Speed limit mph Posted speed limit at the given location

Interchange density Interchanges per mile

The number of interchanges occurring along the road, normalized by the road length

-- = not applicable.

DRAFT COPY

FOR REVIEW.

* = more than one field may need to be recorded

Table 3. Roadway conditions data classification, if applicable

Description

Road weather WR, WNR, F, C

Wet with rain (WR), wet / no rain (WNR), foggy (F), clear (C)

Congestion depends on calculation method

Similar to rural/urban, there is no universal definition, so the calculation method should be stated (e.g., traffic speed < 25 mph). Indication of whether the congestion is recurring or non-recurring may also be useful. Some of the following fields may be used to characterize congestion.

Travel time index -- Ratio of the travel time for the run and the free-flow travel time along the same segment of road. Calculated over a road segment.

Speed-to-free-flow speed ratio

-- Similar to travel time index, but for speed instead of travel time. Could also be calculated with instantaneous speed as well as speed over a segment.

Local density Vehicles per mile

The density within a (constantly-moving) region of the road a specified distance up-/downstream of the vehicle, collected continuously

Time of data collection hh:mm:ss.00 When the data were collected. Can be used to determine whether conditions were peak hour or not, and could be used as a proxy for lighting conditions if unable to collect directly.

Lighting conditions lux Presence of ambient light Law enforcement officer presence

Yes/no Whether a law enforcement officer is present / visible on the roadway

Vehicle mix % Shares of light-duty vs. heavy duty, CAV vs. human-driven vehicles on the roadway

Volume-to-capacity ratio**

-- Ratio of volume to estimated capacity. See Table 8 for macroscopic data fields including volume.

Demand-to-capacity ratio**

-- Ratio of estimated demand to estimated capacity. See Table 8 for macroscopic data fields including volume.

** = limitations exist due to estimation required

Table 4 shows data fields that are useful in building, calibrating and validating car-following models. While some of these data are directly collected by the IRV, others must be defined and calculated in post-processing – for example, determining whether the subject vehicle is following another may be calculated via Δ𝑥𝑥/Δ𝑣𝑣 regimes. As discussed above, the definitions, calculation method, and underlying data used in post-processing should be clear when publishing processed

DRAFT COPY

FOR REVIEW.

data or resulting analysis. Listening session participants agreed that 10 Hz resolution for IRV data is standard, and that onboard accelerometers are preferable to GPS for data precision.

Table 4. Car-following data classification Description Direct collection

(D) or calculated in post-processing (P)?

Position Latitude-longitude Location of the vehicle on the roadway.

D

Speed mph Speed of the subject vehicle D Acceleration m/s2 Acceleration of the subject vehicle D In car following?

Yes/no Is subject vehicle following another? P

Distance to leader

Meters Distance between the subject vehicle and the vehicle ahead of it (if vehicle sensed ahead)

D

Leader speed mph Speed of the vehicle immediately ahead of the IRV

D

Leader acceleration m/s2 Acceleration of the vehicle immediately ahead of the IRV

D

Being followed?

Yes/no Is subject vehicle being followed? P

Distance to follower m Distance between the subject vehicle and its follower (if vehicle sensed behind)

D

Follower speed mph Speed of the vehicle immediately behind the subject vehicle

D

Follower acceleration m/s2 Acceleration of the vehicle immediately behind the subject vehicle

D

Table 5 shows data fields that are specific to data collection about CAV platooning. Since collecting such data would require many instrumented CAVs (whereas other efforts might only require a single IRV to collect information), data for this modeling capability would likely be more expensive to collect than on other capabilities.

Table 5. Platooning data classification Description

In platoon? Yes/no Is the subject vehicle in a string of CAVs following each other?

