Development of tools to model driver behavior in a cooperative and driverless vehicle mixed roadway environment

GEORGIA DOT RESEARCH PROJECT 18-23 Final Report
DEVELOPMENT OF TOOLS TO MODEL DRIVER BEHAVIOR IN A COOPERATIVE
AND DRIVERLESS VEHICLE MIXED ROADWAY ENVIRONMENT
Office of Performance-based Management and Research
600 West Peachtree Street NW | Atlanta, GA 30308 January 2022

TECHNICAL REPORT DOCUMENTATION PAGE

1. Report No.:

2. Government Accession No.:

FHWA-GA-22-1823

N/A

4. Title and Subtitle:

Development of Tools to Model Driver Behavior in a Cooperative

and Driverless Vehicle Mixed Roadway Environment

7. Author(s): Michael Hunter (PI), Ph.D. (https://orcid.org/ 0000-0002-0307-
9127); Angshuman Guin (coPI), Ph.D. (https://orcid.org/ 0000-0001-
6949-5126); Abhilasha Saroj, Ph.D.; Jong In Bae
9. Performing Organization Name and Address: Georgia Tech Research Corp. 505 Tenth St. Atlanta, GA 30318

3. Recipient's Catalog No.: N/A 5. Report Date: January 2022 6. Performing Organization Code: N/A 8. Performing Organization Report No.: 18-23
10. Work Unit No.: N/A 11. Contract or Grant No.: PI#0015718

12. Sponsoring Agency Name and Address: Georgia Department of Transportation (SPR) Office of Performance-based Management and Research 600 West Peachtree St. NW Atlanta, GA 30308

13. Type of Report and Period Covered: Final; March 2019 January 2022
14. Sponsoring Agency Code: N/A

15. Supplementary Notes: Prepared in cooperation with the U.S. Department of Transportation, Federal Highway Administration.

16. Abstract: Promising advances in autonomous vehicle (AV) technology have fueled industry and research fields to dedicate

significant efforts to the study of the integration of AVs into the traffic network. While most studies anticipate a

beneficial role of AVs, contributing to improved traffic efficiency and roadway safety, the underlying assumptions

on the interactions between AVs and human-driven vehicles (HDVs) are often cooperative in nature. The first

portion of this study investigates the impact of aggressive human-driven vehicles' (AHDVs) merging behaviors on

traffic performance in a mixed environment that includes three vehicle types: AVs, HDVs, and AHDVs. This study

is undertaken in an open-source microscopic traffic simulation model, Simulation of Urban Mobility (SUMO).

AHDVs have been modeled in this study to show aggressive merging behaviors at a merge section of a freeway exit

ramp by targeting the farthest reachable AV for lane change as well as forcing a merge immediately in front of the

AV. Results show that the travel-time gains achieved by AHDVs were at the expense of AVs and HDVs, and the

interaction of aggressive HDVs with cooperative AVs could negatively impact overall capacity. The second portion

of this study developed an Excel-based tool exploring the impact of AVs on departure capacity from a signalized

intersection. Through both portions of this study, it was seen that critical indicators of the impact of AVs on traffic

performance are: (1) Is a rise in aggressive interactions witnessed? (2) What are the headways being adopted by

AVs? And (3) What are the spacing and maximum-length platooning characteristics?

17. Keywords:

18. Distribution Statement:

Autonomous Vehicles, Mixed Traffic Simulation, No Restriction

Aggressive Behavior, Queue-Jump

19. Security Classification 20. Security Classification (of this

(of this report):

page):

Unclassified

Unclassified

21. No. of Pages: 22. Price:

167

Free

Form DOT 1700.7 (8-72)

Reproduction of completed page authorized

GDOT Research Project 18-23 Final Report
DEVELOPMENT OF TOOLS TO MODEL DRIVER BEHAVIOR IN A COOPERATIVE AND DRIVERLESS VEHICLE MIXED ROADWAY
ENVIRONMENT By
Michael Hunter, Ph.D. Principal Investigator Angshuman Guin, Ph.D. co-Principal Investigator Abhilasha Saroj, Ph.D.
Researcher Jong In Bae Graduate Research Assistant Georgia Tech Research Corporation
Contract with Georgia Department of Transportation
In cooperation with U.S. Department of Transportation Federal Highway Administration
January 2022
The contents of this report reflect the views of the authors, who are responsible for the facts and accuracy of the data presented herein. The contents do not necessarily reflect the official views or policies of the Georgia Department of Transportation or the Federal Highway Administration. This report does not constitute a standard, specification, or regulation.
ii

Symbol
in ft yd mi
in2 ft2 yd2 ac mi2
fl oz gal ft3 yd3
oz lb T
oF
fc fl
lbf lbf/in2

SI* (MODERN METRIC) CONVERSION FACTORS

APPROXIMATE CONVERSIONS TO SI UNITS

When You Know

Multiply By

To Find

LENGTH

inches

25.4

millimeters

feet

0.305

meters

yards

0.914

meters

miles

1.61

kilometers

AREA

square inches

645.2

square millimeters

square feet

0.093

square meters

square yard

0.836

square meters

acres

0.405

hectares

square miles

2.59

square kilometers

VOLUME

fluid ounces

29.57

milliliters

gallons

3.785

liters

cubic feet

0.028

cubic meters

cubic yards

0.765

cubic meters

NOTE: volumes greater than 1000 L shall be shown in m3

MASS

ounces

28.35

grams

pounds

0.454

kilograms

short tons (2000 lb)

0.907

megagrams (or "metric ton")

TEMPERATURE (exact degrees)

Fahrenheit

5 (F-32)/9

Celsius

or (F-32)/1.8

ILLUMINATION

foot-candles foot-Lamberts

10.76 3.426

lux candela/m2

FORCE and PRESSURE or STRESS

poundforce

4.45

newtons

poundforce per square inch

6.89

kilopascals

Symbol
mm m m km
mm2 m2 m2 ha km2
mL L m3 m3
g kg Mg (or "t")
oC
lx cd/m2
N kPa

Symbol
mm m m km
mm2 m2 m2 ha km2
mL L m3 m3
g kg Mg (or "t")
oC
lx cd/m2
N kPa

APPROXIMATE CONVERSIONS FROM SI UNITS

When You Know

Multiply By

To Find

LENGTH

millimeters

0.039

inches

meters

3.28

feet

meters

1.09

yards

kilometers

0.621

miles

AREA

square millimeters

0.0016

square inches

square meters

10.764

square feet

square meters

1.195

square yards

hectares

2.47

acres

square kilometers

0.386

square miles

VOLUME

milliliters

0.034

fluid ounces

liters

0.264

gallons

cubic meters

35.314

cubic feet

cubic meters

1.307

cubic yards

MASS

grams

0.035

ounces

kilograms

2.202

pounds

megagrams (or "metric ton")

1.103

short tons (2000 lb)

TEMPERATURE (exact degrees)

Celsius

1.8C+32

Fahrenheit

ILLUMINATION

lux candela/m2

0.0929 0.2919

foot-candles foot-Lamberts

FORCE and PRESSURE or STRESS

newtons

0.225

poundforce

kilopascals

0.145

poundforce per square inch

Symbol
in ft yd mi
in2 ft2 yd2 ac mi2
fl oz gal ft3 yd3
oz lb T
oF
fc fl
lbf lbf/in2

**SSI iIsitshethseysmybmolbfoolr fthoer tIhnteerInnatetironnaatlioSnysatleSmysotfeUmniotsf. UAnpiptsro. pAripapteroropurinadtiengrosuhnoduilndgbeshmoaudlde btoecmomadpelytwoitchoSmepcltyionw4ithofSAeScTtiMonE438o0f. ASTM E380. ((RReevvisiseeddMMaracrhch2020030)3)

iii

TABLE OF CONTENTS
EXECUTIVE SUMMARY .............................................................................................. 1 OVERVIEW ................................................................................................................. 1 Report Organization................................................................................................ 4 Recommendations.................................................................................................... 4
CHAPTER 1. INTRODUCTION .................................................................................... 6
CHAPTER 2. BACKGROUND....................................................................................... 9 UNCERTAINTY IN INTERACTIONS BETWEEN AUTONOMOUS VEHICLES AND HUMAN ROAD USERS .......................................................... 9 PREVIOUS RESEARCH ON MODELING AUTONOMOUS VEHICLE BEHAVIORS.......................................................................................................... 10 AVHUMAN INTERACTION MODELING ......................................................... 18 AGGRESSIVE BEHAVIOR MODELING APPROACH ..................................... 22
CHAPTER 3. SELECTION OF SIMULATION MODEL APPROACH ................. 23 INTRODUCTION...................................................................................................... 23 KEY EVALUATION CRITERIA ............................................................................ 23 General Information.............................................................................................. 23 Availability of Driving Models.............................................................................. 23 Driver Behavior Model Parameters ..................................................................... 24 Data Export ............................................................................................................ 24 User Contribution .................................................................................................. 25 Signal Control System ........................................................................................... 25 Overall Comments ................................................................................................. 25
CHAPTER 4. SIMULATION MODELING OF DRIVER BEHAVIOR .................. 28 AGGRESSIVE MERGE BEHAVIOR MODEL DEVELOPMENT .................... 28 Simulation Tool ...................................................................................................... 28 Network Layout ..................................................................................................... 28 Vehicle Classification and Characteristics .......................................................... 29 Aggressive Behavior Model................................................................................... 30 Aggressive Merge Behavior Model: Target Selection Process........................... 31 Aggressive Merge with Maximum Advancement .............................................. 31 Aggressive Merge with Zipper .......................................................................... 32 Position Check .............................................................................................. 33 Can Catch ...................................................................................................... 34 Merge Position .............................................................................................. 36 Target AV Selection Process in Aggressive Merge with Maximum Advancement Case .................................................................................................................. 37
iv

Target Selection Process in Aggressive Merge with Zipper Case.................... 38 Aggressive Merge Behavior Model: Lane Changing ......................................... 44
Merging Process in SUMO-controlled Merge.................................................. 44
EXPERIMENTS ........................................................................................................ 45
Experiment 1: Aggressive Merging with Platooned Arrivals ............................ 46 Objectives.............................................................................................................. 46 Experiment 1 Design............................................................................................ 46 Results................................................................................................................... 47 Discussion............................................................................................................. 47 Low Congestion ................................................................................................ 49 High Congestion ............................................................................................... 52
Experiment 2: Aggressive Merging with Random Arrivals .............................. 55 Objective ............................................................................................................... 55 Experiment Design............................................................................................... 55 Results................................................................................................................... 58 Discussion............................................................................................................. 64
Experiment 3: Comparison of Impact of Demand versus Congestion on Travel Times ...................................................................................................... 66 Objective ............................................................................................................... 66 Results................................................................................................................... 66
Experiment 4: Evaluation of Impact of Aggressive Merging on Capacity....... 75 Experiment Design............................................................................................... 75 Vehicle Classification .......................................................................................... 76 Vehicle Assignment.............................................................................................. 77 Results Visualization............................................................................................ 78 Discussion............................................................................................................. 78 Summary............................................................................................................... 92
CHAPTER 5. DATA COLLECTION FOR DRIVER BEHAVIOR ......................... 94
DATA COLLECTION PLAN .................................................................................. 94
DATA COLLECTION METHOD ........................................................................... 94
Site Selection for Drone Video Data Collection .................................................. 94 Drone Video Data Processing Using the DataFromSky Viewer........................ 95
Manual Georeferencing....................................................................................... 97 Manual Annotation Configuration ..................................................................... 98 Exporting Analysis Data ...................................................................................... 99
OBSERVATIONS/RESULTS................................................................................. 100
CHAPTER 6. SIMPLIFIED CAPACITY ANALYSIS TOOL ................................ 101
INTRODUCTION.................................................................................................... 101
CAV SATURATION FLOW OVERVIEW .......................................................... 101
SCAT SATURATION FLOW MODELS.............................................................. 104
v

Capacity Adjustment Factors for Connected and Automated Vehicles in the Highway Capacity Manual, Draft Phase 1 Report, Pooled Fund Study (Schroeder et al. 2021) .................................................................................... 104
"Modeling Impacts of Cooperative Adaptive Cruise Control on Mixed Traffic Flow in Multi-lane Freeway Facilities" (Liu et al. Error! Reference source not found.).......................................................................... 106
"Autonomous and Connected Cars: HCM Estimates for Freeways with Various Market Penetration Rates" (Shi and Prevedouros 2016).............. 107
"Enhanced Intelligent Driver Model to Access the Impact of Driving Strategies on Traffic Capacity" (Kesting et al. 2010) .................................. 108
"A Mixed Traffic Capacity Analysis and Lane Management Model for Connected and Automated Vehicles: A Markov Chain Method" (Ghiasi et al. 2017) ........................................................................................................ 109
VISSIM Simulation ............................................................................................. 109 INSTRUCTIONS FOR SIMPLIFIED CAPACITY ANALYSIS TOOL ........... 111
Individual Scenario Analysis .............................................................................. 114 Scenario Comparative Analysis.......................................................................... 118 Phase Comparative Analysis............................................................................... 122 SUMMARY .............................................................................................................. 124
CHAPTER 7. CONCLUSIONS AND RECOMMENDATIONS ............................. 126 RECOMMENDATIONS ......................................................................................... 129
APPENDIX A: Observed Headways........................................................................... 131 PIB AT BERKELEY LAKE (33.985340, -84.171123) ......................................... 131 Through Movement Headways .......................................................................... 131 Left-turn Movement Headways.......................................................................... 134 PIB AT MEDLOCK BRIDGE ROAD (33.961047, -84.208518) ......................... 137 Through Movement Headways .......................................................................... 137 Left-turn Movement Headways.......................................................................... 140
APPENDIX B: SPEED FLOW PLOTS FOR EXPERIMENT 4A AND 4B ........ 142
REFERENCES.............................................................................................................. 149
vi

LIST OF FIGURES
Figure 1. Diagram. Roadway layout on merging zone, highlighted in yellow. ................ 29 Figure 2. Diagram. Speed by lane..................................................................................... 29 Figure 3. Diagram. Aggressive merge behavior of AHDVs toward AVs. ....................... 31 Figure 4. Diagram. Aggressive merge with zipper. .......................................................... 33 Figure 5. Diagram. Illustration of AHDV's decision point. ............................................. 34 Figure 6. Diagram. AHDV travel-time calculation. ......................................................... 35 Figure 7. Diagram. Illustration of merge position. ........................................................... 36 Figure 8. Flowchart. Target AV selection process with step numbers in the process
description marked (top). ............................................................................................ 40 Figure 9. Flowchart. Target AV selection process with step numbers in the process
description marked (bottom)....................................................................................... 41 Figure 10. Flowchart. Target selection process in aggressive merge with zipper (top). .. 42 Figure 11. Flowchart. Target selection process in aggressive merge with zipper
(bottom)....................................................................................................................... 43 Figure 12. Diagram. Example of aggressive merge.......................................................... 44 Figure 13. Diagram. Example of SUMO-controlled merge. ............................................ 45 Figure 14. Plot. Base case timespace diagram in uncongested deceleration-lane
scenario. ...................................................................................................................... 49 Figure 15. Plot. Aggressive merge with maximum advancement timespace diagram
in uncongested deceleration-lane scenario.................................................................. 50 Figure 16. Plot. Aggressive merge with zipper timespace diagram in uncongested
deceleration-lane scenario........................................................................................... 51 Figure 17. Plot. Base case timespace diagram in congested deceleration-lane
scenario. ...................................................................................................................... 52 Figure 18. Plot. Aggressive merge with maximum advancement timespace diagram
in congested deceleration-lane scenario...................................................................... 53 Figure 19. Plot. Aggressive merge with zipper timespace diagram in congested
deceleration-lane scenario........................................................................................... 54 Figure 20. Bar plots. Experiment 2: Average travel time in aggressive merge with
maximum advancement scenarios by vehicle type in low traffic-demand condition. .................................................................................................................... 58 Figure 21. Bar plots. Experiment 2: Average travel time in aggressive merge with zipper scenarios by vehicle type in low traffic-demand condition. ............................ 59 Figure 22. Bar plots. Experiment 2: Average travel time in aggressive merge with maximum advancement scenarios by vehicle type in high traffic-demand condition. .................................................................................................................... 60 Figure 23. Bar plots. Experiment 2: Average travel time in aggressive merge with zipper by vehicle type in high traffic-demand condition. ........................................... 61 Figure 24. Bar plots. Experiment 3: Average travel time in aggressive merge with maximum advancement scenarios by vehicle type in low traffic-demand condition. .................................................................................................................... 69 Figure 25. Bar plots. Experiment 3: Average travel time in aggressive merge with zipper scenarios by vehicle type in low traffic-demand condition. ............................ 70
vii

Figure 26. Bar plots. Experiment 3: Average travel time in aggressive merge with maximum advancement scenarios by vehicle type in high traffic-demand condition. .................................................................................................................... 71
Figure 27. Bar plots. Experiment 3: Average travel time in aggressive merge with zipper scenarios by vehicle type in high traffic-demand condition. ........................... 72
Figure 28. Diagram. Queue building at the end of lane B_1. ........................................... 79 Figure 29. Plot. Time vs. average speed & flow plots at 500 ft before the start of the
deceleration lane in experiment 4a: (a) 0%, (b) 25%, (c) 50%, (d) 75%, and (e) 100%...................................................................................................................... 82 Figure 30. Plot. Time vs. average speed & flow plots at the start of the deceleration lane in experiment 4a: (a) 0%, (b) 25%, (c) 50%, (d) 75%, and (e) 100%................. 83 Figure 31. Plot. Time vs. average speed & flow plots at the start of the ramp in experiment 4a: (a) 0%, (b) 25%, (c) 50%, (d) 75%, and (e) 100%............................. 84 Figure 32. Plot. Time vs. average speed vs. flow plots at 500 ft before the start of the deceleration lane in experiment 4b: (a) 0%, (b) 25%, (c) 50%, (d) 75%, and (e) 100%...................................................................................................................... 87 Figure 33. Plot. Time vs. average speed vs. flow plots at the start of the deceleration lane in experiment 4b: (a) 0%, (b) 25%, (c) 50%, (d) 75%, (e) 100%. ...................... 88 Figure 34. Plot. Time vs. average speed vs. flow plots at the start of ramp in experiment 4b: (a) 0%, (b) 25%, (c) 50%, (d) 75%, and (e) 100%. ........................... 89 Figure 35. Maps. Sites chosen for data collection: (a) at the intersection of PIB@ North Berkeley Lake Road, (b) at the intersection of PIB@ Medlock Bridge Road. ........................................................................................................................... 96 Figure 36. Map. Annotated vehicle trajectories in DataFromSky Viewer. ...................... 97 Figure 37. Map. Manual georeferencing in DataFromSky Viewer. ................................. 98 Figure 38. Map. Inserting gates at stop line...................................................................... 99 Figure 39. Screenshot. Exporting analysis data from DataFromSky Viewer. ................ 100 Figure 40. Screenshot. SCAT Individual Scenario Analysis example Data Input Section....................................................................................................................... 115 Figure 41. Screenshot. SCAT Individual Scenario Analysis example Data Phase Layout section. .......................................................................................................... 116 Figure 42. Screenshot. SCAT Individual Scenario Analysis example Analysis Table Output section. ................................................................................................ 117 Figure 43. Screenshot. SCAT Individual Scenario Analysis example Analysis Graphical Output section. ......................................................................................... 117 Figure 44. Screenshot. SCAT Scenario Comparative Analysis example Data Input section.............................................................................................................. 119 Figure 45. Screenshot. SCAT Scenario Comparative Analysis example Capacity Tables section............................................................................................................ 120 Figure 46. Screenshot. SCAT Scenario Comparative Analysis example Capacity Graphs section........................................................................................................... 121 Figure 47. Screenshot. SCAT Phase Comparative Analysis example Data Input section ....................................................................................................................... 122 Figure 48. Screenshot. SCAT Phase Comparative Analysis example Capacity Per AV Scenario section. ................................................................................................ 123
viii

Figure 49. Screenshot. SCAT Phase Comparative Analysis example Capacity Graph section. ........................................................................................................... 124
Figure 50. Map. Through movements PIB at Berkeley Lake. ..................................... 131 Figure 51. Plot. PIB at Berkeley Lake through movement headway data visualization
(top left NB Lane 1, top right NB Lane 2, bottom left SB Lane 1, bottom right SB Lane 2)..................................................................................................... 132 Figure 52. Plot. PIB at Berkeley Lake through movement headway distribution .......... 133 Figure 53. Map. Left-turn movements PIB at Berkeley Lake ..................................... 134 Figure 54. Plot. PIB at Berkeley Lake left-turn movement headway visualization ....... 135 Figure 55. Plot. PIB at Berkeley Lake left-turn movement headway distribution ......... 136 Figure 56. Map. Through movement PIB at Medlock Bridge Road ........................... 137 Figure 57. Plot. PIB at Medlock Bridge Road through movement headway data visualization .............................................................................................................. 138 Figure 58. Plot. PIB at Medlock Bridge Road through movement headway distribution ................................................................................................................ 139 Figure 59. Map. Left-turn movement PIB at Medlock Bridge Road........................... 140 Figure 60. Plot. PIB at Medlock Bridge Road left-turn movement headway data visualization. ............................................................................................................. 141 Figure 61. Plot. PIB at Medlock Bridge Road left-turn movement headway distribution ................................................................................................................ 141 Figure 62. Plot. Speed vs. flow plots at 500 ft before the start of the deceleration lane in experiment 4a (top left 0%, top right 25%, bottom left 50%, bottom middle 75%, bottom right 100%)........................................................................ 142 Figure 63. Plot. Speed vs. flow plots at the start of the deceleration lane in experiment 4a (top left 0%, top right 25%, bottom left 50%, bottom middle 75%, bottom right 100%).................................................................................... 143 Figure 64. Plot. Speed vs. flow plots at the start of the ramp in experiment 4a (top left 0%, top right 25%, bottom left 50%, bottom middle 75%, bottom right 100%). ........................................................................................................... 144 Figure 65. Plot. Speed vs. flow plots at 500 ft before the start of the deceleration lane in experiment 4b (top left 0%, top right 25%, bottom left 50%, bottom middle 75%, bottom right 100%)........................................................................ 145 Figure 66. Plot. Speed vs. flow plots at the start of the deceleration lane in experiment 4b (top left 0%, top right 25%, bottom left 50%, bottom middle 75%, bottom right 100%).................................................................................... 146 Figure 67. Plot. Speed vs. flow plots at the start of the ramp in experiment 4b (top left 0%, top right 25%, bottom left 50%, bottom middle 75%, bottom right 100%). ........................................................................................................... 147
ix

LIST OF TABLES
Table 1. Summary of AV/CAV behavior assumptions and parameters employed. ......... 12 Table 2. List of frequently and infrequently used VISSIM parameters............................ 17 Table 3. Summary of AV Human interaction modeling approaches............................. 20 Table 4. Evaluation criteria summary on simulation tools. .............................................. 27 Table 5. Vehicle assignment for experiments 2 and 3. ..................................................... 57 Table 6. Experiment 2: Paired t-test on travel time in aggressive merge with
maximum advancement scenarios. ............................................................................. 62 Table 7. Experiment 2: Paired t-test on travel time in aggressive merge with zipper
scenarios...................................................................................................................... 63 Table 8. Experiment 3: Paired t-test on travel time in aggressive merge with
maximum advancement scenarios. ............................................................................. 73 Table 9. Experiment 3: Paired t-test on travel time in aggressive merge with zipper
scenarios...................................................................................................................... 74 Table 10. Vehicle assignment on lane A_0 for experiment 4a and experiment 4b. ......... 77 Table 11. Vehicle count by vehicle type at entry point in experiment 4a ........................ 85 Table 12. Vehicle count by vehicle type at ramp in experiment 4a (vehicles / 15
minutes)....................................................................................................................... 86 Table 13. Vehicle count by vehicle type at entry point in experiment 4b ........................ 90 Table 14. Vehicle count by vehicle type at ramp in experiment 4b (vehicles / 15
minutes)....................................................................................................................... 91 Table 15. Through movement headway distribution PIB at Berkeley Lake................ 131 Table 16. Left-turn movement headway distribution PIB at Berkeley Lake. .............. 134
x

ACC AHDV AV CACC CAV CV HCM HDV IDM LMRS LOS PIB SCAT SUMO TraCI TSD TwB VAD

LIST OF ACRONYMS/ABBREVIATIONS Adaptive Cruise Control Aggressive Human-Driven Vehicle Autonomous Vehicle Cooperative Adaptive Cruise Control Connected and Autonomous Vehicle Connected Vehicle Highway Capacity Manual Human-Driven Vehicle Intelligent Driver Model Lane Change Model with Relaxation and Synchronization Level of Service Peachtree Industrial Boulevard Simplified Capacity Analysis Tool Simulation of Urban Mobility Traffic Control Interface TimeSpace Diagram Travel while Braking Vehicle Awareness Device

xi

EXECUTIVE SUMMARY
OVERVIEW Many studies that support an optimistic outlook on the traffic flow impacts of autonomous vehicles (AVs) limit modeled driving behavior modifications to the cooperative actions of the AVs. However, these studies have not considered the impacts on traffic performance of potential aggressive interactions of human-driven vehicles (HDVs) with AVs in a mixed environment (AVs and HDVs). Considering that AVs will not retaliate when they are the target of an aggressive action, it is not hard to postulate that some human drivers may display aggressive behaviors toward AVs, taking advantage of the AVs' collision-avoidance features. Given these potential behaviors, the objective of this effort is to develop and test models of AV interaction with aggressive human drivers.
To aid in understanding the potential impact of aggressive HDV (AHDV) interactions with AVs, this effort has investigated a merging situation at an off-ramp. Three classes of vehicles are simulated: AVs, HDVs, and AHDVs. AHDVs represent human-driven vehicles with aggressive merging-behavior characteristics. To perform this study, AHDV behavior at a merge section of a freeway exit ramp, in a mixed-traffic environment, is simulated using the open-source traffic simulation package SUMO (Simulation of Urban Mobility). Two types of potential AHDV merging behavior when interacting with an AV are modeled: (1) aggressive merge with maximum advancement, and (2) aggressive merge with zipper. The aggressive merge with maximum advancement represents the highest level of aggressive behavior. The AHDVs with this behavior target the farthest
1

reachable AV on the deceleration lane to act as the following vehicle in the receiving lane, i.e., the AHDV will lane change in front of the AV, essentially without regard for the available gap. In the second type, the aggressive merge with zipper, the AHDVs continue to target downstream AVs in the exit lane, but avoid the scenario where the same AV is targeted by multiple AHDVs.
The impacts of the AHDVs' aggressive behaviors in a mixed-traffic environment (i.e., AVs, HDVs, and AHDVs) on different network traffic characteristics, such as travel time and capacity, is demonstrated. Four experiments are conducted to explore the impact of the AHDV behavior on traffic operations. The first experiment observes the change in speed of the target AV, as well as the following traffic, when a platoon of 10 AHDVs merges in front of the AV near a freeway exit. The second and third experiments observe the travel times of exiting AHDVs and other vehicles when AHDVs are randomly distributed throughout the traffic stream with varying percentages of AVs and AHDVs in the traffic composition. The fourth experiment considers the impact on capacity in a similar merging situation where vehicle behavior is set as cooperative or noncooperative utilizing SUMO driver-behavior parameters.
Experiments 1 through 3 showed that the presence of human drivers' aggressive merging behaviors had adverse effects on AVs and HDVs. The adverse effects were more significant in high congestion, when there is a queue on the deceleration lane. The impacts of AHDVs' aggressive merges were muted by the larger headways between vehicles in low congestion when there is no queue on the deceleration lane. Based on the experiment 2 and experiment 3 results, AHDVs had a higher travel-time gain with higher level of aggressive behaviors, which in return had greater adverse effects on the AVs'
2

and the HDVs' travel times. Throughout the experiments, the system-wide travel time tended to be relatively stable, indicating that the AHDV travel-time improvements came at the expense of AVs' and other vehicles' travel times.
Experiment 4 took a closer look at the impact of cooperative behaviorinduced aggressive merges on capacity. It was seen that when most vehicles are either fully cooperative or noncooperative, similar capacities are obtained; however, where a higher percentage of cooperative vehicles are positioned to be targeted by more aggressive vehicles, this aggressive-to-non-aggressive interaction can significantly reduce capacity. In addition, it was seen, similar to experiments 1 through 3, that AHDV gains were achieved at the expense of AVs. Finally, even in those scenarios where the overall capacity was not significantly changed in response to the variation in the percentage of cooperative vehicles in the traffic, increased fluctuations in the flow may potentially negatively impact operations as well as the safety conditions in the upstream traffic.
As a final component of this research, an Excel-based Simplified Capacity Analysis Tool (SCAT) is developed. This tool draws predicted saturation flow rates, at various connected and autonomous vehicle (CAV) market penetration rates, from the literature and a simulation experiment. These saturation flow rates are utilized to determine potential phase capacities at a signalized intersection. While the freeway SUMO experiments focused on the impact of lane changing, SCAT explores the impact of CAV car-following and platooning behaviors. It is seen that a wide variation in capacity predictions may be found throughout the literature, from slight reductions to significant increases in capacity as AV market penetration increases. Across the literature, when considering the car-following aspect of AV operations, it is clear that two key sets of
3

assumptions are driving the predictions: the first is the headways selected by the AVs in a mixed-traffic environment, and the second is the characteristics of AV platoons, i.e., platooned vehicle spacing and maximum platoon length.
The findings of this study suggest that despite the general belief in the benefits of autonomous vehicles, there may be adverse impacts on the non-aggressive vehicle travel times in the presence of human drivers' aggressive merging behaviors in a mixed-traffic environment, especially in congested conditions. Thus, when the potential benefits of the AV are most needed, i.e., at or near capacity, it is possible that human interaction may negate many of the potential savings.
Report Organization Chapter 2 presents efforts found in the literature on AV modeling, such as assumptions made, frequently adjusted parameters, and common characteristics of AVs. Chapter 3 presents a comparison between the PTV VISSIM and SUMO simulation modeling platforms and discusses the selection of SUMO for the merge modeling efforts. Chapter 4 presents how the two aggressive merging models were developed, as well as the four different experiments that investigate the impacts of the aggressive merging models in a mixed traffic environment. Chapter 5 highlights the data collection conducted for the headway utilized to calibrate the model in chapter 6. Finally, chapter 6 presents the Simplified Capacity Analysis Tool.
Recommendations As seen in the report, the high state of uncertainty in AV driving-behavior characteristics and a similar level of uncertainty in the behavior of human-driven vehicles when
4

interacting with AVs, makes it extremely difficult to incorporate AVs into current planning and design processes with any sense of assuredness. However, based on this project, Georgia Department of Transportation (GDOT) can likely achieve an early sense of the ultimate operational impacts of AVs by tracking three primary leading indicators:
1. As AV tests continue, or low market penetration occurs, is a rise in aggressive interactions witnessed?
2. What are the headways being adopted by AV manufacturers, and what are the potential regulatory requirements?
3. Are platoons implemented in AVs, and, if so, what are the spacing requirements and maximum length restrictions, which are again potentially manufacturer and/or regulatory-agency driven?
As the direction of each of these indicators becomes clearer, GDOT will be able to select the more likely futures from the many potential predicted futures, with a higher level of confidence. This would allow AV penetration to begin to influence policy decisions and design decisions, such as queue management at ramp junctions, HV/AVs lane-usage restrictions, optimizing signalized intersections to process AV platoons, etc., in a more informed manner.
5

CHAPTER 1. INTRODUCTION
Many studies that support an optimistic outlook on the traffic-flow impacts of autonomous vehicles (AVs) limit modeled driving behavior modifications to the cooperative actions of the AVs, such as slowing down for merging vehicles. Similarly, lane changes to advance the AV position in the traffic stream relative to other vehicles receive low priority (Aria et al. 2016, Rahman and Abdel-Aty 2018, Stanek et al. 2017).
However, these and similar studies have not considered the impacts on traffic performance of potential aggressive interactions of human-driven vehicles (HDVs) with AVs in a mixed environment (i.e., AVs and HDVs), although such behaviors are likely to occur. For instance, mobility service companies have observed aggressive human driver behaviors directed at their AV test fleets, such as abrupt merging, tailgating, and hostile verbal and hand gestures (Randazzo 2018, Hamilton 2019). Even without AVs in the fleet, aggressive behavior has been observed at merge locations with heavy queuing. For example, within the last few hundred feet of a merge section an aggressive driver may take advantage of the slower acceleration and larger headways of heavy vehicles (Toth 2014). By extension, considering that AVs will not retaliate when they are the target of an aggressive action, it is not hard to postulate that some human drivers may display aggressive behaviors toward AVs, taking advantage of the AVs' collision-avoidance features. Even drivers that do not typically display such behavior may be more aggressive, or ignore common courtesies in vehicle interactions, when interacting with AVs. Given these potential behaviors, the objective of this effort is to develop and test models of aggressive merging behaviors, targeted toward AVs by a subset of humandriven vehicles, in a mixed environment.
6

To this end, the potential impact of merging behaviors on traffic performance is explored in a simulation environment. Three classes of vehicles are simulated: AVs, HDVs, and aggressive human-driven vehicles (AHDVs). AHDVs represent human-driven vehicles with aggressive merging-behavior characteristics. To perform this study, AHDV behavior at a merge section of a freeway exit ramp, in a mixed traffic environment, is simulated using the open-source traffic simulation package SUMO (Simulation of Urban Mobility) (Eclipse Foundation 2020). Two types of potential AHDV merging behavior when interacting with an AV are modeled: (1) aggressive merge with maximum advancement, and (2) aggressive merge with zipper. The aggressive merge with maximum advancement represents the highest level of aggressive behavior. The AHDVs with this behavior target the farthest reachable AV on the deceleration lane to act as the following vehicle in the receiving lane, i.e., the AHDV will lane change in front of the AV, essentially without regard for the available gap. In the second type, the aggressive merge with zipper, the AHDVs continue to target downstream AVs in the exit lane, but avoid the scenario where the same AV is targeted by multiple AHDVs. If an AV has already participated in a targeted merge with an AHDV, then the next AHDV will target the next AV upstream of that AV. Where an AV is not present, the AHDV will select and merge in front of an HDV in a non-aggressive manner, similar to HDVs.
Using simulation experiments, the impacts of the AHDVs' aggressive behaviors in a mixed-traffic environment (i.e., AVs, HDVs, and AHDVs) on different network traffic characteristics, such as travel time, is demonstrated. Three experiments are conducted to explore the impact of the AHDV behavior on traffic operations. The first experiment observes the change in speed of the target AV, as well as the following traffic, when a
7

platoon of 10 AHDVs merges in front of the AV near a freeway exit. The second and third experiments observe the travel times of exiting AHDVs and other vehicles when AHDVs are randomly distributed throughout the traffic stream with varying percentages of AVs and AHDVs in the traffic composition. The results of the three experiments show that the presence of AHDVs' aggressive behaviors lead to increased travel times, indicating higher levels of interruption in the traffic flow in a congested condition. As a final component of this research, an Excel-based Simplified Capacity Analysis Tool (SCAT) is developed. This tool draws predicted saturation flow rates, at various connected and autonomous vehicle (CAV) market penetration rates, from the literature and a simulation experiment. These saturation flow rates are utilized to determine potential phase capacities at a signalized intersection. While the freeway SUMO experiment focused on the impact of lane changing, SCAT explores the impact of CAV car-following and platooning behaviors. The remainder of this report is organized as follows. Chapter 2 provides background information on AV driving behaviors, the interaction between AV and human-driven vehicles, and modeling of aggressive AVs. Chapter 3 presents the process for the selection of the simulation modeling platform. Chapter 4 reviews the development of the selected simulation. Chapter 5 presents the results. Chapter 6 presents the Simplified Capacity Analysis Tool. Finally, chapter 7 summarizes the findings.
8

CHAPTER 2. BACKGROUND
UNCERTAINTY IN INTERACTIONS BETWEEN AUTONOMOUS VEHICLES AND HUMAN ROAD USERS Over the past decade, the rapid advancement in autonomous driving technology in research and in industry has led several automobile manufacturers to develop and deploy various levels of autonomous vehicles. Numerous studies present optimistic roadway performance outlooks given the deployment of autonomous vehicles. However, there is a gap in the understanding of the impacts of the autonomous vehicles' interactions with human drivers, which is crucial for reliably modeling the impacts of AV implementation. This is particularly relevant in the transition phase where roadways are expected to consist of a mixed fleet of AVs and HDVs. Such a mixed fleet may result in significant changes to roadway safety, operational, environmental, and other performance metrics.
A significant source of the current uncertainty stems from the lack of standardization in autonomous vehicles' driving behaviors (National Highway Traffic Safety Administration [NHTSA] 2017, Zhao et al. 2019). However, human drivers' actions toward AVs are also a significant source of uncertainty. For instance, the trends in people's perception and behavior toward AVs are captured in several recent studies. Results of a survey conducted in 2016 indicated that the majority of the respondents would feel uncomfortable driving alongside an AV (Tennant et al. 2016). It has been suggested that given such concerns, AV and HDV interaction behaviors may contribute to traffic disturbances, particularly under low AV market penetration levels (Nishimura et al. 2019). An intersection field study by Rothenbucher et al. (2016) observed changes in pedestrian's and bicyclist's behavior in the presence of AVs. The pedestrians and
9

bicyclists acted in a conservative manner, which is hypothesized to be due to their uncertainty in potential AV behaviors. Further, in a few studies, field experts have shared concerns on the possibility of human drivers displaying aggressive behaviors toward AVs, taking advantages of AVs' conservative behaviors (Mller et al. 2016, Hedlund 2017). These concerns on human drivers' aggressive behaviors targeted to AVs have been observed on real-world roadways. News articles have reported that mobility service companies such as Uber and Waymo have been observing human drivers' behaviors such as aggressive merging, tailgating, and hostile verbal and hand gestures directed toward their autonomous vehicles (Randazzo 2018, Hamilton 2019).
PREVIOUS RESEARCH ON MODELING AUTONOMOUS VEHICLE BEHAVIORS As stated previously, one of the key common challenges experienced in modeling AV behavior is the lack of standardization in driving behavior parameters (NHTSA 2017, General Motors 2015, National Academies of Sciences, Engineering, and Medicine 2017). While it is challenging to anticipate and model AV driving behaviors, the ability to utilize current human driving-behavior models with minimal modifications to model AV driving behaviors has been explored by numerous researchers (Stanek et al. 2017, Wagner 2016). Most studies model AV driving behaviors using traditional car-following models and lane-changing models with customized decision-making processes and modified parameter values that assign certain characteristics to the AVs. The following list provides the frequently assumed behavioral modifications from human driverbehavior models to AV/CAV models.
10

Frequently assumed behavioral modifications for AV/CAV vehicles include the following:
Lower headways. Lower deviation or zero randomness in speed variation from speed limit. Lower reaction time. Slows down for merging vehicles (cooperative lane change). Higher acceleration rate. Table 1 summarizes AV/CAV behavior assumptions and parameter adjustments for a sample of roadway application studies drawn from the literature. A more detailed discussion of several of these studies may be found in chapter 6.
11

Author(s), Year
Aria et al., 2016
Stanek et al., 2017
Mesionis et al., 2020
Richter et al., 2019
Yu et al., 2019
Li et al., 2020

Table 1. Summary of AV/CAV behavior assumptions and parameters employed.

AV/ CAV AV AV AV
AV AV AV

Simulation Software VISSIM
VISSIM 9
AIMSUN

Scenario Roadway

Anticipated AV/CAV Behaviors

Segment of an autobahn with weaving area,
on-ramp, and off-ramp
Two freeway segments in California
20-mile stretch of 3-lane freeway

Freeway Lower headways Fixed range in scanning surroundings Lower speed deviation from the speed
limit Earlier decision point in lane change Cooperative lane change Lower headways Higher acceleration rate Lower reaction time to green light Cooperative lane change Earlier lane change for turns Not accepting lower gap for merge Less likely to overtake other vehicles Lower reaction time Lower headway

SUMO

Freeway segment with an on- ramp and an acceleration lane

Lower headway Lower reaction time No randomness in speed Earlier lane change for merging Slow down for merging vehicle

AIMSUN

3-lane freeway segment with on-ramp and off-ramp

Lower headways Lower acceleration rate

VISSIM

Highway section with on-ramp and off-ramp

Lower standstill distance Lower headways Lower reaction time Smaller oscillation in speed Smaller oscillation during acceleration Cooperative to merging vehicles

Adjusted Parameters (Parameter Name)
Lower headway time (CC1) Higher look ahead (and back) distance Lower speed dependency of oscillation (CC6) Advanced merging Cooperative lane change
Lower standstill distance (CC0) Lower headway time (CC1), Higher threshold for entering following (CC3) Lower negative and positive following threshold (CC4, CC5) Advanced merging Disabled lower gap acceptance for merge Lower probability in overtaking other vehicles Lower reaction time Lower standstill distance (CC0) Lower time headway (tau) Smaller simulation length Removed randomness in speed (SpeedFactor) Zero driver imperfection (sigma) Earlier merging decision point Higher cooperation to merging vehicles Lower time headways Lower minimum gap Lower acceleration rate Lower standstill distance (CC0) Lower headway time (CC1) Lower negative and positive following threshold (CC4, CC5) No speed dependency of oscillation (CC6) Lower oscillation acceleration (CC7) Lower lane changing minimum headway (LC4) Cooperative lane change

12

Author(s), Year Liu, P. and Fan,
W. 2020 Papadoulis et al.,
2019 Li and Wagner,
2019
Toms et al., 2019
Sukennik et al., 2018
Martin-Gasulla et al., 2019

AV/ CAV CAV CAV AV
AV
AV
CAV

Simulation Software

Scenario Roadway

Anticipated AV/CAV Behaviors

Adjusted Parameters (Parameter Name)

Higher acceleration rate

Higher acceleration rate

VISSIM 4-lane freeway segment Lower car following distance

Lower headway (even lower headway for between two

Lower desired headway

CAVs) (CC0, CC1)

3-lane freeway segment Lower time gaps

Lower headway time (CC1)

VISSIM 7 with two on-ramps and Higher distance in observing surrounding Lower minimum time gap in lane changing (MG1)

two off-ramps

vehicles

Higher look ahead (and back) distance

SUMO

3-lane freeway with two on-ramps and an off-ramp

Lower time-gap Lower driver imperfection Higher compliance rate to speed limits Slows down for merging vehicles

Lower time gap (tau) Zero driver imperfection (sigma) Lower deviation from speed limit (SpeedFactor) Higher cooperative behavior (lcCooperative)

Lower standstill distance (CC0)

Lower headway time (CC1)

Lower threshold for entering following (CC3)

Lower headway

Lower negative and positive following thresholds

VISSIM 9

3-lane freeway segment Higher acceleration rate

for 9 km

Lower variation in acceleration

(CC4/CC5) Lower oscillation acceleration (CC7)

Greater acceptable merging gap

Higher standstill acceleration (CC8)

Higher acceleration at 80 km/hr (CC9)

Higher lane changing minimum headway (LC4)

Lower safety distance reduction factor (LC5)

Lower standstill distance (CC0)

VISSIM 10

Urban roads and freeway

Lower headway

Lower headway time (CC1) Lower following variation (CC2)

Lower threshold for entering following (CC3)

Higher speed stability and headway to

VISSIM 11

Straight single-lane freeway

leading vehicle Lower headway when following CAVs Higher headways when following
conventional vehicles (compared to

Lower headway when CAV following another CAV Higher headway CAV following conventional vehicle
(higher than conventional vehicles' headway)

headways of human drivers)

13

Author(s), Year
Tiblijas et al., 2018
Morando et al., 2018
Atkins, 2016
Anagnostopoulos and Kehagia, 2019
Wang and Wang, 2017

AV/ CAV AV
AV CAV
CAV AV

Simulation Software

Scenario Roadway

Anticipated AV/CAV Behaviors Roundabout

VISSIM 11

Single-lane roundabout with 3 or 4 approaches

Lower headways Lower reaction time Lower speed instability Higher acceleration rates

VISSIM 9

Single-lane roundabout with 4 approaches

Shorter gap No randomness in speed Higher acceleration rate Higher capability in observing vehicles
ahead

VISSIM 8

Roundabout

Lower headway Lower variation in acceleration rate Lower safety distance Higher acceleration rate

VISSIM 11

Double lane roundabout

Lower headways Zero speed oscillation Lower acceleration rates Higher look ahead distance Slow down for merging vehicles

VISSIM 7

4 km single-lane roadway

Single Lane Roadway
Lower reaction time, lower headways No speeding

Adjusted Parameters (Parameter Name)
Lower standstill distance (CC0), Lower headway time (CC1), Lower following variation (CC2) Lower negative and positive following threshold (CC4, CC5) Lower speed dependency of oscillation (CC6) Lower oscillation acceleration (CC7) Higher standstill acceleration (CC8) Higher acceleration at 80 km/hr (CC9) Lower standstill distance (CC0) Lower headway time (CC1) Lower following variation (CC2) Lower negative and positive following threshold (CC4, CC5) Zero speed dependency of oscillation (CC6) Higher acceleration rate Higher look-ahead distance Lower standstill distance (CC0) Lower headway time (CC1) Lower oscillation acceleration (CC7) Lower safety distance, higher standstill acceleration (CC8) Lower minimum time gap in lane changing (MG1) Lower minimum headway in lane changing (MG2) Lower standstill distance (CC0) Lower headway time (CC1) Zero following variation (CC2) Zero negative and positive following threshold (CC4, CC5) Zero speed dependency of oscillation (CC6) Higher acceleration rates (CC7, CC8, CC9) Higher look ahead distance Cooperative lane change
Lower standstill distance (CC0) Lower headway time (CC1) Tighter bounds on speed distribution

14

Author(s), Year

AV/ CAV

Lu et al., 2018

AV

Atkins, 2016

CAV

Lu et al., 2020

AV

Atkins, 2016

CAV

Atkins, 2016

CAV

Wang and Wang, 2017

AV

Elvarsson, 2017

AV

Simulation Software
SUMO

Scenario Roadway Single-lane roadway

Anticipated AV/CAV Behaviors
Lower headways Higher acceleration rate Zero driver imperfection

VISSIM 8

Single-lane link

Lower headway Lower variation in acceleration rate Lower safety distance
Multilane Roadway

SUMO

2-lane roadway in a grid Lower headways

network

Higher acceleration rate

VISSIM 8

Multi-lane link

Lower headway Lower variation in acceleration rate Lower safety distance Greater acceptable merging gap

VISSIM 8

Multi-lane link with merge

Lower headway Lower variation in acceleration rate Lower safety distance Higher acceleration rate

VISSIM 7 VISSIM 9

1 km single-lane signalized intersection
Main arterial roadway with two signalized intersections

Signalized Intersection
Lower reaction time Lower headways No speeding
Lower acceleration rate Lower deceleration rate Tighter bound on speed distribution Lower headway

Adjusted Parameters (Parameter Name)
Lower minimum gap when standing (minGap) Lower time headway (tau) Higher acceleration rate Zero driver imperfection (sigma) Lower standstill distance (CC0) Lower headway time (CC1) Lower oscillation acceleration (CC7) Lower safety distance reduction factor (LC5)
Lower offset to the leading vehicle when standing (minGap) Higher acceleration rate Lower time headway (tau) Zero driver imperfection (sigma) Lower standstill distance (CC0) Lower headway time (CC1) Lower oscillation acceleration (CC7) Lower lane changing min. headway (LC4) Lower safety distance reduction factor (LC5) Lower standstill distance (CC0) Lower headway time (CC1) Lower oscillation acceleration (CC7) Higher standstill acceleration (CC8) Higher acceleration rate at 80 km/hr (CC9) Lower minimum time gap in lane changing (MG1) Lower minimum headway in lane changing (MG2)
Lower reaction time Lower standstill distance (CC0) Lower headway time (CC1) Tighter bounds on speed distribution Lower acceleration rate Lower deceleration rate Tighter bound on speed distribution Lower standstill distance (CC0)

15

Author(s), Year Atkins, 2016
Espinosa, 2015
Morando et al., 2017

AV/ CAV CAV
AV
AV

Simulation Software VISSIM 8
VISSIM 8
VISSIM 9

Scenario Roadway Signalized junction
6-lane signalized intersection
3-lane signalized intersection

Anticipated AV/CAV Behaviors

Adjusted Parameters (Parameter Name)

Lower headway

Lower standstill distance (CC0)

Lower variation in acceleration rate

Lower headway time (CC1)

Lower safety distance

Lower oscillation acceleration (CC7)

Higher acceleration rate

Higher standstill acceleration (CC8)

Increased range in distance and in number

of surrounding vehicles to observe

Higher look ahead (and back) distance

surrounding conditions

Lower headway time (CC1),

Lower headway

Cooperative lane change

Slow down for merging vehicle

Advanced merging

Earlier decision point for merge

Lower standstill distance (CC0)

Lower headways Zero speed oscillation Higher acceleration rate Greater look ahead distance

Lower headway time (CC1) Lower following variation (CC2) Lower negative and positive following threshold (CC4, CC5) Zero speed dependency of oscillation (CC6) Higher acceleration rates (CC8, CC9)

Higher look ahead distance

16

A common simulation platform utilized for modeling AVs is VISSIM. Table 2 summarizes the list of frequently and infrequently used modified VISSIM parameters.