Platoon length Feet Length of the platoon in terms of distance, from front bumper of lead vehicle to rear bumper

DRAFT COPY

FOR REVIEW.

classification

Description

Platoon size Vehicles The number of vehicles in the platoon Subject vehicle position

1, 2, … , n The position of the subject vehicle in the platoon (1=leader)

String leader speed mph Speed of the first vehicle in the platoon (from that vehicle’s instrumentation)

String leader acceleration m/s2 Acceleration of the first vehicle in the platoon (from that vehicle’s instrumentation)

Table 6 shows data that would be useful in building, calibrating and validating lane-changing models. As noted for car-following data, some of this information is difficult to collect directly and must be derived from post-processing or observation. Of particular difficulty is determining when a vehicle or driver’s desire to change lanes begins, since a change may be desired before any lateral movement occurs.

Table 6. Lane-changing data classification Description Direct collection

(D) or calculated in post-processing (P)?

Is a lane change desired?

Changing right / changing left / no

Whether the subject vehicle has determined that a lane change would be necessary or desirable. Note that this is generally prior to the vehicle physically starting to change lanes.

P

MLC or DLC? MLC / DLC Is the lane change mandatory (MLC) or discretionary (DLC)?

P

Distance to MLC event

Meters When in MLC, how far does the vehicle have until the mandatory event (e.g., lane drop)?

D

Time elapsed in

MLC/DLC

Seconds Time from start of MLC/DLC to completion of maneuver (recorded per lane-change instance).

P

Current gap in adjacent [right / left] lane

Meters The distance between the nearest two vehicles in the lane to the [right / left] of the subject vehicle

D

Gap accepted Meters The gap between the two vehicles in the desired lane that the subject vehicle entered between (recorded per lane-change instance)

D

Speed of vehicles in adjacent lane to the [right / left] mph Useful not only if the subject vehicle is changing lanes, but if the adjacent vehicle is attempting to merge into the subject vehicle’s lane

D

DRAFT COPY

FOR REVIEW.

classification

Description Direct collection

(D) or calculated in post-processing (P)?

Courtesy yield / forced merge required to change lanes

Yes/no Whether vehicles in adjacent lane slowed down to allow subject vehicle to change lanes

P

Vehicle in adjacent lane attempting to merge?

Yes/no This is difficult to observe until the lateral motion starts, but could be indicated by turn signal activation (see below).

P

Subject vehicle yielded / was cut off?

Yes/no Whether subject vehicle slowed down to allow a vehicle from another lane to merge into subject lane

P

Blinker activation

Yes/no Whether the turn signal was engaged --for both subject vehicle and the surrounding vehicles.

D

In determining the effects of CAV behavior on traffic, collecting manual drivers’ responses to those behaviors are as important as collecting the behavior itself – perhaps more important, if CAV market penetration is low. Table 7 describes various aspects of driver behavior that could be captured either for a non-CAV instrumented vehicle, or potentially information about surrounding drivers that could be captured by an instrumented CAV.

Table 7. Driver data Data field Description

Steering wheel inputs

Angle/momentum of steering wheel (collected continuously)

Gas pedal inputs How much/if the gas pedal is depressed (collected continuously) Brake pedal inputs How much/if the brake pedal is depressed, for IRVs, or for surrounding vehicles, if the brake light is engaged (collected continuously).

Driver state detection*

Cameras for eye-tracking and sensors that determine if a driver’s hands are on the steering wheel could be used to determine whether or not a driver is distracted or paying attention. (collected continuously)

Age How old the driver is, how experienced of a driver they are (collected once per driver)

Gender Male/female (collected once per driver)

* = more than one field may need to be recorded

Non-IRV Data

Some traffic data – particularly macroscopic information – are better collected by roadside instrumentation or other methods aside from an instrumented vehicle. Some fields may also be available as historical data from a source such as a state DOT. Table 8 lists some of these fields

DRAFT COPY

FOR REVIEW.

not easily collectible by an IRV. In addition, the types of roadway data present in Table 2 and Table 3 may provide context for such roadside data; however, the level of aggregation may be different.