Table 2. List of frequently and infrequently used VISSIM parameters.

Frequently Used VISSIM Parameters
Standstill Distance (CC0)
Headway Time (CC1)
Following Variation (CC2)
Negative Following Threshold (CC4)
Positive Following Threshold (CC5)
Speed Dependency of Oscillation (CC6)
Oscillation Acceleration (CC7) Standstill Acceleration (CC8) Acceleration at 80km/hr (CC9)

Studies That Used Parameter
Stanek et al. 2017, Mesionis et al. 2020, Li et al. 2020Liu and Fan 2020, Toms et al. 2019, Sukennik et al. 2018, Tiblijas et al. 2018, Morando et al. 2018, Atkins 2016, Anagnostopoulos and Kehagia 2019, Wang and Wang 2017, Elvarsson 2017, Morando et al. 2017
Aria et al. 2016, Stanek et al. 2017, Li et al. 2020, Liu and Fan 2020, Toms et al. 2019, Sukennik et al. 2018, Tiblijas et al. 2018, Morando et al. 2018, Atkins 2016, Anagnostopoulos and Kehagia 2019, Wang and Wang 2017, Espinosa 2015, Morando et al. 2017
Stanek et al. 2017, Sukennik et al. 2018, Tiblijas et al. 2018, Morando et al. 2018, Anagnostopoulos and Kehagia 2019, Morando et al. 2017
Stanek et al. 2017, Li et al. 2020, Toms et al. 2019, Tiblijas et al. 2018, Morando et al. 2018, Anagnostopoulos and Kehagia 2019, Morando et al. 2017
Li et al. 20, Toms et al. 2019, Tiblijas et al. 2018, Morando et al. 2018, Anagnostopoulos and Kehagia 2019, Morando et al. 2017
Aria et al. 2016, Li et al. 2020, Tiblijas et al. 2018, Morando et al. 2018, Anagnostopoulos and Kehagia 2019, Morando et al. 2017
Li et al. 2020Toms et al. 2019, Tiblijas et al. 2018, Atkins 2016, Anagnostopoulos and Kehagia 2019, Morando et al. 2017
Toms et al. 2019, Tiblijas et al. 2018, Atkins 2016, Anagnostopoulos and Kehagia 2019, Morando et al. 2017
Toms et al. 2019, Tiblijas et al. 2018, Atkins 2016, Anagnostopoulos and Kehagia 2019

17

Frequently Used VISSIM Parameters
Look Ahead (and Back) Distance
Cooperative Lane Change
Infrequently Used VISSIM Parameters
Threshold for Entering Following (CC3)
Advanced Merging
Minimum Time Gap in Lane Changing (MG1)
Lane Changing Minimum Headway (LC4/MG2) Lower Safety Distance Reduction Factor (LC5)

Studies That Used Parameter
Aria et al. 2016, Morando et al. 2018, Espinosa 2015, Morando et al. 2017 Aria et al. 2016, Li et al. 2020, Morando et al. 2018, Espinosa 2015
Studies Used Parameter
Stanek et al. 2017, Toms et al. 2019, Sukennik et al. 2018 Aria et al. 2016, Mesionis et al. 2020, Espinosa 2015
Mesionis et al. 2020, Atkins 2016
Toms et al. 2019, Atkins 2016
Toms et al. 2019, Atkins 2016, Elvarsson 2017

The two most common customizations to model AV driving behaviors are: (1) cooperative responses to other road users, and (2) conservative driving behavior. Examples of cooperative responses and conservative driving behaviors include AVs slowing down to allow vehicles to merge in front of them and AVs not changing lanes for speed gain, respectively (Nishimura et al. 2019, Liu et al. 2017, Hua et al. 2020).

AVHUMAN INTERACTION MODELING Models of AV driving commonly assume conservative behaviors where AVs interact with pedestrians, such as AVs responding to pedestrians much earlier relative to human perception (Kapania et al. 2019); or when interacting with HDVs, AVs reduce speed or change lanes to allow the HDV to merge (Stanek et al. 2017; Liu et al. 2018). Additionally, in a few studies, AVs are modeled to adjust their driving behaviors and decision-making process based on the observed or predicted behaviors of human drivers (Wei et al. 2013, Tian et al. 2018). Despite differences in approaches to

18

modeling AV driving behaviors, there is a common goal of determining AV driving decisions based on cooperative behavior with neighboring road users (Schwarting et al. 2018). The customizations of AV characteristics are based on the common view of conservative and cooperative AV driving behaviors (Mller et al. 2016). Table 3 presents a summary of AV-and-human-driver interaction modeling approaches reviewed in the literature survey. In the table, the discussed car-following models and the corresponding modifications are used to model vehicles' longitudinal movement, as the car-following models govern the speed and headway controls. Similarly, the discussed lane-changing models and the corresponding modifications are used to model vehicles' lateral movements, as the lane-changing models govern the decisionmaking process in vehicle merges. The following abbreviations are used in the table; CV Connected Vehicle, CACC Cooperative Adaptive Cruise Control, and ACC Adaptive Cruise Control. The `Additional Modification' column refers to any additional change that was made by the author(s) to the adopted car-following model or lane-changing model. Two key observations from the table below are: (1) VISSIM and SUMO are the dominant simulation platforms that were utilized among the reviewed studies, and (2) studies that utilized SUMO used SUMO's default lane-changing model without any additional modification. The absence of additional modification on lane-changing model suggests aggressive lane changing behaviors were not considered.
19

Table 3. Summary of AV Human interaction modeling approaches.

Author(s) Rahman and Abdel-Aty
2018 Liu et al.
2018 Nishimura et al. 2019
Stanek et al. 2017
Tian et al. 2018
Wagner 2016
Wei et al. 2013
Zhao et al. 2019
Zhou et al. 2017

Simulation Tool
VISSIM
AIMSUM
Scenargie
VISSIM
SUMO
Driving Simulator (Scenario Built with MATLAB and PreScan)
-

Scenario Settings

Scenario Layout

Mixed Traffic

3-lane Freeway
4-lane Freeway Mainline & Single-lane On-ramp
3-lane Straight Roadway 20 Freeway Miles, 15 Freeway
Interchanges, 3 Parallel Arterial Corridors,
32 Intersections
Single-lane Roundabout
City Network
2-lane Freeway Mainline and Single-lane
On-ramp

Yes (CV Platoon and Human Vehicles) Yes (ACC/CACC Platoon and
Human Vehicles) Yes (AV and
Human Vehicles)
Yes (AV and Human Vehicles)
Yes (1 AV and 1 Human Vehicle)
Yes (AV and Human Vehicles)
Yes (AV and Human Vehicle)

3-lane Roadway

Yes (AV and Human Vehicles)

3-lane Freeway

Yes (AVs and Human Vehicles)

AV-HDV Interaction
Yes Yes Yes
Yes
Yes No Yes
Yes Yes

Longitudinal Movement

Use of Existing Car-following
Model

Additional Modification

Wiedemann 99 / IDM1

Gap Control

Multiple Sources

(IDM, Gipps, Newell, and

-

Shladover)

IDM

Acceleration Control

Lateral Movement

Use of Existing Lanechanging Model

Additional Modification

VISSIM Default

Merge Control

Lateral Movement Logics Developed by
Shladover
LMRS2

CACC Operation Rules
Merge Control

Wiedemann 74

Acceleration and Gap Control

VISSIM Default

Merge Control

Discrete-time Model
Helly's Model
Markov Decision Process

Speed Control

Game Theoretic Decision-making
Model
SUMO Default
Lateral decision made in the developed algorithm

Driver Type Estimation
-
-

Default Carfollowing Model

Speed Control

Default Lane-change Model

-

CIDM3

-

CIDM

Merge Control

20

Scenario Settings

Author(s)

Simulation Tool

Scenario Layout

Mixed Traffic

Aria et al. 2016

VISSIM

3-lane Freeway / Yes (AVs and 4-lane Freeway Human Vehicles)

Tiblijas et al. 2018

VISSIM

Various Sizes of Yes (AVs and Roundabouts Human Vehicles)

Hua et al.

Cellular 2-lane Freeway / Yes (CAVs and

2020

Automata 3-lane Freeway Human Vehicles)

Guo et al. 2020

Python-based Simulation

3-lane Freeway with On-ramp and Off-ramp

Yes (CAVs and Human Vehicles)

Liu et al. 2017

Cellular Automata

3-lane Freeway

Yes (AVs and Human Vehicles)

1 Intelligent driver model. 2 Lane-change model with relaxation and synchronization.

3 Cooperative intelligent driver model.

4 TakagiSugeno fuzzy model.

5 Quality or value-based learning algorithm. 6 NagelSchreckenberg model.

7 Symmetric two-lane cellular automaton model.

AV-HDV Interaction
Yes Yes Yes
Yes
Yes

Longitudinal Movement

Use of Existing Car-following
Model

Additional Modification

Wiedemann 99

Gap Control

Wiedemann 74
TS Model4
Enhanced Q-learning Algorithm Rules from NaSch6
Model

Gap Control Gap and Speed
Control
Trajectory Planning
-

Lateral Movement

Use of Existing Lanechanging Model

Additional Modification

Wiedemann 99

Merge Control

Wiedemann 74

-

TS Model

Lane Selection

Enhanced Q-learning5 Algorithm

Merge Control

STCA7 Model

Merge Control

21

AGGRESSIVE BEHAVIOR MODELING APPROACH To model different levels of aggressive behaviors between AVs and HDVs, Liu et al. (2017) developed different levels of aggressive lane-changing modes. In their effort, HDVs are assigned to the higher aggressive lane-changing mode (Liu et al. 2017). In another study, aggressive behavior in AV and HDV interaction is introduced by allowing HDVs to force lane changes that caused the following AVs to slow down to create a sufficient gap for the merge (Liu et al. 2018). Studies that allowed HDVs to behave aggressively toward AVs observed greater traffic-flow instability with increased penetration levels of AVs (Liu et al. 2018) or a higher number of incomplete trips with AVs traveling at extremely low or high speed (Nishimura et al. 2019). These results are drastically different from findings of improved safety and reduced travel times in other AV studies that did not consider aggressive behaviors (Rahman and Abdel-Aty 2018, Aria et al. 2016). The possibility of such behaviors targeted at a given vehicle type is not without precedence. For example, mobility service companies such as Uber or Lyft reported that HDVs will exhibit aggressive behaviors specifically toward AVs (Randazzo 2018, Hamilton 2019).
Most current studies allow the HDVs to increase aggressiveness based on the availability of gap distance, regardless of the type of the following vehicle type in the target lane (Nishimura et al. 2019; Liu et al. 2018). In order to investigate the interaction impacts of HDVs' biased aggressive behaviors toward AVs, the levels of aggressive driving behaviors of the HDVs should vary based on the target vehicle type. Modeling different aggressive driving behaviors based on vehicle type requires flexible simulation models that allow for real-time adjustments of driving behavior characteristics and parameters based on vehicle types.
22

CHAPTER 3. SELECTION OF SIMULATION MODEL APPROACH
INTRODUCTION A key objective of this study is to model the interaction between aggressive humandriven vehicles and AVs. Modeling AV and aggressive human-driven vehicle driving behaviors, and developing various fleet-penetration scenarios, requires simulation software capabilities such as flexibility in driving models and real-time interaction with agents during simulation run-time. This section presents the evaluation criteria utilized for simulation software selection. Two software packages are considered, SUMO and VISSIM (Eclipse Foundation 2020, PTV Group 2021). Table 4 lists the key evaluation criteria identified for AV modeling. The following sections explain the reasoning for each of the criteria and the capabilities of the two software packages SUMO and VISSIM corresponding to each evaluation criterion.
KEY EVALUATION CRITERIA General Information Both software packages have interfaces that enable run-time communication with agents, such as vehicles and signal control systems. SUMO's source code is publicly available and accessible as an open-source platform, whereas VISSIM offers source codes on AddIns.
Availability of Driving Models Different car-following models or lane-changing models may be more appropriate to model the AV characteristics based on the given scenario. For example, an area with higher interaction with other vehicles, such as a roundabout, may require a different modeling complexity than a signal-controlled intersection with protected-only
23

movements. SUMO offers a higher number of car-following and lane-changing models compared to VISSIM. However, both software packages offer significant flexibility by allowing users to modify existing models or import customized models. Critically, such capabilities allow users to assign variations in vehicle behaviors by vehicle types.
Driver Behavior Model Parameters Through the literature survey it is seen that due to the lack of standards in AV driving behaviors many AV modeling approaches rely on assumptions and expectations of AVs' anticipated behaviors. As a result, studies that simulate AVs may share certain characteristics such as conservative, cooperative, or cautious; however, they differ in the driving-behavior models and the parameter values utilized. SUMO's drivingbehavior models have an arguably higher level of flexibility given a higher number of model parameters and the open-source nature of the software. However, as stated, both software packages allow the control of parameters during run-time. Such capability is particularly critical in AV modeling as it allows the AV model to adapt to different roadway conditions.
Data Export The outputs that record the states of the simulation agents, such as vehicles, lanes, and signal control systems, are used to measure key performance indices to test the study hypothesis. Both software packages allow the user to extract various outputs that can be used to measure model performance. Critically, both models allow for the output of vehicle trajectory data, from which numerous other measures may be derived.
24

User Contribution VISSIM has been widely used in the public and private sectors as well as academic and research settings, while SUMO has been primarily used for academic and research efforts. A large user community may be helpful in determining experimental designs and model development as previous similar efforts may be used for reference. Efficiency may also be gained through the adoption of previously developed models or findings. For example, a previously developed driving-behavior model can be imported to SUMO or VISSIM for AV modeling.
SUMO has an online community in which users exchange knowledge and contribute to improve SUMO functionality. SUMO's ACC and CACC car-following models are developed by a research team (Lopez et al. 2018). Such interactions among users and between users and developers help enable referencing and learning from previous studies.
Signal Control System Signal control systems are often used to simulate different traffic patterns and flow rates. Complex traffic control systems can be established via the respective interface in SUMO and VISSIM, such as the vehicle preemption signal control system. VISSIM offers more variations of signal control systems that are readily available to users.
Overall Comments Each software package has advantages and disadvantages that can be weighted differently based on the study objective. Based on the identified criteria for modeling AV driving behavior, SUMO's key advantages include the source code availability and the number of driving-behavior models and parameters. These enable the
25

identification of key parameters to model AV characteristics and provide significant flexibility for modeling AV behavior. VISSIM's key advantages include a larger user base and greater variety of signal control systems that can aid in more efficient implementations of complex traffic control systems. The research team selected SUMO for this study because SUMO's key advantages were critical in modeling aggressive human driving behavior and the subsequent AV response. Additionally, being able to access the driver behavior model and parameters via the source code helped the team to identify key parameters to control to model the aggressive merging behaviors as well as the AV response.
26

Table 4. Evaluation criteria summary on simulation tools.

Evaluation Criteria

SUMO (Lopez et al. 2018) VISSIM (PTV Group 2021)

1. General Information

Compatibility with Operating Systems

Windows, Linux, macOS

Windows, Linux

Source Code

Written in C++

Written in C++

Source Code Availability

Yes; online source code library publicly accessible

AddIn source codes available

Interface

TraCI

COM & EDM

Programming Language Compatible with Python,

Compatible with Python,

Compatible with Interface

C++, MATLAB, Java

C++, MATLAB, Java

Accessibility

Open Source (publicly accessible)

Commercial

2. Availability of Driver Models

Car-following Models

14 car-following models imported

2 car-following models imported

Lane-changing Models

3 lane-changing models imported

2 lane-changing models imported

Import-customized Models

Yes

Yes

Modification on Existing Models

Yes

Yes

3. Driver Behavior Model Parameters

Vehicle Attribute Parameters

38 parameters

10 parameters

Lane-changing Model Parameters

23 parameters

14 parameters

Parameters Adjustable in Simulation?

Yes

Yes

4. Data Export

Export File Format

xml

Various Types

Output Data Type

Vehicle-based, detectorbased, simulation-based, traffic light-based, lane-
based

Vehicle-based, pedestrianbased, traffic light-based, lane-based, detector-based

5. User Contribution

User Base

Academic/Research Sector

Academic/Research Sector, Private Sector, Public Sector

Yes; online discussion

Allows External Contribution?

forum, imported a userdeveloped car-following

No

model

6. Traffic Light System

Fixed Time Method

Yes

Yes

Coordinated Method

Yes

Yes

Actuation Method

Yes

Yes

Ring-Barrier Sequence

No

Yes

Optimization

No

Yes

27

CHAPTER 4. SIMULATION MODELING OF DRIVER BEHAVIOR
AGGRESSIVE MERGE BEHAVIOR MODEL DEVELOPMENT Simulation Tool For this study, SUMO (version 1.6.0) is used to simulate the merging behavior scenarios. The SUMO's Traffic Control Interface (TraCI) is utilized for modeling vehicle interaction behaviors. TraCI provides access to the values of the simulation objects during run-time, enabling customization of vehicle behaviors (Eclipse Foundation 2020). In the current application, to model aggressive AHDV merge behaviors, TraCI is used to retrieve the real-time speed of target vehicles, control the AHDVs' speed for overtaking the AV, and force the merge in front of the target AV by accepting low front and rear gaps. In this study, Python 3.8 (Python 2019) is used for developing the TraCI scripts.
Network Layout Figure 1Figure 1. Diagram. Roadway layout on merging zone, highlighted in yellow.
shows the modeled network layout. It consists of two through lanes with a 600 ft deceleration lane to an exit ramp. The aggressive merge behavior occurs in the area near the freeway exit ramp, i.e., the merging zone. The two-lane freeway extends for 1 mile upstream of the beginning of the merge zone, allowing for sufficient space for vehicles to queue during congestion without spilling out of the network. The outflow from the ramp is controlled by a simple two-phase, pretimed traffic signal, with the splits and cycle length set dependent on the modeled scenario; they are specified in each experiment. The ramp junction with the cross street is 1,500 ft downstream from the ramp gore. Figure 2. Diagram. Speed by lane.
28

shows the simulated speed for each lane and the lane labels used in the study.

Figure 1. Diagram. Roadway layout on merging zone, highlighted in yellow.

Lane A_0: 70mph (29.05m/s) Lane A_1: 60mph (29.05m/s)

Lane B_0: 70mph (29.05m/s) Lane B_1: 60mph (29.05m/s) Lane B_2: 55mph (24.59m/s)

Figure 2. Diagram. Speed by lane.
Vehicle Classification and Characteristics To study the interactions between the different vehicle types, three vehicle classes are defined based on driving behaviors, as follows:
1. AVs: Exhibit cooperative driving behaviors; i.e., AVs slow to extend the leading gap, allowing merging vehicles to more easily enter their lane. When AVs exit the freeway, they will change lanes at the start of the deceleration lane. AV behavior is fully controlled by SUMO.
2. AHDVs: Travel on a high-speed lane (Lane B_1, figure 2) until they merge into the deceleration lane. These vehicles exhibit aggressive merge behavior toward AVs by accepting smaller gaps (minimum half-vehicle in length to begin a merge) in front of the AV compared to a merge in front of an HDV. AHDVs will always seek to merge in front of the farthest reachable downstream AV in the deceleration lane. When making an aggressive merge, the AHDV behavior is controlled through TraCI. When an AV is not

29

reachable, the AHDV merge behavior will be non-aggressive and controlled by SUMO. 3. HDVs: Exhibit the same cooperative driving behaviors as that of AVs but are not targeted by AHDVs. HDV behavior is fully controlled by SUMO.
Aggressive Behavior Model The objective of the aggressive merging behavior is to perform an aggressive lane change in front of a target vehicle. AHDVs' aggressive merge behaviors consist of customizing two key behaviors of AHDVs--target selection behavior and merging behavior. The objective of the target selection behavior is to identify the optimal target vehicle. When queueing occurs on the deceleration lane, the targeting behavior allows AHDVs to travel on the higher-speed lane (Lane B_1) until merging in front of target vehicles in Lane B_2, thus allowing AHDVs to make queue-jumps. After a target is selected, the AHDVs adjust their speed, within the constraints of the presence of other vehicles in front of them on the same lane and seek to merge in front of the target vehicle. As mentioned in chapter 1, two merge types are considered in the study based on the levels of aggressiveness in the targeting behavior: (1) aggressive merge with maximum advancement, representing the highest level of aggressive merge; and (2) aggressive merge with "zipper" action, representing a moderate level of aggressive merge. The merging behavior process is similar for these two merge types; however, they differ from each other in their target behavior process. The next section describes the target behavior process for these two types of aggressive merges.
30

Aggressive Merge Behavior Model: Target Selection Process Aggressive Merge with Maximum Advancement The objective of the target selection behavior in the aggressive merge with maximum advancement is to identify the optimal target vehicle, which in this study is considered the AV farthest downstream in the deceleration lane. When queueing occurs on the deceleration lane (Lane B_2), this targeting behavior allows AHDVs to travel on the higher-speed adjacent lane (Lane B_1) until merging in front of target vehicles in the deceleration lane, thus allowing AHDVs to queue-jump. To implement this behavior, an AHDV's initial target is the closest AV on the target lane. After a target is selected, the AHDV adjusts its speed to overtake the AV, within the constraints of the speed of the leading vehicle in the same lane (if one is present) or the lane speed. Once the AHDV is in the vicinity (to be defined subsequently) of the target vehicle, the AHDV checks if the next downstream AV is reachable prior to the end of the deceleration lane. By repeating this process, the AHDV merges in front of the farthest reachable downstream AV. As every AHDV targets the farthest AV, this behavior often results in multiple AHDVs merging in front of the same AV, as shown in Figure 3. Diagram. Aggressive merge behavior of AHDVs toward AVs..
Figure 3. Diagram. Aggressive merge behavior of AHDVs toward AVs. (AHDV deep blue vehicles, AV light blue vehicles, and HDV white vehicles) If an AHDV's target AV becomes no longer reachable due to a speed change or interference from other vehicles in the AHDV's lane, then the AHDV seeks to merge in front of an HDV. However, the merge in front of the HDV no longer uses
31

aggressive gap selection; rather, it is fully controlled by SUMO. If SUMO is unable to successfully complete the merge and an AV from upstream on the deceleration lane begins to overtake the AHDV (which may occur when congestion results in a lower speed on the mainline lane than that on the deceleration lane), the AHDV returns to its aggressive behavior and merges in front of the approaching AV. Aggressive Merge with Zipper In the aggressive merge with zipper case, to target a vehicle for merge, AHDVs first check whether there is an AHDV merge occurring downstream. If there is an aggressive merge downstream, the AHDVs do not target the same AV affected by the previous merge but rather target any following AV behind the last merge's target AV, as shown in figure 4. This selection of a new target vehicle, different from the last merge's target vehicle, makes this aggressive merge with zipper case less aggressive compared to the aggressive merge with maximum advancement case. Targeting the following vehicle of the last merge results in shorter queue-jumping distance than the queue-jumping distance in the aggressive merge with maximum advancement case. If there is no relevant merge downstream or AHDVs cannot reach the optimal target vehicle, AHDVs target the farthest reachable AV by going through the same target selection process as the aggressive merge with maximum advancement case. The aggressive merge with zipper case represents a moderate level of aggressive behaviors in AHDVs.
32

Figure 4. Diagram. Aggressive merge with zipper. (AHDV deep blue vehicles, AV light blue vehicles, and HDV white vehicles) There are three essential computations used to model the targeting behavior--position check, can catch, and merge position check. These three functions are executed every time step to update the target vehicle based on the position and speed changes in AHDVs and their target vehicles.
Position Check To merge into the deceleration lane, the AHDV must decrease to the speed of the vehicle in front of the target AV, as the target AV and its leading vehicle represent the lagging and leading vehicles, respectively, for the gap that will be entered by the AHDV. The objective of position check is to determine if the AHDV has reached the position where it must decide whether to target the next downstream AV or keep the current target and start braking to prepare for the merge. Figure 5. Diagram. Illustration of AHDV's decision point. (AHDV deep blue vehicles, AV light blue vehicles, and HDV white vehicles)
shows an illustration of this decision point position check process. The braking distance is the distance the AHDV will travel to reduce its speed to the merge speed by the time it is one vehicle length downstream of the AV. Travel while Braking (TwB), is the distance that the target vehicle travels while the AHDV travels the braking distance. When the braking distance is equal to or less than the sum of TwB and vehicle length, the position check function returns `true' and the target vehicle ID is sent to the can catch function to determine whether the next potential target vehicle
33

is reachable. If the position check function returns `false', it indicates that the AHDV has not yet reached the decision point position and, thus, continues to travel at its current speed.
Figure 5. Diagram. Illustration of AHDV's decision point. (AHDV deep blue vehicles, AV light blue vehicles, and HDV white vehicles) Can Catch The objective of the can catch function is to determine if the AHDV can reach the front of the target vehicle, to allow for a merge, before the deceleration lane-end point. This is determined by evaluating the current position and speed conditions, and comparing the travel time of the AHDV and target vehicles to the end of the lane. Can catch is applied at every time step to confirm that the current target vehicle may still be reached, allowing for potential changing conditions due to congestion. In addition, can catch is utilized when the AHDV evaluates if it will switch from its current target to the next downstream AV. The travel-time comparison between AHDV and the target vehicle is based on the current position and speed data, as shown in the equations below. The AHDV must be able to reach the lane endpoint before the potential target vehicle. The travel time of the target vehicle can be calculated by dividing the remaining distance until the lane end by the current speed.
34

(1) Calculating the travel time of an AHDV depends on its current position. As shown in figure 6, the remaining distance is divided into two regions.
Figure 6. Diagram. AHDV travel-time calculation. (AHDV deep blue vehicles, AV light blue vehicles, and HDV white vehicles) The distance in red indicates the braking distance from the AHDV's current speed to the target speed (equation 2), with the merge occurring at the end of the deceleration lane. The distance in green indicates the distance that the AHDV needs to travel at its current speed until it starts braking (equation 3). Thus, the AHDV's travel time (equation 4) is the sum of the travel time over the fixed-speed distance (indicated in green in figure 5) and the travel time over the braking distance (indicated in red in figure 5). If the AHDV's travel time is less than the target vehicle's travel time, the can catch function returns `true'.
(2)
(3)
(4)
35

(5) As stated, can catch is executed every time step for the current target AV. If the can catch function returns `false', it indicates the AHDV is no longer able to catch the current target AV vehicle. When this occurs, merge control is released to SUMO, which will execute a non-aggressive merge maneuver into the deceleration lane as soon as possible. However, while SUMO is seeking a merge opportunity, the AHDV continues to search for an AV within 20 ft downstream or approaching from the upstream if the deceleration lane is moving faster than the mainline lane. If an AV is identified, the TraCI logic will be reinitiated. Merge Position This function checks whether an AHDV is within the position to initiate an aggressive merge. Once the AHDV is in position, the merge process initiates. As shown in figure 7, an AHDV executes the merge process if its front bumper is anywhere between the center of the target vehicle and the head of the leading vehicle to the target vehicle (red region in figure 7).
Figure 7. Diagram. Illustration of merge position. (AHDV deep blue vehicles, AV light blue vehicles,
and HDV white vehicles)
36

Target AV Selection Process in Aggressive Merge with Maximum Advancement Case The flowchart of the AHDV's target AV selection process is shown in figure 8 and figure 9. The target selection process is undertaken every time step. The following steps are the general procedure:
1. Vehicle ID list is updated to contain the IDs for all AHDVs currently in the merging zone. Vehicles entering the merging zone are added and vehicles that have merged into the deceleration lane are removed.
2. When an AHDV first arrives at the upstream start of the merging zone, the deceleration lane condition is reviewed. If the deceleration lane is empty, the AHDV changes lane without any further consideration.
3. If the deceleration lane is not empty, the AHDV checks for the presence of any AV.
4. If there is no AV, the AHDV continues to search for any AV upstream and downstream of its current location while allowing SUMO to execute a merge whenever it is possible. This process continues until either SUMO executes the merge or the AHDV finds an AV in the traffic.
5. If there is more than one AV in the deceleration lane, the AHDV initially identifies the nearest downstream AV.
6. Next, using the can catch function, the merge feasibility of the AHDV with the nearest AV is checked. If the AHDV cannot merge in front of the nearest AV, it indicates that there is no AV that the AHDV can catch. The process returns to step 4.
7. If the AHDV can catch the nearest AV, the position of the AHDV is checked (using the position check function) to determine whether the AHDV is ready
37

to search for the next AV downstream. If the AHDV is not in such position, it continues to travel until being checked again in the next time step. 8. If the AHDV is in such position, the next AV downstream is identified and checked for merge feasibility using the can catch function. 9. If the AHDV can catch the next AV in downstream, the target AV is updated. If not, the current target AV is maintained. 10. In every time step, the can catch function is used to determine if the AHDV can still catch the current target AV. 11. If the AHDV can no longer catch the current target AV, the AHDV first searches to check whether the nearest reachable AV is downstream. If there is one, the AHDV updates its target. 12. If there is no reachable downstream AV, the AHDV searches for the nearest reachable AV that is upstream. If there is one, the AHDV updates its target. If there is no such AV, the process returns to step 4. 13. The process continues until all AHDVs have been checked; then the simulation time advances.
Target Selection Process in Aggressive Merge with Zipper Case The flowchart of the target selection process in the aggressive merge with zipper case
is shown in figure 10 and figure 11. The first four steps in the target selection process are the same as the maximum advancement case. The following steps are a divergence from the maximum advancement case at step 5:
5. If there is more than one AV, the AHDV searches for any previous merge. If there is no previous merge, the AHDV finds the nearest AV and follows
38

the same steps in the aggressive merge with maximum advancement case (step 8). 6. If there is a previous merge, the AHDV searches for the follower AV of the previous merge. If there is no follower vehicle, it indicates that the lane behind the merge is empty. Thus, the a AHDV allows SUMO to execute a lane change whenever possible. 7. If the AHDV finds a follower AV to the previous merge, the merge feasibility is checked with the `can catch' function. If the AHDV can catch the follower AV, the follower AV is selected as the target vehicle. 8. If the AHDV cannot catch the follower AV, the AHDV searches for the nearest AV and the same steps are followed as in the aggressive merge with maximum advancement case. However, if the AHDV is ahead of an HDV, the AHDV allows SUMO to merge it in front of the HDV.
39

a

b

c

d

e

f

g

h

Figure 8. Flowchart. Target AV selection process with step numbers in the process description marked (top).

40

a

b

c

d

e

f

g

h

Figure 9. Flowchart. Target AV selection process with step numbers in the process description marked (bottom). 41

a

b

c

d

e

f g

h

i

j

l k

Figure 10. Flowchart. Target selection process in aggressive merge with zipper (top).

42

a

b

c

d

e

f g

h

i

j

k

l

Figure 11. Flowchart. Target selection process in aggressive merge with zipper (bottom). 43

Aggressive Merge Behavior Model: Lane Changing After the targeting process is complete and the AHDV is positioned next to the target AV, the lane change process is initiated. As shown in figure 12, AHDVs merge in front of the target AV as soon as the AHDV's rear bumper crosses the front bumper of the AV, forcing the AV to decelerate to meet its desired spacing. For this aggressive merge, the TraCI moveTo command is utilized. The moveTo command in SUMO manually moves the position of a vehicle by the specified coordinate shift and, critically, it does not require the vehicle to satisfy any gap requirements.
Figure 12. Diagram. Example of aggressive merge. (AHDV deep blue vehicles, AV light blue vehicles) Merging Process in SUMO-controlled Merge In order to assign the cooperative characteristic in AVs and HDVs, SUMO's `lcCooperativeSpeed' parameter is set to 1. Setting this parameter to 1 allows the neighboring vehicles to slow down cooperatively for merging vehicles. When the algorithm requests that SUMO control the AHDV merging process, the neighboring vehicle (an HDV, as an aggressive merge would be undertaken for an AV) starts slowing down cooperatively to create a sufficient gap for AHDVs to merge. However, when TraCI is utilized to implement an aggressive merge, the AVs do not exhibit a cooperative
44

behavior, as they are unaware the AHDV will merge until it begins to encroach into the AV lane. Only upon this encroachment will the AV begin to slow. Thus, a SUMO-controlled merge requires a sufficient gap before a lane change is performed, whereas the aggressive merges (using the moveTo command) are not affected by the gap availability. This results in the SUMO-controlled merge often requiring a longer time period for the merge, possible requiring slowing of the merging vehicle to find a suitable gap to complete. An example of the spacing between the lagging vehicle and the merging vehicle in a SUMO-controlled merge, which requires longer gaps to merge, is shown in figure 13.
Figure 13. Diagram. Example of SUMO-controlled merge. (AHDV deep blue vehicles, AV light blue vehicles)
EXPERIMENTS Four experiments were conducted to study the developed aggressive merging models. The first experiment simulates a platoon of 10 AHDVs performing the aggressive merges, for two levels of congestion on the deceleration lane. The second and third experiments simulate AHDVs spread out in the mixed traffic flow performing the aggressive merges, for two levels of traffic demands. The distinction between the second and third experiments is the level of congestion in the deceleration lane, resulting from changing the signal timing at the ramp end intersection. The fourth experiment evaluates
45

the impact of aggressive merging on capacity. For all four experiments, a base case was also created without any aggressive vehicle behaviors.
Experiment 1: Aggressive Merging with Platooned Arrivals Objectives The objectives of the initial scenario are to: (1) demonstrate the aggressive merging behavior models under two different traffic conditions--uncongested deceleration lane and congested deceleration lane; and (2) visualize the difference in the impacts on the deceleration lane traffic between aggressive merging and SUMO-controlled merging.
Experiment 1 Design A platoon of 10 AHDVs is introduced into the traffic stream on lane A_0, the left-most freeway lane. The platoon vehicles change their lane to the lane adjacent to the deceleration lane as soon as they enter the merging zone. The AHDVs then seek to merge into the deceleration lane, utilizing the targeting and lane-changing behavior as discussed previously. An entry volume of 1,350 vehicles/hour was used on lane A_0 with a 50 percent AV ratio. All vehicles on lane A_0 were exit vehicles. The uncongested deceleration-lane experiment was conducted before the queue started forming on the deceleration lane. The congested deceleration-lane experiment was conducted after a queue formed on the ramp and extended to the deceleration lane. For comparison, the base case introduces an equivalent platoon of 10 AHDVs, although functioning as HDVs, seeking to exit with the merge behavior controlled by SUMO.
46

Results Figure 14 figure 19 show the timespace diagrams (TSDs) for the merge zone, with the AHDV platoon trajectories indicated in red, and the AV and HDV trajectories in blue. AV and HDV travel occur on the deceleration lane, while AHDV travel may occur on the deceleration lane or adjacent mainline lane. Each graph starts at the beginning of the merge zone, at approximately x=5200 ft The ramp gore is at 5900 ft, and the intersection with the cross street is approximately at 7500 ft.
Discussion The impacts of the AHDVs' aggressive merging behaviors can be observed in two ways: (1) the AHDVs' reduced travel times, and (2) the speed changes in the traffic on the target (i.e., deceleration) lane. The AHDVs' reduced travel times are shown by the time steps in which each red line ends. In each congestion scenario the AHDV platoon enters the merge zone at approximately the same time, i.e., at approximately t = 100 seconds for the uncongested scenarios--i.e., base (figure 14), aggressive merge with maximum advancement (figure 15), and aggressive merge with zipper (figure 16)--and t = 960 for the congested scenarios--i.e., base (figure 17), aggressive merge with maximum advancement (figure 18), and aggressive merge with zipper (figure 19). However, in each aggressive merge scenario, the platoon of AHDVs departs from the intersection at the end of the ramp (top of the TSD) earlier than in the base case with the non-aggressive HDV platoon. This is accomplished by the AHDVs queue-jumping (as seen by the crossing of the red and blue trajectories) by driving further downstream on the mainline, then performing aggressive merges near the ramp gore. The impact on the speed of the vehicles behind the merged AHDVs is witnessed by a flattening of the slopes on the
47

vehicle trajectories. In the vicinity of the merge activity (highlighted in yellow) the speeds of the following vehicles are reduced by approximately 17 mph in the uncongested scenario, and in the congested scenario following vehicles are forced by the AHDVs to briefly come to a complete stop to avoid colliding with the merging vehicles. The travel time and speed impacts are more clearly seen in the congested decelerationlane scenario compared to the uncongested scenario. This is due to the spacing between vehicles. Since vehicles were more spread out in the uncongested scenario, the impacts of aggressive merges were muted by the larger headways between the vehicles; in the congested scenario, the impacts of aggressive merges were directly passed along to the following vehicles. The next two experiments investigate the impacts of the aggressive merging behaviors with AHDVs spread throughout the traffic flow.
48

Low Congestion
Figure 14. Plot. Base case timespace diagram in uncongested deceleration-lane scenario. 49

Figure 15. Plot. Aggressive merge with maximum advancement timespace diagram in uncongested deceleration-lane scenario.
50

Figure 16. Plot. Aggressive merge with zipper timespace diagram in uncongested deceleration-lane scenario.
51

High Congestion
Figure 17. Plot. Base case timespace diagram in congested deceleration-lane scenario. 52

Figure 18. Plot. Aggressive merge with maximum advancement timespace diagram in congested deceleration-lane scenario.
53

Figure 19. Plot. Aggressive merge with zipper timespace diagram in congested deceleration-lane scenario.
54

Experiment 2: Aggressive Merging with Random Arrivals Objective While the initial experiment investigated the impact of AHDVs in a platoon, the objective of this experiment is to investigate the impact of aggressive merging behavior under conditions where the AHDVs are distributed throughout the traffic stream.
Experiment Design The roadway layout for this experiment is the same as the previous experiment. Traffic volume is balanced in the mainline lanes entering the merge zone. All exiting vehicles enter the merge zone already positioned in lane A_1 (see figure 2). Thus, all vehicles in the left-most lane A_0 are through vehicles only, while vehicles on lane A_1 consist of both through and exit vehicles. In this experiment, 35 percent of the traffic is assumed to exit; thus, 70 percent of the lane A_1 vehicles were assigned as exit vehicles, consisting of AVs, HDVs, and AHDVs (percentages described subsequently). All exit vehicles except for the AHDVs shift over to the deceleration lane B_2 when they reach the merging zone, at the start of the deceleration lane. The AHDVs continue to travel on lane B_1 and make a lane change to B_2 by either aggressive merge or SUMO-controlled merge, as defined previously.
Two levels of traffic demand were considered in this experiment--high traffic demand (1,200 vehicles/hour/lane) and low traffic demand (600 vehicles/hour/lane). For each traffic-demand level, five different AV ratios (percentage of the total traffic that is AV) and five different AHDV/HDV ratios (including the base case with no AHDVs) of exiting traffic not assigned as AV were considered, as shown in table 5.
55

The distinction between experiment 2 and experiment 3 (presented in the next section) is in the signal timing at the ramp end intersection. In this experiment, 50 seconds of green time and 70 seconds of red time are used for both the lower traffic-demand and higher traffic-demand conditions. This results in no queue spillback to the deceleration lane in the low traffic-demand case, but there was queueing on the deceleration lane in the high traffic-demand case. The base case consists of only AV and HDV. Each scenario has 10 replicate runs. Figure 20 figure 23 show the average travel time of exit vehicles by vehicle type in each scenario. Note that the y-axis scales are different in the two sets of figures to accommodate the wider range of travel times in high traffic-demand conditions. Table 6 and table 7 show the paired t-test results on the travel times of AHDVs compared to the travel times of AVs and HDVs. The `Difference' column shows whether the difference is statistically significant (marked as TRUE if significantly different and FALSE otherwise).
56

Lower Traffic Demand

AV Ratio 10% 20% 30% 40% 50% 10% 20% 30% 40% 50%

Table 5. Vehicle assignment for experiments 2 and 3.

Start Lane Direction AHDV Ratio 0% (Base)
25% 50% 75% 100% 0% (Base) 25% 50% 75% 100% 0% (Base) 25% 50% 75% 100% 0% (Base) 25% 50% 75% 100% 0% (Base) 25% 50% 75% 100% 0% (Base) 25% 50% 75% 100% 0% (Base) 25% 50% 75% 100% 0% (Base) 25% 50% 75% 100% 0% (Base) 25% 50% 75% 100% 0% (Base) 25% 50% 75% 100%

Lane A_0

Through

HDV AV

540

60

540

60

540

60

540

60

540

60

480

120

480

120

480

120

480

120

480

120

420

180

420

180

420

180

420

180

420

180

360

240

360

240

360

240

360

240

360

240

300

300

300

300

300

300

300

300

300

300

1080

120

1080

120

1080

120

1080

120

1080

120

960

240

960

240

960

240

960

240

960

240

840

360

840

360

840

360

840

360

840

360

720

480

720

480

720

480

720

480

720

480

600

600

600

600

600

600

600

600

600

600

Through

HDV AV

162

18

162

18

162

18

162

18

162

18

144

36

144

36

144

36

144

36

144

36

126

54

126

54

126

54

126

54

126

54

108

72

108

72

108

72

108

72

108

72

90

90

90

90

90

90

90

90

90

90

324

36

324

36

324

36

324

36

324

36

288

72

288

72

288

72

288

72

288

72

252

108

252

108

252

108

252

108

252

108

216

144

216

144

216

144

216

144

216

144

180

180

180

180

180

180

180

180

180

180

Lane A_1
HDV 378 284 189 95 0 336 252 168 84 0 294 221 147 74 0 252 189 126 63 0 210 158 105 53 0 756 567 378 189 0 672 504 336 168 0 588 441 294 147 0 504 378 252 126 0 420 315 210 105 0

Exit AV 42 42 42 42 42 84 84 84 84 84 126 126 126 126 126 168 168 168 168 168 210 210 210 210 210 84 84 84 84 84 168 168 168 168 168 252 252 252 252 252 336 336 336 336 336 420 420 420 420 420

AHDV 0 95 189 284 378 0 84 168 252 336 0 74 147 221 294 0 63 126 189 252 0 53 105 158 210 0 189 378 567 756 0 168 336 504 672 0 147 294 441 588 0 126 252 378 504 0 105 210 315 420

Higher Traffic Demand

57

Results
Figure 20. Bar plots. Experiment 2: Average travel time in aggressive merge with maximum advancement scenarios by vehicle type in low traffic-demand condition. 58

Figure 21. Bar plots. Experiment 2: Average travel time in aggressive merge with zipper scenarios by vehicle type in low traffic-demand condition.
59

Figure 22. Bar plots. Experiment 2: Average travel time in aggressive merge with maximum advancement scenarios by vehicle type in high traffic-demand condition.
60

Figure 23. Bar plots. Experiment 2: Average travel time in aggressive merge with zipper by vehicle type in high traffic-demand condition.
61

Table 6. Experiment 2: Paired t-test on travel time in aggressive merge with maximum advancement scenarios.

AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100

Volume = 600 veh/hr/ln, AV10

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.304 FALSE 0.429 FALSE

0.078 FALSE 0.958 FALSE

0.059 FALSE 0.060 FALSE

0.159 FALSE

-

Volume = 600 veh/hr/ln, AV30

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.016

TRUE

0.007

TRUE

0.105 FALSE 0.410 FALSE

0.808 FALSE 0.071 FALSE

0.080 FALSE

-

Volume = 600 veh/hr/ln, AV50

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.476 FALSE 0.409 FALSE

0.003

TRUE

0.010

TRUE

0.764 FALSE 0.387 FALSE

0.624 FALSE

-

Volume = 1200 veh/hr/ln, AV20

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

1.15E-11 TRUE 5.30E-12 TRUE

4.91E-12 TRUE 1.56E-12 TRUE

5.53E-09 TRUE 8.54E-09 TRUE

1.51E-10 TRUE

-

Volume = 1200 veh/hr/ln, AV40

AHDV vs. AV

AHDV vs. HDV

P-Value 4.54E-14 1.18E-11 6.70E-11 1.09E-10

Difference TRUE TRUE TRUE TRUE

P-Value Difference 2.87E-13 TRUE 9.95E-12 TRUE 5.94E-11 TRUE
-

Volume = 600 veh/hr/ln, AV20

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.480

FALSE

0.065

FALSE

0.846

FALSE

0.749

FALSE

0.348

FALSE

0.044

TRUE

0.238

FALSE

-

Volume = 600 veh/hr/ln, AV40

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.043

TRUE

0.069

FALSE

0.282

FALSE

0.461

FALSE

0.457

FALSE

0.132

FALSE

0.157

FALSE

-

Volume = 1200 veh/hr/ln, AV10

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

3.20E-09 TRUE 5.45E-10 TRUE

1.43E-11 TRUE 6.33E-12 TRUE

1.65E-08 TRUE 2.56E-09 TRUE

1.37E-08 TRUE

-

Volume = 1200 veh/hr/ln, AV30

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

9.46E-13 TRUE 4.56E-13 TRUE

6.38E-12 TRUE 2.02E-12 TRUE

1.41E-11 TRUE 2.19E-11 TRUE

2.90E-11 TRUE

-

Volume = 1200 veh/hr/ln, AV50

AHDV vs. AV

AHDV vs. HDV

P-Value 2.35E-12 2.42E-11 1.44E-11 1.31E-09

Difference TRUE TRUE TRUE TRUE

P-Value Difference 3.47E-12 TRUE 1.67E-11 TRUE 3.10E-11 TRUE
-

62

Table 7. Experiment 2: Paired t-test on travel time in aggressive merge with zipper scenarios.

AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100

Volume = 600 veh/hr/ln, AV10

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.304 FALSE 0.429 FALSE

0.078 FALSE 0.958 FALSE

0.05

FALSE 0.060 FALSE

0.15

FALSE

Volume = 600 veh/hr/ln, AV30

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.016

TRUE

0.007

TRUE

0.105 FALSE 0.410 FALSE

0.811 FALSE 0.071 FALSE

0.081 FALSE

Volume = 600 veh/hr/ln, AV50

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.476 FALSE 0.409 FALSE

0.003

TRUE

0.010

TRUE

0.643 FALSE 0.417 FALSE

0.619 FALSE

Volume = 1200 veh/hr/ln, AV20

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference 8.65E-07 TRUE 7.89E-07 TRUE 2.12E-05 TRUE 3.16E-05 TRUE 9.28E-06 TRUE 7.57E-05 TRUE 3.80E-06 TRUE
Volume = 1200 veh/hr/ln, AV40

AHDV vs. AV

AHDV vs. HDV

P-Value 9.95E-11 1.99E-09 7.77E-06 1.15E-05

Difference TRUE TRUE TRUE TRUE

P-Value 1.96E-10 1.19E-09 1.31E-05

Difference
TRUE TRUE TRUE

Volume = 600 veh/hr/ln, AV20

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.480

FALSE

0.065

FALSE

0.846

FALSE

0.749

FALSE

0.350

FALSE

0.044

TRUE

0.253

FALSE

Volume = 600 veh/hr/ln, AV40

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.043

TRUE

0.069

FALSE

0.282

FALSE

0.462

FALSE

0.458

FALSE

0.132

FALSE

0.129

FALSE

Volume = 1200 veh/hr/ln, AV10

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference 1.06E-07 TRUE 3.58E-08 TRUE 3.49E-07 TRUE 6.26E-08 TRUE 8.66E-08 TRUE 6.87E-08 TRUE 1.08E-06 TRUE
Volume = 1200 veh/hr/ln, AV30

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

8.07E-11 TRUE 1.34E-10 TRUE

0.001

TRUE

0.001

TRUE

1.51E-06 TRUE 2.02E-06 TRUE

1.50E-06 TRUE

Volume = 1200 veh/hr/ln, AV50

AHDV vs. AV

AHDV vs. HDV

P-Value 5.05E-13 1.11E-07 1.97E-07 2.58E-07

Difference TRUE TRUE TRUE TRUE

P-Value 1.29E-12 6.41E-08 3.99E-07

Difference TRUE TRUE TRUE

63

Discussion The travel times of exit vehicles in low traffic-demand scenarios, as shown in figure 20 and figure 21, were not significantly impacted by the aggressive merges, with no clear trends being apparent. The travel times between vehicle types or AV penetration rates never differ by more than a few seconds. No queue formed on the deceleration lane in low traffic demand, so most AHDVs performed lane changes to the deceleration lane immediately since the deceleration lane was empty. In a few cases where AHDVs performed aggressive merges, the impacts of the aggressive merges may have been muted because of the existence of large headways between the vehicles. As a result, the paired t-test in low traffic-demand conditions showed that AHDVs had no significant difference in travel time compared to AVs and HDVs in most scenarios. These results are in concurrence with the findings in experiment 1. In figure 15 and figure 16 (uncongested deceleration lane), the impact to the non-AHDV is clearly more muted than the impact seen in figure 18 and figure 19 (congested merge lane). The impacts of the aggressive merges were not passed down to the following vehicles in low traffic demand.
In high traffic-demand conditions, the AHDVs' travel times are significantly lower than the travel times of AVs and HDVs in all scenarios with the aggressive merges, as shown in table 6 and table 7. However, the overall average exit times remained relatively constant, implying that as the AHDVs were able to improve their travel time, the AVs and HDV suffered increased travel time. The HDVs' travel time did not increase to the same extent as the AVs'; however, they did see travel time increases, even though they were never "targeted" by the AHDV. The HDV increase results from HDVs in the deceleration lane following AVs that are targeted.
64

In aggressive merge with maximum advancement cases, it is also seen that the AHDV travel times show (figure 22) an increasing trend at the lower AV ratios (10 and 20 percent). However, the trend reverses when the AV ratios were high (3050 percent). The reason for this behavior is that when there is a smaller number of AVs to target, more AHDVs merged via SUMO-control (i.e., non-aggressive merge), which requires a longer time to complete. As the availability of target AVs increased with higher AV ratios, more AHDVs successfully completed aggressive merges by targeting AVs.
It was also observed during the simulation run that multiple AHDVs targeted the same AV on the deceleration lane, forcing the target AV, as well as the following traffic, to come to a complete stop, similar to the observations for the experiment 1 congested conditions.
However, the AHDV travel times in the aggressive merge with zipper cases in high congestion showed an increasing trend with higher AHDV ratios in all scenarios. It should be noted that there is no direct relationship between the AV ratios and AHDV ratios since the target selection is affected by both AV ratios and the position of the AHDVs. If an AHDV needs to target a following AV to the previous merge and the AHDV is closer to an HDV compared to the target AV, the AHDV will merge to the HDV via SUMO-controlled merge. Therefore, more AHDVs merged via SUMOcontrolled merge as the AHDV ratios increased, which resulted in the increasing trend in travel time in all scenarios.
The bar charts suggested that the aggressive merge with maximum advancement had greater impact on the AV and HDV travel times than the aggressive merge with zipper in high flow rate conditions. In aggressive merge with maximum advancement cases,
65

multiple AHDVs targeted the same AV on the deceleration lane, leading the target AV as well as the following traffic to come to a complete stop. Such behavior was also shown in the preliminary experiment. The blue slopes after the merge in figure 18 became flat, indicating a complete stop due to the merge. However, in the aggressive merge with zipper cases, the target vehicle moved forward after a single AHDV's merge.
The net impacts of AHDVs' aggressive merging behaviors on all exit vehicles (AHDV, AV, and HDV) are shown in the `All Exit Vehicle Travel Time' bar charts in figure 20 figure 23. In low traffic-demand conditions, the net impact was insignificant since most AHDVs changed their lanes to the deceleration lane immediately after reaching the merging zone and the few cases of aggressive merges left little impacts on the target AVs and the following traffic. In high traffic-demand conditions, the net impact was insignificant due to the discussed trade-off effects. The AHDVs' travel-time decreases were achieved at the expense of the travel-time increases of the AVs and HDVs.
Experiment 3: Comparison of Impact of Demand versus Congestion on Travel Times Objective The objective of this experiment is to differentiate between the impact due to increased demand or congestion. Thus, in this experiment the ramp intersection signal times were adjusted such that the low-demand volume resulted in queuing on the deceleration lane and the high-demand volume had no queuing.
Results Similar to experiment 2, figure 24 figure 27 show the average travel time by exit vehicle type in each scenario. While the absolute travel times change due to the signal
66

timing updates, trade-offs are again seen between the AHDVs and the AVs/HDVs. Except, the trade-off between the AHDVs and AVs/HDVs now occurs at the lower volume case, with no obvious trends in the high-volume case. Also similar to experiment 2, the difference in the scenarios with queuing on the deceleration lane (i.e., the low-volume demand in this experiment) are predominately statistically significant, while the scenarios without queueing on the deceleration lane (i.e., high-volume scenarios) are not statistically significant, as shown in table 8 and table 9.
In aggressive merge with maximum advancement cases, the only significant difference in trends was seen in the AHDV delay across AV ratios, which was increasing throughout the low volume in experiment 3. Based on observations of the simulation, it was seen that the change in signal timing resulted in a slower-moving queue, increasing the time required for an AHDV to merge into the deceleration lane, even with aggressive merges. This resulted in more AHDV stacking in the adjacent lane, waiting to merge, and a higher sensitivity to the number of AHDVs.
In the aggressive merge with zipper cases, shown in figure 25, the AHDV travel times are similar in the higher AHDV ratio scenarios due to the same reason discussed above. Since each AHDV can merge in front of a single AV or a single HDV, the AHDV line becomes longer in lower AV ratios, regardless of the AHDV ratios. However, in higher AV ratio scenarios, more AHDVs can perform aggressive merges, which requires less gap compared to SUMO-controlled merges. As a result, the AHDV travel time becomes lower in the lower AHDV ratio with higher AV ratio scenarios.
From this experiment, in context with the previous experiments, it is seen that the presence of queuing (or near- or over-capacity conditions) is a critical factor in the impact
67

of the AHDVs, as this presents significant opportunities for the aggressive behavior. The absolute volume has a lesser impact.
68

Figure 24. Bar plots. Experiment 3: Average travel time in aggressive merge with maximum advancement scenarios by vehicle type in low traffic-demand condition.
69

Figure 25. Bar plots. Experiment 3: Average travel time in aggressive merge with zipper scenarios by vehicle type in low traffic-demand condition.
70

Figure 26. Bar plots. Experiment 3: Average travel time in aggressive merge with maximum advancement scenarios by vehicle type in high traffic-demand condition.
71

Figure 27. Bar plots. Experiment 3: Average travel time in aggressive merge with zipper scenarios by vehicle type in high traffic-demand condition.
72

Table 8. Experiment 3: Paired t-test on travel time in aggressive merge with maximum advancement scenarios.

AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100

Volume = 600 veh/hr/ln, AV10

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

3.52E-13 TRUE 8.61E-13 TRUE

1.00E-07 TRUE 6.09E-07 TRUE

1.34E-07 TRUE 2.33E-06 TRUE

0.001

TRUE

-

Volume = 600 veh/hr/ln, AV30

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

1.88E-11 TRUE 4.08E-11 TRUE

3.22E-10 TRUE 2.80E-10 TRUE

1.34E-06 TRUE 1.02E-05 TRUE

2.74E-05 TRUE

-

Volume = 600 veh/hr/ln, AV50

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

2.71E-11 TRUE 3.24E-11 TRUE

2.69E-11 TRUE 8.97E-11 TRUE

4.77E-09 TRUE 4.39E-08 TRUE

1.24E-06 TRUE

-

Volume = 1200 veh/hr/ln, AV20

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.876 FALSE

0.314 FALSE

0.116

FALSE

0.250 FALSE

0.967 FALSE

0.220 FALSE

0.036

TRUE

-

Volume = 1200 veh/hr/ln, AV40

AHDV vs. AV

AHDV vs. HDV

P-Value 0.850 0.480 0.776 0.568

Difference FALSE FALSE FALSE FALSE

P-Value 0.610 0.292 0.109

Difference FALSE FALSE FALSE
-

Volume = 600 veh/hr/ln, AV20

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

5.87E-12 TRUE 1.83E-10 TRUE

7.20E-07 TRUE 1.73E-06 TRUE

0.001

TRUE

0.004

TRUE

0.080

FALSE

-

Volume = 600 veh/hr/ln, AV40

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

4.19E-09 TRUE 6.29E-10 TRUE

1.00E-05 TRUE 3.94E-06 TRUE

9.43E-07 TRUE 5.80E-07 TRUE

1.18E-05 TRUE

-

Volume = 1200 veh/hr/ln, AV10

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.154

FALSE

0.446

FALSE

0.247

FALSE

0.555

FALSE

0.040

TRUE

0.858

FALSE

0.079

FALSE

-

Volume = 1200 veh/hr/ln, AV30

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.416

FALSE

0.273

FALSE

0.065

FALSE

0.869

FALSE

0.685

FALSE

0.111

FALSE

0.109

FALSE

-

Volume = 1200 veh/hr/ln, AV50

AHDV vs. AV

AHDV vs. HDV

P-Value 0.318 0.415 0.177 0.960

Difference FALSE FALSE FALSE FALSE

P-Value 0.603 0.437 0.931

Difference FALSE FALSE FALSE
-

73

Table 9. Experiment 3: Paired t-test on travel time in aggressive merge with zipper scenarios.

AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100
AHDV Ratio
25 50 75 100

Volume = 600 veh/hr/ln, AV10

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

4.92E-05 TRUE 4.50E-07 TRUE

0.075 FALSE

0.001

TRUE

0.043

TRUE

0.001

TRUE

0.007

TRUE

Volume = 600 veh/hr/ln, AV30

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

3.91E-09 TRUE 2.62E-08 TRUE

0.027

TRUE

0.001

TRUE

0.049

TRUE

0.025

TRUE

0.050 FALSE

Volume = 600 veh/hr/ln, AV50

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

1.37E-07 TRUE 8.04E-08 TRUE

1.91E-07 TRUE 4.81E-08 TRUE

0.007

TRUE

0.354 FALSE

0.221 FALSE

Volume = 1200 veh/hr/ln, AV20

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.876 FALSE

0.314 FALSE

0.208 FALSE 0.1174 FALSE

0.878 FALSE

0.166 FALSE

0.164 FALSE

Volume = 1200 veh/hr/ln, AV40

AHDV vs. AV

AHDV vs. HDV

P-Value 0.855 0.522 0.981 0.097

Difference FALSE FALSE FALSE FALSE

P-Value 0.611 0.305 0.183

Difference FALSE FALSE FALSE

Volume = 600 veh/hr/ln, AV20

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.001

TRUE

0.005

TRUE

0.006

TRUE

0.020

TRUE

0.005

TRUE

0.029

TRUE

0.009

TRUE

Volume = 600 veh/hr/ln, AV40

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

5.33E-08 TRUE 4.52E-09 TRUE

0.001

TRUE

0.001

TRUE

0.224

FALSE

0.321

FALSE

0.283

FALSE

Volume = 1200 veh/hr/ln, AV10

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.154

FALSE

0.446

FALSE

0.283

FALSE

0.634

FALSE

0.047

TRUE

0.790

FALSE

0.251

FALSE

Volume = 1200 veh/hr/ln, AV30

AHDV vs. AV

AHDV vs. HDV

P-Value Difference P-Value Difference

0.416

FALSE

0.273

FALSE

0.064

FALSE

0.838

FALSE

0.705

FALSE

0.159

FALSE

0.558

FALSE

Volume = 1200 veh/hr/ln, AV50

AHDV vs. AV

AHDV vs. HDV

P-Value 0.360 0.381 0.114 0.559

Difference FALSE FALSE FALSE FALSE

P-Value 0.651 0.577 0.941

Difference FALSE FALSE FALSE

74

Experiment 4: Evaluation of Impact of Aggressive Merging on Capacity Experiment Design Experiment 4 examines the impacts of aggressive characteristics in vehicles near a freeway exit on the capacity of the exit lane. Vehicles with two levels of cooperative characteristics were used. The two levels were implemented with the lowest value and the highest value of the SUMO built-in parameter, lcCooperativeSpeed. Varying levels of cooperative behavior were emulated by changing the ratio of noncooperative (lcCooperativeSpeed = 0) and cooperative (lcCooperativeSpeed = 1) vehicles from 0 to 1. When a vehicle with cooperative behavior is available in the traffic, SUMO tries to perform the lane-changing for a vehicle in front of the farthest downstream cooperative vehicle that is reachable. This operation is very similar to the aggressive merging logic developed for AHDV in the previous experiments, albeit with a slightly less degree of control available to the modeler on which vehicles behave as aggressive vehicles than that achieved in experiments 1 to 3 with the explicit modeling of AHDVs.
The experiment is conducted in a two-lane freeway stretched out for 2 miles. A 2,000-ft deceleration lane is added at the end of the 2-mile freeway segment, which is then followed by an exit ramp, as shown in figure 1. All vehicles are seeking to exit the freeway using the ramp. As shown in figure 2, the left-lane (A_0) traffic travels with higher speed than the right-lane (A_1) traffic, creating the opportunity for the left-lane traffic to queue-jump. Both A_1 and A_0 traffic had a mix of vehicles with lcCooperativeSpeed of 1 and 0, indicating the highest level of cooperative characteristic and the lowest level of cooperative characteristic, respectively. The A_1 traffic volume was maintained the same throughout the simulation, whereas the A_0 traffic volumes
75

were increased every 900 seconds. Details on the vehicle assignment are provided in the following section.
The A_0 traffic's route was changed from traveling on the freeway mainline to taking the exit ramp when they arrive at the merging zone indicated in figure 1. Once the route change assignment is complete, the A_0 vehicles begin their lane-changing process. The lane-changing is controlled by SUMO. Some vehicles immediately change their lanes to lane B_1, while others continue to travel on lane B_0 depending on the availability of gaps on lane B_1. For the right lane, all A_1 traffic shifts uninterrupted to lane B_2.
It should be noted that this study assumes the same headways for all vehicle types-- regardless of the level of cooperative characteristics. This eventually becomes a critical factor in explaining how the flow rate was not affected by the aggressive characteristics--but rather, there was a trade-off between the vehicle types that exited on the ramp.
Vehicle Classification Two types of vehicles were considered in the experiments: human-driven vehicles with a SUMO lcCooperativeSpeed value of 0 (referred to as HV0 hereafter), and human-driven vehicles with lcCooperativeSpeed value of 1 (referred to as HV1 hereafter).
The value of 1 in the lcCooperativeSpeed parameter for a particular vehicle allows the vehicle's speed to be adjusted during the merge process. This is especially relevant for receiving-lane vehicles. The vehicle in the receiving lane adjusts it speed and cooperates with the merging vehicle, enabling the merging vehicle to perform the lane change. On the other hand, the value of 0 in the parameter results in no speed adjustment, and
76

consequently no cooperation, to make the merge or to allow the merge. The merge in this case is completely dependent on a pre-existing sufficient gap in front of the vehicle in the receiving lane.
Vehicle Assignment The traffic on lane A_1 had a flow rate of 1,400 vehicles/hour throughout the simulation. Two scenarios were tested with different types of base vehicles in lane A_1. The lane A_1 traffic comprised all HV0 in the first scenario (experiment 4a), while that lane had all HV1 in the second scenario (experiment 4b). The traffic in lane A_0 was a mixture of HV0 and HV1. The ratios of HV0 and HV1 were varied across runs and the volumes were increased every15 minutes (900 seconds) within each run. The vehicle assignment matrix is shown in table 10. Each of the two experiments had five different sub-scenarios with five levels of the HV0/HV1 ratios, generating 10 runs (single trial per scenario).

Table 10. Vehicle assignment on lane A_0 for experiment 4a and experiment 4b.

Time Step (Seconds)
0 900 1800 2700 3600 4500 5400 6300 7200 8100

Total Volume on A_0 (vehicle /
15minute)
0 100 150 200 250 300 350 400 450 500

0% (vehicle / 15-minute)

HV0 HV1

0

0

100 0

150 0

200 0

250 0

300 0

350 0

400 0

450 0

500 0

25% (vehicle / 15-minute)

HV0 HV1

0

0

75 25

113 38

150 50

188 63

225 75

263 88

300 100

338 113

375 125

50% (vehicle / 15-minute)

HV0 HV1

0

0

50 50

75 75

100 100

125 125

150 150

175 175

200 200

225 225

250 250

75% (vehicle / 15-minute)

HV0 HV1

0

0

25 75

38 113

50 150

63 188

75 225

88 263

100 300

113 338

125 375

100% (vehicle / 15-minute)

HV0 HV1

0

0

0 100

0 150

0 200

0 250

0 300

0 350

0 400

0 450

0 500

77

Results Visualization The impacts of aggressive merging were investigated by studying the flow and speed metrics. Speed vs. flow plots, time vs. average speed plots, and time vs. flow plots were used to visualize the changes in response to the increase in demand over time (in 15minute increments). The data for three locations--500 ft before the start of the deceleration lane on lane A_1, the start of the deceleration lane (on lane B_2), and the start of the ramp--were plotted. Time vs. average speed plots and time vs. flow plots were combined into dual-axis plots (see appendix B).
The 15-minute vehicle counts (table 11 table 14) were measured at the start of the simulation where the vehicles entered the simulation and at the start of the ramp. The vehicle counts are also divided into the vehicle types by lane (A_0 HV0, A_0 HV1, A_1 HV0, and A_1 HV1) to measure the trade-off effects on each vehicle type.
Discussion In experiment 4a, figure 62 figure 67 in appendix B show the speed vs. flow plots at various locations across all HV1 to HV0 ratio cases. The 1-minute aggregate count observations were multiplied by 60 to generate the corresponding estimated hourly flow rates. As shown on the plots, the change in HV1 to HV0 ratios on lane A_0 did not lead to significant changes in capacity when all A_1 traffic consisted of HV0 (the condition of experiment 4a). The same headways were specified for cooperative vehicles (HV1) and noncooperative vehicles (HV0) in the simulation. The headways also remained the same before and after performing the merge. Even though the flow on lane B_2 (and thus lane A_1) was interrupted by the merge activity, the overall capacity of the exit lane was not
78

affected significantly since the merging vehicles had the same headways and the headway distribution on the exit lane remained the same. While the average capacity was not affected, the change in HV1 to HV0 ratios (figure 28) did affect the variability and level of fluctuation in the flow, as can be seen by comparing the plots across the five different levels in figure 29 and figure 31. In the 0 percent HV1 case, since A_0 traffic only consisted of HV0, most A_0 vehicles were not able to change their lane to the deceleration lane but started building a queue at the end of lane B_1 (shown in figure 28). Lane changes occurred only when there were gaps between the platoons on lane B_2 caused by stochastic variation in the vehicle insertion.
Figure 28. Diagram. Queue building at the end of lane B_1. In 25100 percent cases, a greater instability in flow was observed (see figure 29 and figure 30). These instabilities occur when an HV1 on lane B_1 (typically near the back of the queue) changed to lane B_2, using a gap caused by stochastic variation in the traffic. Once in lane B_2, the merged HV1 vehicle would allow vehicles waiting in the queue on lane B_1 to merge in front of it due to its cooperative characteristics, essentially clearing a portion of the B_1 queue. However, these instabilities did not last long nor occur frequently since they only occurred when there was a sufficient gap between the vehicle platoons on lane B_2 to allow the initial HV1 to merge. The total ramp volumes were unaffected by the HV1 ratios, as can be seen in figure 31 and table 12.
79

Unlike experiment 4a, experiment 4b (i.e., all vehicles in lane A-1 are HV1) did experience a change in capacity (figure 65 figure 66 in appendix B). At the lower HV1 percentages from lane A_0, a lower ramp capacity (i.e., the number of vehicles that were able to successfully merge and exit) was observed. At the lower HV1 penetration rates the plots in figure 32 figure 33 show that the flow on lane A_1 and B_2 was more frequently interrupted compared to experiment 4a (figure 29 and figure 30) since all A_1 traffic was HV1 and merges occurred freely and without building a long queue on lanes B-1 or B_0. The ability of traffic originating from A_1 to successfully exit the freeway was reduced over time, as seen in table 13, due to the merging vehicle originating from lane A_0 consuming a larger portion of the available capacity, and the merge maneuvers resulting in longer headways. This is reflective of the results in experiments 1, 2, and 3 where the AVs were seen to yield to the more aggressive vehicles. Interestingly, as the percentage of HV1s increased, the ramp capacity increased, reaching a level equivalent to experiment 4a. That is, when most vehicles are either fully cooperative or noncooperative, similar capacities are obtained; however, where a higher percentage of cooperative vehicles are positioned to be targeted by more aggressive vehicles, this aggressive-to-non-aggressive interaction can significantly reduce capacity.
Additionally, the served vehicles in the aggressive-to-non-aggressive interaction tend to be the aggressive vehicle. This is seen through the increasing queue length on lane A_1 as the percentage of noncooperative merging vehicles increased. These findings are congruent with the findings of experiments 2 and 3 where AHDVs benefited in reduced travel time by targeting AVs, while the AVs and the following traffic's travel time
80

increased. The AHDVs' gains were achieved at the expense of AVs and the other following traffic. Such trade-off trends in aggressive vehicles taking advantage of cooperative vehicles are a potential significant issue in freeway control. For instance, heavy trucks are often viewed as a merge target in congested condition. In a scenario where an AV heavy truck on an exit lane is targeted by multiple aggressive drivers, the exit lane flow is likely to be interrupted as seen in experiment 4b. On the other hand, in a scenario where an AV heavy truck is unable to merge into the exit lane due to the uncooperative behaviors by the exitlane vehicles, the adjacent lane flow will be disrupted as demonstrated in experiment 4a. The aggressiveness experienced by AVs will potentially not be limited to AV trucks but may be experienced by any AV. Even where the overall capacity may not be significantly changed, the increased fluctuations in the flow will potentially negatively impact the operations as well as the safety conditions in the upstream traffic.
81

(a)

(b)

(c)

(d)

(e)

Figure 29. Plot. Time vs. average speed and flow plots at 500 ft before the start of the deceleration lane in experiment 4a: (a) 0%, (b) 25%, (c) 50%, (d) 75%, and (e) 100%.

82

(a)

(b)

(c)

(d)

(e)

Figure 30. Plot. Time vs. average speed and flow plots at the start of the deceleration lane in experiment 4a: (a) 0%, (b) 25%, (c) 50%, (d) 75%, and (e) 100%.

83

(a)

(b)

(c)

(d)

(e)

Figure 31. Plot. Time vs. average speed and flow plots at the start of the ramp in experiment 4a: (a) 0%, (b) 25%, (c) 50%, (d) 75%, and (e) 100%.

84

Entry Point

Entry Point

Table 11. Vehicle count by vehicle type at entry point in experiment 4a (vehicles / 15 minutes).

0
A_0 HV0 0 A_0 HV1 0 A_0 Total 0 A_1 HV0 350 A_1 HV1 0 A_1 Total 350
A_0 HV0 0 A_0 HV1 0 A_0 Total 0 A_1 HV0 350 A_1 HV1 0 A_1 Total 350
A_0 HV0 0 A_0 HV1 0 A_0 Total 0 A_1 HV0 350 A_1 HV1 0 A_1 Total 350
A_0 HV0 0 A_0 HV1 0 A_0 Total 0 A_1 HV0 350 A_1 HV1 0 A_1 Total 350
A_0 HV0 0 A_0 HV1 0 A_0 Total 0 A_1 HV0 350 A_1 HV1 0 A_1 Total 350

Time Step

900 1800 2700 3600 4500 5400 6300 7200

0%

100 150 200 250 300 350 326 166

0

0

0

0

0

0

0

0

100 150 200 250 300 350 326 166

350 351 349 350 351 350 350 350

0

0

0

0

0

0

0

0

350 351 349 350 351 350 350 350

25%

75

113 150 188 225 263 300

26

100 151

50

63

75

88

100

9

175 264 200 251 300 351 400

35

350 351 351 351 350 351 351 351

0

0

0

0

0

0

0

0

350 351 351 351 350 351 351 351

50%

50

75

100 125 150 175 126 113

100 150 100 125 150 175 127 112

150 225 200 250 300 350 253 225

350 351 349 350 351 350 350 350

0

0

0

0

0

0

0

0

350 351 349 350 351 350 350 350

75%

25

38

50

63

75

88

75

60

100 151 150 188 225 263 225 178

125 189 200 251 300 351 300 238

350 351 351 351 350 351 351 350

0

0

0

0

0

0

0

0

350 351 351 351 350 351 351 350

100%

0

0

0

0

0

0

0

0

100 150 200 250 300 350 321 158

100 150 200 250 300 350 321 158

350 351 349 350 351 350 350 350

0

0

0

0

0

0

0

0

350 351 349 350 351 350 350 350

8100
180 0 180 350 0 350
184 61 245 305 0 305
81 82 163 350 0 350
17 48 65 350 0 350
0 223 223 350 0 350

Entry Point

Entry Point

Entry Point

85

Ramp

Table 12. Vehicle count by vehicle type at ramp in experiment 4a (vehicles / 15 minutes).

Ramp

0
A_0 HV0 0 A_0 HV1 0 A_0 Total 0 A_1 HV0 286 A_1 HV1 0 A_1 Total 286
A_0 HV0 0 A_0 HV1 0 A_0 Total 0 A_1 HV0 286 A_1 HV1 0 A_1 Total 286
A_0 HV0 0 A_0 HV1 0 A_0 Total 0 A_1 HV0 286 A_1 HV1 0 A_1 Total 286
A_0 HV0 0 A_0 HV1 0 A_0 Total 0 A_1 HV0 286 A_1 HV1 0 A_1 Total 286
A_0 HV0 0 A_0 HV1 0 A_0 Total 0 A_1 HV0 286 A_1 HV1 0 A_1 Total 286

Time Step

900 1800 2700 3600 4500 5400 6300 7200 8100

0%

82

130 166 179 183 190 157 159 209

0

0

0

0

0

0

0

0

0

82

130 166 179 183 190 157 159 209

350 349 343 345 341 339 366 365 321

0

0

0

0

0

0

0

0

0

350 349 343 345 341 339 366 365 321

25%

59

110 136 164 107 237

0

179 282

79

143

67

45

41

56

0

64

109

138 253 203 209 148 293

0

243 391

341 341 326 294 361 197 516 243

83

0

0

0

0

0

0

0

0

0

341 341 326 294 361 197 516 243

83

50%

39

72

80

83

128

50

105 107

33

78

144

87

88

102

52

126

81

59

117 216 167 171 230 102 231 188

92

341 351 349 344 279 403 277 316 424

0

0

0

0

0

0

0

0

0

341 351 349 344 279 403 277 316 424

75%

20

37

47

41

44

48

64

18

50

79

146 146 114 130 143 188

46

150

99

183 193 155 174 191 252

64

200

341 345 333 358 333 327 259 458 323

0

0

0

0

0

0

0

0

0

341 345 333 358 333 327 259 458 323

100%

0

0

0

0

0

0

0

0

0

80

141 197 192 257 115 132 254 195

80

141 197 192 257 115 132 254 195

350 347 323 335 271 408 389 272 334

0

0

0

0

0

0

0

0

0

350 347 323 335 271 408 389 272 334

Ramp

Ramp

Ramp

86

(a)

(b)

(c)

(d)

(e)

Figure 32. Plot. Time vs. average speed and flow plots at 500 ft before the start of the deceleration lane in experiment 4b: (a) 0%, (b) 25%, (c) 50%, (d) 75%, and (e) 100%.

87

(a)

(b)

(c)

(d)

(e)

Figure 33. Plot. Time vs. average speed and flow plots at the start of the deceleration lane in experiment 4b: (a) 0%, (b) 25%, (c) 50%, (d) 75%, (e) 100%.

88

(a)

(b)

(c)

(d)

(e)

Figure 34. Plot. Time vs. average speed and flow plots at the start of ramp in experiment 4b: (a) 0%, (b) 25%, (c) 50%, (d) 75%, and (e) 100%.

89

Table 13. Vehicle count by vehicle type at entry point in experiment 4b (vehicles / 15 minutes).

Entry Point

Entry Point

Entry Point

Time Step

0 900 1800 2700 3600 4500 5400 6300 7200 8100

0%

A_0 HV0 0 100 150 200 250 300 350 401 450 394

A_0 HV1 0 0 0 0 0 0 0 0 0 0

A_0 Total 0 100 150 200 250 300 350 401 450 394

A_1 HV0 0 0 0 0 0 0 0 0 0 0

A_1 HV1 350 350 351 349 350 242 95 87 93 87

A_1 Total 350 350 351 349 350 242 95 87 93 87

25%

A_0 HV0 0 75 113 150 188 225 263 300 338 322

A_0 HV1 0 100 151 50 63 75 88 100 113 108

A_0 Total 0 175 264 200 251 300 351 400 451 430

A_1 HV0 0 0 0 0 0 0 0 0 0 0

A_1 HV1 350 350 351 351 351 350 201 71 39 76

A_1 Total 350 350 351 351 351 350 201 71 39 76

50%

A_0 HV0 0 50 75 100 125 150 175 200 225 219

A_0 HV1 0 100 150 100 125 150 175 200 225 219

A_0 Total 0 150 225 200 250 300 350 400 450 438

A_1 HV0 0 0 0 0 0 0 0 0 0 0

A_1 HV1 350 350 351 349 350 351 349 150 131 97

A_1 Total 350 350 351 349 350 351 349 150 131 97

75%

A_0 HV0 0 25 38 50 63 75 88 100 113 125

A_0 HV1 0 100 151 150 188 225 263 300 338 374

A_0 Total 0 125 189 200 251 300 351 400 451 499

A_1 HV0 0 0 0 0 0 0 0 0 0 0

A_1 HV1 350 350 351 351 351 350 351 205 155 138

A_1 Total 350 350 351 351 351 350 351 205 155 138

100%

A_0 HV0 0 0 0 0 0 0 0 0 0 0

A_0 HV1 0 100 150 200 250 300 350 401 450 499

A_0 Total 0 100 150 200 250 300 350 401 450 499

A_1 HV0 0 0 0

0

0

0

0

0

0

0

A_1 HV1 350 350 351 349 350 351 350 226 137 146

A_1 Total 350 350 351 349 350 351 350 226 137 146

Entry Point

Entry Point

90

Ramp

Ramp

Table 14. Vehicle count by vehicle type at ramp in experiment 4b (vehicles / 15 minutes).

Time Step

0 900 1800 2700 3600 4500 5400 6300 7200 8100

0%

A_0 HV0 0 86 143 189 232 281 309 312 312 309

A_0 HV1 0 0 0 0 0 0 0 0 0 0

A_0 Total 0 86 143 189 232 281 309 312 312 309

A_1 HV0 0 0 0 0 0 0 0 0 0 0

A_1 HV1 286 345 343 264 119 95 86 92 86 90

A_1 Total 286 345 343 264 119 95 86 92 86 90

25%

A_0 HV0 0 65 106 146 180 216 242 262 248 270

A_0 HV1 0 86 142 67 59 70 72 73 95 87

A_0 Total 0 151 248 213 239 286 314 335 343 357

A_1 HV0 0 0 0 0 0 0 0 0 0 0

A_1 HV1 286 336 357 328 240 122 84 55 69 61

A_1 Total 286 336 357 328 240 122 84 55 69 61

50%

A_0 HV0 0 43 71 95 122 145 164 171 184 194

A_0 HV1 0 86 141 109 119 144 160 169 165 171

A_0 Total 0 129 212 204 241 289 324 340 349 365

A_1 HV0 0 0 0 0 0 0 0 0 0 0

A_1 HV1 286 335 358 333 287 220 157 139 107 77

A_1 Total 286 335 358 333 287 220 157 139 107 77

75%

A_0 HV0 0 22 35 49 60 73 84 90 100 98

A_0 HV1 0 86 142 149 181 222 246 265 292 296

A_0 Total 0 108 177 198 241 295 330 355 392 394

A_1 HV0 0 0 0 0 0 0 0 0 0 0

A_1 HV1 286 336 359 333 293 238 187 177 142 125

A_1 Total 286 336 359 333 293 238 187 177 142 125

100%

A_0 HV0 0 0 0 0 0 0 0 0 0 0

A_0 HV1 0 85 144 191 240 287 335 366 407 410

A_0 Total 0 85 144 191 240 287 335 366 407 410

A_1 HV0 0 0 0

0

0

0

0

0

0

0

A_1 HV1 286 346 349 329 296 247 194 168 130 123

A_1 Total 286 346 349 329 296 247 194 168 130 123

Ramp

Ramp

Ramp

91

Summary This chapter models aggressive merging behaviors in human drivers toward AVs in a mixed traffic environment. The existing literature review suggests that the general outlook on autonomous vehicles is optimistic in that most studies anticipate enhanced roadway performance and safety in a mixed traffic environment. However, these studies had a common assumption--autonomous vehicles and human roadway users will have cooperative interactions. This study asks the question of `what happens if the interactions are not always cooperative between autonomous vehicles and human drivers?'.
Experiments 1 through 3 showed that the presence of human drivers' aggressive merging behaviors had adverse effects on AVs and HDVs. The adverse effects had more significance in high congestion, when there is a queue in the deceleration lane. The impacts of AHDVs' aggressive merges were muted by the larger headways between vehicles in low congestion when there is no queue in the deceleration lane. Based on the experiment 2 and experiment 3 results, AHDVs had a higher travel-time gain with higher level of aggressive behaviors, which in return had greater adverse effects on the AVs' and the HDVs' travel times. However, AHDVs had a greater travel-time reduction with higher AV ratios when the traffic on the deceleration lane was moving relatively quicker. When the traffic on the deceleration lane was not moving quickly, AHDVs ended up blocking the other AHDVs from performing the aggressive merges regardless of the AV ratios.
Experiment 4 took a closer look at the impact of cooperative behaviorinduced aggressive merges on capacity. It was seen that when most vehicles are either fully cooperative or noncooperative similar capacities are obtained; however, where a higher
92

percentage of cooperative vehicles are positioned to be targeted by more aggressive vehicles, this aggressive-to-non-aggressive interaction can significantly reduce travel time. In addition, it was seen, similar to experiments 1 through 3, that AHDV gains were achieved at the expense of AVs. Finally, even in those scenarios where the overall capacity was not significantly changed in response to the variation of the percentage of cooperative vehicles in the traffic, increased fluctuations in the flow may potentially negatively impact operations as well as the safety conditions in the upstream traffic. The findings of this study suggest that despite the general beliefs in the benefits of autonomous vehicles, there may be adverse impacts on the non-aggressive vehicle travel times in the presence of human drivers' aggressive merging behaviors in a mixed traffic environment, especially in congested conditions. Thus, when the potential benefits of the AV are most needed, i.e., at or near capacity, it is possible that human interaction may negate many of the potential savings. While there are certainly limitations to the study, one of the most noteworthy limitations may be a lack of validation. As the interaction between AVs and human-driven vehicles is rare--and some may argue non-existent or at least still "novel"--it is impossible to validate the behavioral assumptions made. However, this same limitation exists for all mixed-fleet studies. It is the goal of this effort to provide a meaningful data point to the range of potential behavioral, and subsequently operational, outcomes.
93

CHAPTER 5. DATA COLLECTION FOR DRIVER BEHAVIOR
DATA COLLECTION PLAN The initial objective of the data collection task was to obtain trajectory data for drivers performing aggressive merges and use these data to finetune the aggressive merge model. However, several physical site-specific limitations prevented the collection of these data. The Georgia Department of Transportation's (GDOT) permanent cameras on roadside poles did not provide a view that could be used for trajectory data extraction. These views suffered from excessive occlusion of the vehicles in the lanes away from the camera. Drone data collection was contemplated as an alternative. However, such efforts were thwarted by the restrictions on the airspace due to nearby airports and helipads and also the lack of cooperation from nearby business owners. The data collection therefore was focused on supplementing the effort on studying the impact of headways (which are affected by the aggressive behavior as well as other automated vehicle behavior, such as platooning) on capacity, which will be presented in chapter 6. The data collection effort measured the saturation headways at two typical intersections in Georgia during the PM peak period on weekdays.
DATA COLLECTION METHOD Site Selection for Drone Video Data Collection For collecting data, two sites were chosen on mainline Peachtree Industrial Boulevard (PIB). The sites were chosen in a way that the drone can be docked within the GDOT right of way and away from any no-fly zones (figure 35). At the two intersections shown
94

in figure 35, the objective was to obtain trajectories of the vehicles departing from a standing queue when the signal indication changed to green. To ensure sufficient demand to achieve these conditions, the data collection was performed during the PM-peak hours between 3 PM and 6 PM on weekdays.
The data were collected at an elevation of approximately 350400 ft above the ground. Hence, no interactions with the overhead wires were expected. However, special considerations had to be made to maneuver around the wires while taking off and landing the equipment. In the event of a breeze, the equipment would get offset from the data collection spot to balance the effect of the wind and hence the equipment had to be readjusted accordingly from time to time in reaction to the automatic mid-air adjustments. The equipment was not operated on a day with any heavy rain or thunderstorm forecast.
Drone Video Data Processing Using the DataFromSky Viewer The field-collected drone video data were processed to extract vehicle trajectories using the services of an external vendor, DataFromSky (DataFromSky 2021a) via their online service portal. The platform uses artificial intelligence (AI) and computer vision to detect vehicle movements and produce annotated vehicle trajectories. The processed data are returned from the platform in the form of a data package called a "tracking log" with a file extension, ".tlgx". To extract vehicle trajectories and measure other traffic-flow characteristics, tracking logs are further processed in the DataFromSky Viewer software (DataFromSky 2021). Figure 36 shows a sample of the annotated vehicle trajectories for one of the intersections, loaded from a tracking log in DataFromSky Viewer.
95

(a)
(b) Figure 35. Maps. Sites chosen for data collection: (a) at the intersection of PIB@ North Berkeley Lake Road, (b) at the intersection of PIB@ Medlock Bridge Road.
(red lines = GDOT right-of-way boundaries, X = docking station for drone) Source: Google Maps
96

Figure 36. Map. Annotated vehicle trajectories in DataFromSky Viewer. Source: Google Maps
After loading a tracking log in DataFromSky Viewer, the post-processing of the data is performed using the following three steps:
1. Manual georeferencing. 2. Manual annotation configuration. 3. Exporting analysis data.
Manual Georeferencing Georeferencing ensures that the video footage is properly mapped, oriented, and scaled to allow accurate calculation of trajectory data, including position, speed, and acceleration. A minimum of three points in the footage scene are assigned coordinates extracted from Google Maps. If acceptable positioning accuracy is achieved, the points are shown in green with precision indication in the DataFromSky Viewer, as illustrated in figure 37; otherwise, the points are flagged in red.
97

Figure 37. Map. Manual georeferencing in DataFromSky Viewer. Source: Google Maps
Manual Annotation Configuration This step involves inserting data collection points. For this study, data collection points were gates positioned at stop lines for headway measurements. Figure 38 shows two gates labeled as EB_Ln1 and EB_Ln2 for the two through-movement lanes. When a vehicle crosses a gate, data are collected, including the vehicle type, time of exit, and speed.
98

Figure 38. Map. Inserting gates at stop line.Source: Google Maps Exporting Analysis Data The last step involves exporting the analysis data to a comma-separated values (CSV) file for further analysis using other methods/tools as needed by the research study. As shown in figure 39, the options include exporting entire trajectories and exporting gate-crossing events.
99

Figure 39. Screenshot. Exporting analysis data from DataFromSky Viewer. OBSERVATIONS/RESULTS The output of the post-processing analysis of the trajectories was the individual vehicle headways. A deconvolution analysis was performed in Python using the GaussianMixture function in the scikit-learn module (INRIA 2021, NumFOCUS 2021, NumPy 2021, Xu 2021) to separate out the headways of vehicles that relate to the saturation flow from the other headways. The average saturation headways for the through movements were found to be in the range of 1.84 to 2.28 in the different lanes at the different intersection approaches. The average saturation headways for the protected left turns were in the range of 1.89 to 2.33. The detailed lane-by-lane results are presented in appendix A.
100

CHAPTER 6. SIMPLIFIED CAPACITY ANALYSIS TOOL
INTRODUCTION This chapter presents a Simplified Capacity Analysis Tool for exploring the potential impact of various levels of CAV market penetration on signalized intersection capacity. SCAT is an Excel-based tool that provides through and left-turn movement capacity estimates for user-selected phase timings. While numerous CAV development efforts are underway, with varying degrees of success, there is no accepted representative CAV technology nor is there a generally accepted (or governmentally required) set of CAV behavioral characteristics for vehicles that may ultimately be deployed on the public roadways. As such, it is not possible to develop a single, authoritative estimate of the impact of CAVs on capacity. Thus, SCAT draws on findings from the literature, as well as a project-based simulation, presenting 10 different potential CAV impact scenarios. The analyst may utilize SCAT to explore a range of potential futures and understand the sensitivity of current intersection, as well as future designs, to potential CAV operating characteristics.
CAV SATURATION FLOW OVERVIEW The following section discusses the CAV saturation flow estimates included in SCAT. While the saturation flow modeling approaches in the literature differ widely, there are several overarching vehicle behavioral components covered by each. The key components of most models (generally microscopic) are their approach to car following, platooning, and lane changing.
101

Car following refers to the behavior of a following vehicle behind a lead vehicle, within a lane. The output of a car-following model is the following vehicle's acceleration, that is, should the following vehicle accelerate, decelerate, or maintain its current speed. There are enumerable approaches to developing car-following models, but commonly they consist of some function of a desired or minimum time gap, the spacing between vehicles, speed, and desired or maximum accelerations and decelerations. However, other parameters or traffic-condition characteristics may also be part of a car-following algorithm.
Platooning is arguably a special case of car following. However, platooning vehicles tend to travel in lock-step, that is, the reaction time between vehicles is practically (if not actually) reduced to zero. In addition, headways may be significantly lower than the minimum found in most car-following models. To implement platooning, it is assumed that the following vehicle is either connected (i.e., in communication) with the lead vehicle or has sufficient sensors to allow for a reaction time nearing 0 seconds. Many CAV models will impose limits on the length of platoons. This may be either due to assumed technology limits or as a safety constraint where breaks in platoons are deemed necessary to allow for interaction with human-driven vehicles in a mixed-fleet environment.
Lane changing, while influenced by car following and platooning, is the process by which a vehicle decides whether and if to implement a lane change. Commonly, lane changing is considered as discretionary (e.g., a vehicle changes lanes to advance its position in the traffic stream) or mandatory (e.g., a lane change is required to enter a freeway from an on-ramp). Lane change models may also incorporate behavioral
102

changes, such as cooperative braking, by the vehicle in the destination lane. Lanechanging models are critical in multilane facilities and often a determining factor in the capacity of bottlenecks, weaving areas, merges, diverges, etc.
Within the literature, freeway-based CAV models are significantly more common than arterial models. Given the current lack of arterial models, the majority of the models included in SCAT are developed and calibrated for freeway scenarios. However, SCAT is focused on the capacity of vehicles departing from an approach, ignoring the effects of lane changing. Thus, the freeway models utilized were for developed basic freeway segment saturation flows rather than weaving areas, limiting the influence of the model's CAV lane-changing behaviors. The model impacts are focused on the changes in carfollowing and platooning related to the market penetration of CAVs. While arterialspecific models would be preferred, the referenced models should give a sense of the variation in capacity that may be witnessed for departing vehicles at a signalized intersection.
However, a direct application of any one of these models to a specific intersection would likely provide findings with minimum reliability given the significant uncertainty in the characteristics and deployment timeline of CAV technology. Rather, a more productive use of SCAT (or direct reference of the literature) is to explore the sensitivity of projected traffic demands and designs across the range of future predictions. These models provide a sense of the various assumed CAV headways, platooning, and other characteristics. Designers and policy makers can also consider the impacts of various timelines for increasing market penetration rates. Ultimately, testing a design against multiple potential
103

CAV futures aids in understanding its robustness in the face of significant uncertainty and the potential and cost to "future-proof" designs.
SCAT SATURATION FLOW MODELS Prior to describing the use of SCAT, the models included are briefly discussed. While the term CAV is utilized as a broad descriptor in this chapter, it will be seen that the selection of models includes a range of vehicle types: connected vehicle (CV), AV, CAV, and cooperative adaptive cruise control (CACC). It will also be seen that adaptive cruise control (ACC) or CACC models are often utilized for the car-following behavior in a CAV model. The discussion provided for each model will utilize the term from the given reference.
Capacity Adjustment Factors for Connected and Automated Vehicles in the Highway Capacity Manual, Draft Phase 1 Report, Pooled Fund Study (Schroeder et al. 2021) This project sought to develop CAV capacity adjustment factors for use in the Highway Capacity Manual (HCM). The effort utilized an agent-based simulation modeling approach implemented in VISSIM, developing capacity adjustment factors for freeway segments (i.e., basic, merge, diverge, and weaving), signalized intersections (i.e., through movements and protected and permitted left-turn movements), two-way stop-controlled intersections, and roundabouts (i.e., yield control entry). The CAVs modeled were assumed to be SAE1 Level 4 or 5, that is, for the facilities being modeled the vehicle was
1 Society of Automotive Engineers Levels of Driving AutomationTM are defined in SAE J3016 from Level 0 (no driving automation) to Level 5 (full driving automation).
104

assumed to operate with no human intervention. Capacity adjustment factors were developed over CAV penetration rates from 0 to 100 percent.
As with all CAV modeling efforts, limited field data are available and CAV technology is in a continual state of flux. In Schroeder et al. (2021), it is highlighted that a key objective was the development of a minimum achievable gap. Developing such a gap required a number of assumptions regarding "CAV capability, Human-driven vehicle capability, Platooning behavior, Left-turn behavior, Inter-platoon gaps, Intra Platoon gaps, Maximum platoon size, System reliability, and Traffic Stream Composition." Assumptions were based on a review of the literature and best judgment. The number of required assumptions should not be taken as a criticism of this effort; it is simply a reflection of the current state of uncertainty in the ultimate characteristics of CAVs and a source of the differences seen in the capacity impact estimates throughout the literature.
A signalized intersection of a four-lane roadway (40-mph speed limit) with a two-lane roadway (30-mph speed limit), with all approaches having a left-turn bay, was utilized as the base model. A 100-second cycle was utilized with volume demands set to approximate a 0.7 volume-to-capacity ratio. The human-driven vehicles were modeled using Wiedemann 74 driving behavior, with parameters adjusted to match the base saturation flow provided by the HCM. The CAV car-following model is based on a Cooperative Adaptive Cruise Control algorithm developed by Milanes and Shladover (2014). The VISSIM application programming interface (API) is used to implement CAV-based platoon and lane-changing behavior. Ideal conditions are assumed, such as, "no interaction with non-motorized road users, no adverse weather impacts, and a facility without driveways or access points impacting saturation flow rates." This effort also
105

found no significant impact to startup and clearance lost times based on the CAV penetration rate. As one of the few studies to directly consider lost time, the assumption of no-impact is applied to all models included in SCAT.
Included within SCAT are the Schroeder et al. developed capacity adjustment factors for the through movement and protected and permitted left turns. The Schroeder et al. (2021) document also included development of saturation flow rate adjustments for permitted left turns. However, these are not included within SCAT as the adjustment factors are specific to the intersection signal timing and left-turn movement opposing volume, and thus not generally applicable. However, for a given volume set and signal timing, if desired, a SCAT user may update the saturation flows on the SCAT Saturation Flow Adjustment Worksheet, using the factors from Schroeder et al. and a base saturation flow (i.e., 0 percent AVs) calculated using the HCM for the given conditions.
"Modeling Impacts of Cooperative Adaptive Cruise Control on Mixed Traffic Flow in Multi-lane Freeway Facilities" (Liu et al. Error! Reference source not found.) The effort by Liu et al. (2018) models CACC vehicles on freeway facilities. This effort focuses on the "disengagement of CACC strings"; that is, the forming and releasing of platoons of CACC vehicles in a mixed (human-driven and CACC) vehicle environment. Liu et al. considers managed-lane scenarios as well as the implementation of vehicle awareness devices (VADs), which enable a manually driven vehicle to be a CACC platoon leader. The values utilized in SCAT are based on the homogenous freeway segment results found in Liu et al. as this provides the closest approximation for the departure from a signal (i.e., not incorporating significant lane changing or weaving).
106

However, while not incorporated into SCAT, Liu et al. (2018) also include significant effort related to bottleneck behavior at ramp junctions.
Similar to Schroeder et al. (2021), the CACC car-following is based on Milanes and Shladover (2014). In addition, the "NGSIM oversaturated flow human driver model" of Yeo et al. (2008) is utilized. Liu et al. implement a logic allowing a CACC vehicle to join a platoon of existing CACC vehicles, utilizing a reduced headway and thus higher flow rates. Where a platoon is at the maximum-allowed platoon length, the next CACC vehicle will initiate a new platoon, becoming a platoon leader. As part of the effort, a managed lane limited to CACC is considered. Finally, several updates are proposed to the lanechanging rules. Lastly, the 0 percent CACC model is calibrated to field conditions while CACC behavior is based on the literature and best judgment.
"Autonomous and Connected Cars: HCM Estimates for Freeways with Various Market Penetration Rates" (Shi and Prevedouros 2016) This effort considers the impact of driverless vehicles on level of service (LOS) as measured in the Highway Capacity Manual, with a concentration on freeway conditions. To determine the impact on LOS, Shi and Prevedouros (2016) focus on the driverless vehicle car-following headway and penetration rate. For the traffic stream, Shi and Prevedouros utilize a weighted average of the car-following headways for human-driven and driverless vehicles. A driverless vehicle headway of 0.5 second is assumed. Critically, platoon size is not limited, which is a constraint in many other efforts intended to aid the ability of human-driven vehicles to successfully operate in a facility with a high percentage of driverless vehicles. Thus, as the penetration rate approaches 100 percent, the saturation flow rate approaches 7,200 vehicles per hour per lane. The assumptions of
107

Shi and Prevedouros result in significantly higher capacities than any of the other literature included in this effort.
"Enhanced Intelligent Driver Model to Access the Impact of Driving Strategies on Traffic Capacity" (Kesting et al. 2010) The intelligent driver model is a commonly implemented and enhanced model for ACC as well as the car-following component of CAV models. The enhanced IDM utilized in SCAT provides an advancement over the original IDM model by Kesting et al. (2010). The IDM seeks to provide "controllable stability properties" with "smooth transitions between acceleration and deceleration behavior" based on six parameters: desired speed, free acceleration, desired time gap, jam distance, maximum acceleration, and desired deceleration. The IDM provides a continuous function that combines free-road driving and a deceleration model to maintain a desired safety gap. The enhanced IDM improves upon the original model by addressing instability that could be introduced by certain lane-changing behavior. A constant acceleration heuristic is introduced to address overreaction in braking that may occur in the original IDM. While a number of scenarios are considered within the enhanced IDM paper, SCAT integrates the results for a freeway segment outflow from a traffic jam, as this is most analogous to an intersection approach departure. A critical caveat to these results is that Kesting et al. provides capacities only for ACC penetration rates of 0 to 50 percent. Thus, the results in SCAT should not be applied for penetrations greater than 50 percent. Additionally, results in SCAT are given for a default set of traffic conditions (0 percent trucks) and driving behaviors (safety time gap, maximum acceleration, and comfortable acceleration). While not dramatically different, estimated capacities given differing driving behavior assumptions were seen in
108

Kesting et al. (2010) to vary by up to several hundred vehicles per hour as the penetration rate increased based on the selected parameter values.
"A Mixed Traffic Capacity Analysis and Lane Management Model for Connected and Automated Vehicles: A Markov Chain Method" (Ghiasi et al. 2017) Ghiasia et al. (2017) provide an analytical approach for determining the capacity of a highway segment at various CAV market penetration levels. Utilizing a Markov chain approach (i.e., a stochastic modeling approach where the likelihood of the next event is dependent on the previous event), Ghiasia et al. model the spatial headway distributions of the traffic stream. A key element of the model is reflecting the various leaderfollower pairings (i.e., CAVCAV, CAVHuman Driven [HV], HVCAV, and HVHV) in their stochastic model. However, as with all other efforts reported, this effort relies on a set of assumed distributions for these leaderfollower headway pairings, particularly with CAVs. This effort also includes platooning intensity, a measure of the likelihood of vehicles platooning. Platooning intensity allows the model to account for differing platooning strategies, for instance, CAVs seeking other CAVs to create platoons versus platooning opportunities based on a random ordering of vehicles in the traffic stream. (All other models discussed assume platooning opportunities based on random ordering of vehicles.) Ghiasi et al. (2017) is one of the limited number of efforts that demonstrates that increasing capacity with increasing CAV penetration is not guaranteed and that for a given set of "conservative CAV technology scenarios" capacity may decrease.
VISSIM Simulation The final model included in the analysis is based on a VISSIM simulation completed as part of the current study. The model utilizes results from the CoEXist project (CoEXist
109

2021b). The CoEXist project was a European effort to "strengthen the capabilities of urban road authorities for the planning and integration of connected and automated vehicles on their networks." (CoEXist 2021a). As part of the CoEXist project, PTV Group developed for VISSIM a series of new features and parameters set for the modeling of CAVs (Sukennik 2018, Sukennik and Kautzsch 2018).
PTV Group developed three AV models: AV Cautious, AV Normal, and AV Aggressive. For each of these models, a set of Wiedemann 99 CC0 through CC9 parameters were calibrated for CAVs. CC0 through CC9 are driving-behavior parameters of the Weidemann 99 car-following model; interested readers are directed to the final report of GDOT Research Project 18-33, VISSIM 11 Simulation Guidance, for a detailed discussion of each parameter and parameter calibration (Hunter 2021). In addition, parameters were developed for the Wiedemann 74 model, which is generally utilized for arterial operations; however, robust calibration was not undertaken for these parameters and they are not yet recommended for use. In addition, recommendations for the necessary and free lane-change CAV parameter sets were generated, including characteristics such as maximum and accepted deceleration, inclusion of advanced merging and cooperative lane change, and safety distance factor, minimum headway, and maximum cooperation for braking. Updates to driver behaviors at signals were also defined (i.e., behavior at amber, behavior at red, reaction time distribution, reduced safety distance factor, reduced safety start upstream of the stop line, and reduced safety end upstream of stop line.) PTV Group has introduced the ability for vehicle classspecific platooning, enabling the modeling of CAVs at close spacings. Importantly, maximum
110

platoon lengths may be set, with platoon splitting where the number of CAVs in a row exceeds the platoon limit.
For the VISSIM simulation-based saturation flows given in SCAT, a single-lane approach of an intersection was modeled. A 100-second cycle was utilized with a 30-second phase on the subject approach. Demand was set to ensure a constant standing queue. Saturation flow was calculated by measuring the departure headway of the fourth through twelfth vehicle on the approach, each cycle. For the saturation flows reported in SCAT, the Weidemann 99 AV normal settings were utilized with a maximum platoon length of seven vehicles. AV market penetration rates were modeled from 0 to 100 percent, in 10 percent increments. Ten replications were completed for each penetration rate. Finally, to better represent Georgia conditions, the base model (0 percent AVs) headways were calibrated utilizing the data collection at Peachtree Industrial Blvd and Medlock Bridge Rd, as discussed in Chapter 5 of the report. While not provided in SCAT, model runs were also completed using the AV Aggressive setting. However, the saturation values were only slightly higher than the AV Normal. This is likely due to the single-lane approach eliminating any impact of aggressive merging and utilizing the same platoon length.
INSTRUCTIONS FOR SIMPLIFIED CAPACITY ANALYSIS TOOL The use of SCAT is intended to be straightforward. SCAT is set to provide the capacity of each phase at a signalized intersection based on 10 different CAV models, at penetration rates from 0 to 100 percent. A simple eight-phase dual-ring control scheme is assumed, with protected-only lefts. The analyst provides the phase length and number of
111

lanes per movement, and yellow, red clear, and lost time. Currently the model does not incorporate permitted lefts, shared through plus left-turn lanes, or right-turn-on-red. For each analysis all lanes are assigned the same per-lane saturation flow. To explore different saturation values, it is necessary to run the analysis for each CAV model assumption separately. Finally, multi-lane analysis assumes a linear increase in capacity, with no degradation in service due to lane changes, unbalanced lane flows, etc. That is, the capacity for two lanes is taken to be double that of one lane, the capacity for three lanes is triple one lane, etc. The capacity calculation utilized is shown in equation 6:
(6) The saturation flow is based on the literature or simulation results, and the phase lengths, lost time, and number of lanes per movement are provided by the analyst.
112

SCAT has three analysis sections:
1. Individual Scenario Analysis allows an analyst to explore the impact on phase capacity of different AV, CAV, or CACC models.
2. Scenario Comparative Analysis allows the analyst to compare the capacities for two saturation flow models, for all phases.
3. Phase Comparative Analysis allows for the comparison of all capacity scenarios across a single phase.
In addition, SCAT allows for the adjustment of all models to the same base saturation flow, that is, the saturation flow with 0 percent CAVs is set to the same value for all models. When drawn from the literature, each saturation flow model has its own assumed base saturation flow, ranging from approximately 1,900 veh/hr/ln to 2,400 veh/hr/ln. To help explore the relative difference with increasing or decreasing CAV penetration rates, SCAT enables the normalization of base saturation flows. However, caution should be exercised in the interpretation of these values. The applied normalization is a simple linear adjustment to all saturation values for a given model. That is, if the reported base value in the source literature is 2,100 veh/hr/ln for a given model, and the analyst wishes to consider all models at a base saturation flow of 2,000 veh/hr/ln, then 100 veh/hr/ln will be subtracted from the saturation flow value at all penetration levels. This adjustment is intended to provide convenience for comparing models' relative rates of change. However, the models have not been executed with the new base saturation flow as in the original literature source. It some models it is likely that the linear adjustment assumption is an oversimplification of the impact of changing the base saturation flow. For instance, another reasonable assumption could be to reduce the base rate for the given model to the
113

same value (i.e., set the 0 percent penetration rate to 2,000 veh/hr/ln) and make proportionally smaller changes to the saturation flow as the penetration rate increases. At 100 percent CAV, the saturation flow would be unchanged from the original source literature, as the base saturation flow (i.e., all human drivers) has little influence on the 100 percent CAV market penetration saturation flow. Individual Scenario Analysis To complete the Individual Scenario Analysis, enter the following information in the Data Input Section (figure 40):
1. Enter the desired Phase Lengths (in seconds). 2. Enter values for Yellow, Red Clear, and Lost Time under Other Signal Data. 3. Enter values for the Number of Lanes for each Phase. 4. Select the saturation flow Analysis Option to be analyzed. 5. Select the checkbox under Base Saturation Flow if all models are to be set to the
same base saturation flow.
114

Figure 40. Screenshot. SCAT Individual Scenario Analysis example Data Input Section.
The analyst can confirm that the signal control has been correctly input by reviewing the Data Phase Layout section (figure 41). Separate checks are provided to confirm that the rings have the same cycle length and that the phase pairs on each side of the barrier have the same sum. The analyst should confirm each of these reads "OK".
115

Figure 41. Screenshot. SCAT Individual Scenario Analysis example Data Phase Layout section.
The calculated capacity values for the selected analysis option will be shown in Analysis Table Output (figure 42) and graphically in Analysis Graphical Output (figure 43).
116

Figure 42. Screenshot. SCAT Individual Scenario Analysis example Analysis Table Output section.
Figure 43. Screenshot. SCAT Individual Scenario Analysis example Analysis Graphical Output section. 117

Scenario Comparative Analysis To complete the Scenario Comparative Analysis, enter the following information in the Data Input section (figure 44):
1. Select the checkboxes for two scenarios for comparison under Scenario Selection and Select.
2. Enter the AV penetration rates as Range Low and Range High values under AV Penetration Range (default is 0 to 100).
3. Confirm there are no errors in the range selection, i.e., Error Checks read "OK". All other data are drawn from the Individual Scenario Analysis.
118

Figure 44. Screenshot. SCAT Scenario Comparative Analysis example Data Input section.
Next, the capacity values for each phase, for each or the two selected models will be provided in the Capacity Tables section (figure 45) and the graphical results will be provided in the Capacity Graphs section (figure 46). It is critical to note that the Capacity Tables and Capacity Graphs will reflect the Base Saturation Flow adjustment selection in the Individual Scenario Analysis, Data Input Section.
119

Figure 45. Screenshot. SCAT Scenario Comparative Analysis example Capacity Tables section.
120

Figure 46. Screenshot. SCAT Scenario Comparative Analysis example Capacity Graphs section.
121

Phase Comparative Analysis To complete the Phase Comparative Analysis, complete the following steps in the Data Input section (figure 47):
1. Enter the Phase to be compared across models under Selected Phase. 2. Enter the AV penetration rates as Range Low and Range High values under
Desired AV Rates (default 0 to 100). 3. Confirm there are no errors in the range selection, i.e., Error Checks read "OK". All other data are drawn from Individual Scenario Analysis.
Figure 47. Screenshot. SCAT Phase Comparative Analysis example Data Input section
Next, the capacity values for the given phase, for each model will be provided in the Capacity Per AV Scenario table (figure 48) and the graphical results will be provided in
122

the Capacity Graph section (figure 49). As before, it is critical to note that the Capacity Tables and Capacity Graph will reflect the Base Saturation Flow adjustment selection in the Individual Scenario Analysis, Data Input Section.
Figure 48. Screenshot. SCAT Phase Comparative Analysis example Capacity Per AV Scenario section.
123

Figure 49. Screenshot. SCAT Phase Comparative Analysis example Capacity Graph section.
SUMMARY This chapter presented a Simplified Capacity Analysis Tool for exploring the potential impact of various levels of CAV market penetration on signalized intersection capacity. As seen, SCAT is an Excel-based tool that provides capacity estimates for user-selected phase timings. To reflect the lack of a single accepted representative CAV technology model, SCAT--drawing on the literature and a simulation modeling effort--incorporates results from a selection of saturation flow models across CAV market penetration rates from 0 to 100 percent. The analyst may utilize SCAT to explore a range of potential futures and understand the sensitivity of current intersections, as well as future designs, to potential CAV operating characteristics.
124

To allow for a broader application, next steps in the development of SCAT should incorporate left-turnpermitted movements and shared lanes. Additionally, an ability for analysts to enter a given intersection volume set to be compared directly against model capacities should be added, automating the creation of volume-to-capacity ratios for the various models. Finally, as the development of CAV technology and traffic models is in constant flux, a frequent review and update of the selected models should be undertaken. As new models are developed based on additional field data, recent technology advances, changes in legislation related to required AV characteristics, etc., the addition of these to the SCAT saturation flow estimates will allow for an increasingly robust analysis.
125

CHAPTER 7. CONCLUSIONS AND RECOMMENDATIONS
Many studies support an optimistic outlook on the traffic-flow impacts of autonomous vehicles based on models that assume both AVs and human-driven vehicles express cooperative behaviors. However, these studies have not considered the impacts on traffic performance of potential aggressive interactions of HDVs with AVs in a mixed environment (i.e., AVs and HDVs). Concerns of such interactions occurring are not unwarranted as mobility service companies have observed aggressive human-driver behaviors directed at their AV test fleets, as well as the already existing aggressive behavior that may be observed at merge locations with heavy queuing.
To aid in understanding the potential impact of aggressive HDV with AV interactions, this effort has investigated a merging situation at an off-ramp. Three classes of vehicles are simulated: AVs, HDVs, and aggressive human-driven vehicles. AHDVs represent human-driven vehicles with aggressive merging-behavior characteristics. To perform this study, AHDV behavior at a merge section of a freeway exit ramp, in a mixed traffic environment, is simulated using the open-source traffic simulation package SUMO (Eclipse Foundation 2020). Two types of potential AHDV merging behavior when interacting with an AV are modeled: (1) aggressive merge with maximum advancement, and (2) aggressive merge with zipper. The aggressive merge with maximum advancement represents the highest level of aggressive behavior. The AHDVs with this behavior target the farthest reachable AV on the deceleration lane to act as the following vehicle in the receiving lane, i.e., the AHDV will lane change in front of the AV, essentially without regard for the available gap. In the second type, the aggressive merge with zipper, the
126

AHDVs continue to target downstream AVs in the exit lane, but avoid the scenario where the same AV is targeted by multiple AHDVs.
The impacts of the AHDVs' aggressive behaviors in a mixed-traffic environment (i.e., AVs, HDVs, and AHDVs) on different network traffic characteristics, such as travel time and capacity, is demonstrated. Four experiments are conducted to explore the impact of the AHDV behavior on traffic operations. The first experiment observes the change in speed of the target AV, as well as the following traffic, when a platoon of 10 AHDVs merges in front of the AV near a freeway exit. The second and third experiments observe the travel times of exiting AHDVs and other vehicles when AHDVs are randomly distributed throughout the traffic stream with varying percentages of AVs and AHDVs in the traffic composition. The fourth experiment considers the impact on capacity in a similar merging situation where vehicle behavior is set as cooperative or noncooperative utilizing SUMO driver-behavior parameters.
Experiments 1 through 3 showed that the presence of human drivers' aggressive merging behaviors had adverse effects on AVs and HDVs. The adverse effects were more significant in high congestion, when there is a queue on the deceleration lane. The impacts of AHDVs' aggressive merges were muted by the larger headways between vehicles in low congestion when there is no queue on the deceleration lane. Based on the experiment 2 and experiment 3 results, AHDVs had a higher travel time gain with higher level of aggressive behaviors, which in return had greater adverse effects on the AVs' and the HDVs' travel times. Throughout the experiments, the system-wide travel time tended to be relatively stable, indicating that the AHDV travel-time improvements came at the expense of AVs' and other vehicles' travel time.
127

Experiment 4 took a closer look at the impact of cooperative behaviorinduced aggressive merges on capacity. It was seen that when most vehicles are either fully cooperative or noncooperative, similar capacities are obtained; however, where a higher percentage of cooperative vehicles are positioned to be targeted by more aggressive vehicles, this aggressive-to-non-aggressive interaction can significantly reduce capacity. In addition, it was seen that, similar to experiments 1 through 3, AHDV gains were achieved at the expense of AVs. Finally, even in those scenarios where the overall capacity was not significantly changed in response to the variation in the percentage of cooperative vehicles in the traffic, increased fluctuations in the flow may potentially negatively impact operations as well as the safety conditions in the upstream traffic.
As a final component of this research, an Excel-based Simplified Capacity Analysis Tool is developed. This tool draws predicted saturation flow rates, at various connected and autonomous vehicle market penetration rates, from the literature and a simulation experiment. These saturation flow rates are utilized to determine potential phase capacities at a signalized intersection. While the freeway SUMO experiments focused on the impact of lane changing, SCAT explores the impact of CAV car-following and platooning behaviors. It is seen that a wide variation in capacity predictions may be found throughout the literature, from slight reductions to significant increases in capacity as AV market penetration increases. Across the literature, when considering the car-following aspect of AV operations, it is clear that two key sets of assumptions are driving the predictions: the first is the headways selected by the AVs in a mixed traffic environment, and the second is the characteristic of AV platoons, i.e., platooned vehicle spacing and maximum platoon length.
128

The findings of this study suggest that despite the general beliefs in the benefits of autonomous vehicles, there may be adverse impacts on the non-aggressive vehicle travel times in the presence of human drivers' aggressive merging behaviors in a mixed-traffic environment, especially in congested conditions. Thus, when the potential benefits of the AV are most needed, i.e., at or near capacity, it is possible that human interaction may negate many of the potential savings.
RECOMMENDATIONS Given the high state of uncertainty in AV driving-behavior characteristics and a similar level of uncertainty in the behavior of human-driven vehicles when interacting with AVs, it is extremely difficult to incorporate AVs into current planning and design processes with any sense of assuredness. In the near-term this uncertainty will likely only increase with the development of more AV models, countless future predictions, trial AV deployment successes and failures, etc. However, based on this project, GDOT can likely achieve an early sense of the ultimate operational impacts of AVs by tracking three primary leading indicators:
1. As AV tests continues, or low market penetration occurs, is a rise in aggressive interactions witnessed?
2. What are the headways being adopted by AV manufactures, and what are the potential regulatory requirements?
3. Are platoons implemented in AVs, and if so, what are the spacing requirements and maximum length restrictions, which are again potentially manufacturer and/or regulatory agency driven?
129

As the direction of each of these indicators becomes clearer, GDOT will be able to select the more likely futures from the many potential predicted futures, allowing AV penetration to begin to influence design and policy decisions in a more informed manner. For example, if it is seen that human-driven vehicles begin to express aggressive interaction with AVs, then GDOT may need to revisit signal control at ramp junctions, where eliminating queueing on the freeway is a priority to minimize targeting opportunities. Additionally, design changes such as increasing use of delineator posts immediately upstream of the gore area may be required. Similarly, as platooning parameters clarify, signal control may be revisited, optimizing detection and control strategies to incorporate processing of maximum platoon lengths, that is, optimal control will minimize splitting platoons. Lastly, this study did not address potential safety impacts that could arise from aggressive human-driven vehicle AV interaction. Future efforts need to investigate potential safety impacts and begin to develop recommendations for design, operations, or policy mitigations.
130

APPENDIX A: OBSERVED HEADWAYS PIB AT BERKELEY LAKE (33.985340, -84.171123) Through Movement Headways

Figure 50. Map. Through movements PIB at Berkeley Lake. Source: Google Maps

Table 15. Through movement headway distribution PIB at Berkeley Lake.

Movement Lane

NB

Lane 1

Lane 2

SB

Lane 1

Lane 2

Gaussian Dist 1 mu1 SD1 Weighting 1 1.976 0.576 0.811 1.839 0.522 0.719 2.088 0.675 0.832 2.107 0.710 0.814

Gaussian Dist 2 mu2 SD2 Weighting 2 3.469 1.525 0.189 3.117 1.186 0.281 4.781 1.858 0.168 5.369 2.152 0.186

131

Figure 51. Plot. PIB at Berkeley Lake through movement headway data visualization (top left NB Lane 1, top right NB Lane 2, bottom left SB Lane 1, bottom right SB Lane 2).
132

Figure 52. Plot. PIB at Berkeley Lake through movement headway distribution (top left NB Lane 1, top right NB Lane 2, bottom left SB Lane 1, bottom right SB Lane 2).
133

Left-turn Movement Headways

Figure 53. Map. Left-turn movements PIB at Berkeley Lake. Source: Google Maps

Table 16. Left-turn movement headway distribution PIB at Berkeley Lake.

Movement Lane

EBL WBL SBL

Lane 1 Lane 1 Lane 1 Lane 2

Gaussian Dist 1

mu1 SD1 Weighting 1

2.326 0.628

0.786

2.248 0.490

0.757

1.889 0.296

0.613

1.907 0.394

0.607

Gaussian Dist 2

mu2 SD2 Weighting 2

4.764 1.893

0.214

4.746 1.920

0.243

3.175 0.673

0.387

2.838 0.554

0.393

134

Figure 54. Plot. PIB at Berkeley Lake left-turn movement headway visualization (top left Eastbound Lane, top right Westbound Lane, bottom left SB Lane 1, bottom right SB Lane 2).
135

Figure 55. Plot. PIB at Berkeley Lake left-turn movement headway distribution (top left Eastbound Lane, top right Westbound Lane, bottom left SB Lane 1, bottom right SB Lane 2).
136

PIB AT MEDLOCK BRIDGE ROAD (33.961047, -84.208518) Through Movement Headways

Figure 56. Map. Through movement PIB at Medlock Bridge Road. Source: Google MapTable 17. Through movement headway distribution PIB at
Medlock Bridge Road.

Movement Lane

WB

Lane 1

Lane 2

EB

Lane 1

Lane 2

Gaussian Dist 1 mu1 SD1 Weighting 1 2.281 0.758 0.712 2.207 0.729 0.799 2.205 0.672 0.754 1.939 0.613 0.774

Gaussian Dist 2 mu2 SD2 Weighting 2 5.087 1.873 0.288 5.157 1.764 0.201 4.746 1.811 0.246 4.011 1.652 0.226

137

Figure 57. Plot. PIB at Medlock Bridge Road through movement headway data visualization (top left WB Lane 1, top right WB Lane 2, bottom left EB Lane 1, bottom right EB Lane 2).
138

Figure 58. Plot. PIB at Medlock Bridge Road through movement headway distribution (top left WB Lane 1, top right WB Lane 2, Bottom Left EB Lane 1, bottom right EB Lane 2).
139

Left-turn Movement Headways

Figure 59. Map. Left-turn movement PIB at Medlock Bridge Road. Source: Google Maps

Table 18. Left-turn movement headway distribution PIB at Medlock Bridge Road.

Movement
EBL WBL

Lane
Lane 1 Lane 1

Gaussian Dist 1

mu1 SD1 Weighting 1

2.120 0.426

0.647

2.250 0.491

0.759

Gaussian Dist 2

mu2 SD2 Weighting 2

3.251 0.755

0.353

4.759 1.919

0.241

140

Figure 60. Plot. PIB at Medlock Bridge Road left-turn movement headway data visualization.
Figure 61. Plot. PIB at Medlock Bridge Road left-turn movement headway distribution (left EB lane, right WB lane).
141

APPENDIX B: SPEED FLOW PLOTS FOR EXPERIMENT 4A AND 4B
Figure 62. Plot. Speed vs. flow plots at 500 ft before the start of the deceleration lane in experiment 4a (top left 0%, top right 25%, bottom left 50%, bottom middle 75%, bottom right 100%). 142

Figure 63. Plot. Speed vs. flow plots at the start of the deceleration lane in experiment 4a (top left 0%, top right 25%, bottom left 50%, bottom middle 75%, bottom right 100%).
143

Figure 64. Plot. Speed vs. flow plots at the start of the ramp in experiment 4a (top left 0%, top right 25%, bottom left 50%, bottom middle 75%, bottom right 100%).
144

Figure 65. Plot. Speed vs. flow plots at 500 ft before the start of the deceleration lane in experiment 4b (top left 0%, top right 25%, bottom left 50%, bottom middle 75%, bottom right 100%).
145

Figure 66. Plot. Speed vs. flow plots at the start of the deceleration lane in experiment 4b (top left 0%, top right 25%, bottom left 50%, bottom middle 75%, bottom right 100%).
146

Figure 67. Plot. Speed vs. flow plots at the start of the ramp in experiment 4b (top left 0%, top right 25%, bottom left 50%, bottom middle 75%, bottom right 100%).
147

ACKNOWLEDGMENTS The information, data, or work presented herein was funded in part by the Georgia Department of Transportation in cooperation with U.S. Department of Transportation Federal Highway Administration as GDOT Research Project 18-23. The authors thank Mr. Alan Davis in the Office of Traffic Operations in GDOT for his support during the project. The views and opinions of the authors expressed herein do not necessarily state or reflect those of the State of Georgia or any agency thereof. This paper does not constitute a standard, specification, or regulation.
148

REFERENCES
Adebisi, A., Guo, Y., Schroeder, B., Ma, J., Cesme, B., Bibeka, A., and Morgan, A. (2021). "Highway Capacity Manual Capacity Adjustment Factor Development for Connected and Automated Traffic at Signalized Intersections." Journal of Transportation Engineering, Part A: Systems, 148(3). Available online: https://doi.org/10.1061/JTEPBS.0000631.
Anagnostopoulos, A. and Kehagia, F. (2019). "CAVs and Roundabouts: Research on Traffic Impacts and Design Elements." 47th European Transport Conference, Dublin, Ireland, October 911. Available online: http://dx.doi.org/10.1016/j.trpro.2020.09.008.
Aria, E., Olstam, J., and Schwietering, C. (2016). "Investigation of Automated Vehicle Effects on Driver's Behavior and Traffic Performance." Transportation Research Procedia, 15, pp. 761770. Available online: https://doi.org/10.1016/j.trpro.2016.06.063.
Atkins. (2016). Research on the Impacts of Connected and Autonomous Vehicles (CAVs) on Traffic Flow: Stage 2: Traffic Modelling and Analysis: Technical Report. Department for Transport, London, UK. Available online: https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_ data/file/530093/impacts-of-connected-and-autonomous-vehicles-on-traffic-flowtechnical-report.pdf.
CoEXist. (2021a). "AV-Ready' Transport Models and Road Infrastructure for the Coexistence of Automated and Conventional Vehicles." CoEXist (website). https://ec.europa.eu/inea/en/horizon-2020/projects/h2020-transport/automated-roadtransport/coexist, last accessed December 13, 2021.
CoEXist. (2021b). "Resources." CoEXist (website). Available online: https://www.h2020-coexist.eu/resources/, last accessed December 13, 2021.
DataFromSky. (2021a). "Data From Sky." (website). Czech Republic. Available online: https://datafromsky.com/, last accessed December 13, 2021.
DataFromSky. (2021). "DataFromSky Viewer." (website). Czech Republic. Available online: https://datafromsky.com/news/datafromsky-viewer-new-version-new-features/, last accessed December 13, 2021.
Eclipse Foundation. (2020). "Simulation of Urban MObility." (website) Available online: https://www.eclipse.org/sumo/.
149

Elvarsson, A.B. (2017). "Modeling Urban Driving and Stopping Behavior for Automated Vehicles." Semester Project, IVT, ETH, Zrich, Switzerland.
Espinosa, M. (2015). Safety Evaluation of Signalized Intersections with Automated Vehicles at Various Penetration Levels Based on Conflict Analysis of Simulated Traffic. Master's thesis, Ryerson University, Toronto, Canada. Available online: https://doi.org/10.32920/ryerson.14652957.
General Motors. (2015). "General Motors Promotes DSRC for Connected Vehicle (V2X) Technology Development in China." Corporate Newsroom, November 24. Available online: https://media.gm.com/media/cn/en/gm/news.detail.html/content/Pages/news/cn/en/2015/ Nov/1124_v2x.html, last accessed January 10, 2020.
Ghiasi, A., Hussain, O., Qian, S., and Li, X. (2017). "A Mixed Traffic Capacity Analysis and Lane Management Model for Connected and Automated Vehicles: A Markov Chain Method." Transportation Research Part B: Methodological, 106, pp. 266292. Available online: https://doi.org/10.1016/j.trb.2017.09.022.
Guo, J., Cheng, S., and Liu, Y. (2020). "Merging and Diverging Impact on Mixed Traffic of Regular and Autonomous Vehicles." IEEE Transactions on Intelligent Transportation Systems, 22(3), pp. 16391649. Available online: http://dx.doi.org/10.1109/TITS.2020.2974291.
Hamilton, I.A. (2019). "Uber Says People Are Bullying Its Self-Driving Cars with Rude Gestures and Road Rage." Business Insider, June 13. Available online: https://www.businessinsider.com/uber-people-bullying-self-driving-cars-2019-6, last accessed January 10, 2020.
Hedlund, J. (2017). Autonomous Vehicles Meet Human Drivers: Traffic Safety Issues for States. Governors Highway Safety Association, Washington, DC. Available online: http://www.ghsa.org/sites/default/files/2017-01/AV%202017%20-%20FINAL.pdf.
Hua, X., Yu, W., Wang, W., and Xie, W. (2020). "Influence of Lane Policies on Freeway Traffic Mixed with Manual and Connected and Autonomous Vehicles." Journal of Advanced Transportation, 2020, [3968625]. Available online: https://doi.org/10.1155/2020/3968625.
Hunter, M. (2021). Vissim 11 Simulation Guidance. Final Report, FHWA-GA-11-1833, Georgia Department of Transportation, Atlanta, GA.
150

INRIA. (2021). "sklearn.mixture.GaussianMixture." Scikit-learn. (website) Available online: https://scikitlearn.org/stable/modules/generated/sklearn.mixture.GaussianMixture.html, last accessed December 13, 2021.
Kapania, N.R., Govindarajan, V., Borrelli, F., and Gerdes, J.C. (2019). "A Hybrid Control Design for Autonomous Vehicles at Uncontrolled Intersections." 2019 IEEE Intelligent Vehicles Symposium (IV), pp. 16041611. Available online: https://doi.org/10.1109/IVS.2019.8814116.
Kesting, A., Treiber, M., Helbing, D. (2010). "Enhanced Intelligent Driver Model to Access the Impact of Driving Strategies on Traffic Capacity." Philosophical Transactions of the Royal Society A, 368(1928), pp. 45854605. Available online: https://doi.org/10.1098/rsta.2010.0084.
Li, D. and Wagner, P. (2019). "Impacts of Gradual Automated Vehicle Penetration on Motorway Operation: A Comprehensive Evaluation." European Transport Research Review, 11(1), article 36. Available online: http://dx.doi.org/10.1186/s12544-019-0375-3.
Li, S., Seth, D., and Cummings, M.L. (2020). "Traffic Efficiency and Safety Impacts of Autonomous Vehicle Aggressiveness." 99th Annual Transportation Research Board Meeting, Washington, DC, January. Available online: https://hal.pratt.duke.edu/sites/hal.pratt.duke.edu/files/u36/AV%20Aggressiveness%20T RB%20Paper%20final.pdf.
Liu, P. and Fan, W. (2020). "Exploring the Impact of Connected and Autonomous Vehicles on Freeway Capacity Using a Revised Intelligent Driver Model." Transportation Planning and Technology, 43(3), pp. 279292. Available online: https://doi.org/10.1080/03081060.2020.1735746.
Liu, Y., Guo, J., Taplin, J., and Wang, Y. (2017). "Characteristic Analysis of Mixed Traffic Flow of Regular and Autonomous Vehicles Using Cellular Automata." Journal of Advanced Transportation, 2017, [8142074]. Available online: https://doi.org/10.1155/2017/8142074.
Liu, H., Kan, X.D., Shladover, S.E., Lu, X.Y., and Ferlis, R.E. (2018). "Impact of Cooperative Adaptive Cruise Control on Multilane Freeway Merge Capacity." Journal of Intelligent Transportation Systems, 22(3), pp. 263275. Available online: http://dx.doi.org/10.1080/15472450.2018.1438275.
151

Lopez, P.A., Behrisch, M., Bieker-Walz, L., Erdmann, J., Fltterd, Y.P., Hilbrich, R., Lcken, L., Rummel, J., Wagner, P., and Wiebner, E. (2018). "Microscopic Traffic Simulation Using SUMO." 21st International Conference on Intelligent Transportation Systems (ITSC), Maui, HI, November 47. Available online: https://elib.dlr.de/127994/1/08569938.pdf.
Lu, Q., Tettamanti, T., Hrcher, D., and Varga, I. (2020). "The Impact of Autonomous Vehicles on Urban Traffic Network Capacity: An Experimental Analysis by Microscopic Traffic Simulation." Transportation Letters, 12(8), pp. 540549. Available online: https://doi.org/10.1080/19427867.2019.1662561.
Lu, Q., Tettamanti, T., and Varga, I. (2018). "Impacts of Autonomous Vehicles on the Urban Fundamental Diagram." 5th International Conference on Road and Rail Infrastructure, CETRA 2018, Zadar, Croatia, May 1719. Available online: http://real.mtak.hu/id/eprint/83699.
Martin-Gasulla, M.P., Sukennik, P., and Lohmiller, J. (2019). "Investigation of the Impact on Throughput of Connected Autonomous Vehicles with Headway Based on the Leading Vehicle Type." Transportation Research Record: Journal of the Transportation Research Board, 2673(5), pp. 617626. Available online: https://doi.org/10.1177%2F0361198119839989.
Mesionis, G., Brackstone, M., and Gravett, N. (2020). "Microscopic Modeling of the Effects of Autonomous Vehicles on Motorway Performance." Transportation Research Record: Journal of the Transportation Research Board, 2674(11), pp. 697707. Available online: https://doi.org/10.1177%2F0361198120949243.
Milanes, V. and Shladover, S. (2014). "Modeling Cooperative and Autonomous Adaptive Cruise Control Dynamic Responses using Experimental Data." Transportation Research Part C: Emerging Technologies, 48, pp. 285300. Available online: https://doi.org/10.1016/j.trc.2014.09.001.
Morando, M.M., Tian, Q., Truong, L.T., and Vu, H.L. (2018). "Studying the Safety Impact of Autonomous Vehicles Using Simulation-Based Surrogate Safety Measures." Journal of Advanced Transportation, 2018, [6135183]. Available online: https://doi.org/10.1155/2018/6135183.
Morando, M.M., Truong, L.T., and Vu, H.L. (2017). "Investigating Safety Impacts of Autonomous Vehicles Using Traffic Micro-simulation." Australasian Transport Research Forum 2017, Auckland, New Zealand, November 2729. Available online: https://www.australasiantransportresearchforum.org.au/papers/2017.
152

Mller, L., Risto, M., and Emmenegger, C. (2016). "The Social Behavior of Autonomous Vehicles." UbiComp '16: Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing: Adjunct, September, pp. 686689. Available online: https://doi.org/10.1145/2968219.2968561.
National Academies of Sciences, Engineering, and Medicine. (2017). Strategies to Advance Automated and Connected Vehicles. The National Academies Press, Washington, DC. Available online: https://doi.org/10.17226/24873.
National Highway Traffic Safety Administration (NHTSA). (2017). Automated Driving Systems 2.0: A Vision for Safety. DOT HS 812 442, U.S. Department of Transportation, Washington, DC. Available online: https://www.nhtsa.gov/sites/nhtsa.gov/files/documents/13069aads2.0_090617_v9a_tag.pdf.
Nishimura, Y., Fujita, A., Hiromori, A., Yamaguchi, H., Higashino, T., Suwa, A., Urayama, H., Takeshima, S., and Takai, M. (2019). "A Study on Behavior of Autonomous Vehicles Cooperating with Manually-Driven Vehicles." 2019 IEEE International Conference on Pervasive Computing and Communications (PerCom), Kyoyo, Japan, March 1214. Available online: http://sigiss.work/percom2019/papers/p212-nishimura.pdf.
NumPy Steering Council. (2021). "NumPy." (website) Available online: https://numpy.org/, last accessed December 13, 2021.
NumFOCUS. (2021). "pandas." (website) Available online: https://pandas.pydata.org/, last accessed December 13, 2021.
Papadoulis, A., Quddus, M., and Imprialou, M. (2019). "Evaluating the Safety Impact of Connected and Autonomous Vehicles on Motorways." Accident Analysis & Prevention, 124, pp. 1222. Available online: https://doi.org/10.1016/j.aap.2018.12.019.
PTV Group. (2021). "PTV Vissim." (website) Available online: https://www.ptvgroup.com/en/solutions/products/ptv-vissim/, last accessed December 13, 2021.
Python. (2019). "Python 3.8.0". (website) Available online: https://www.python.org/downloads/release/python-380/
Rahman, M.S. and Abdel-Aty, M. (2018). "Longitudinal Safety Evaluation of Connected Vehicles' Platooning on Expressways." Accident Analysis & Prevention, 117, pp. 381 391. Available online: https://doi.org/10.1016/j.aap.2017.12.012.
153

Randazzo, R. (2018). "A Slashed Tire, a Pointed Gun, Bullies on the Road: Why Do Waymo Self-Driving Cars Get So Much Hate?" USA Today, December 13. Available online: https://www.usatoday.com/story/money/cars/2018/12/13/waymo-self-drivingvehicles-face-harassment-road-rage-phoenix-area/2288833002/, last accessed January 10, 2020.
Richter, G., Grohmann, L., Nitsche, P., and Lenz, G. (2019). "Anticipating Automated Vehicle Presence and the Effects on Interactions with Conventional Traffic and Infrastructure." SUMO User Conference 2019, EPiC Series in Computing, 62, pp. 230 243. Available online: http://dx.doi.org/10.29007/s6m7.
Rothenbucher, D., Li, J., Sirkin, D., Mok, B., and Ju, W. (2016). "Ghost Driver: A Field Study Investigating the Interaction Between Pedestrians and Driverless Vehicles." 25th IEEE International Symposium on Robot and Human Interactive Communication (ROMAN), New York, NY, August 2631, pp. 795802. Available online: https://doi.org/10.1109/ROMAN.2016.7745210.
Schroeder, B., Morgan, A., Ryus, P., Cesme, B., Bibeka, A., Rodegerts, L., and, Ma, J. (2021). Capacity Adjustment Factors for Connected and Automated Vehicles in the Highway Capacity Manual. Draft Phase 1 Report, Pooled Fund Study, Contract #B35967, Oregon DOT, Salem, OR.
Schwarting, W., Alonso-Mora, J., and Rus, D. (2018). "Planning and Decision-Making for Autonomous Vehicles." Annual Review of Control, Robotics, and Autonomous Systems, 1, pp. 187210. Available online: https://doi.org/10.1146/annurev-control060117-105157.
Shi, L. and Prevedouros, P. (2016). "Autonomous and Connected Cars: HCM Estimates for Freeways with Various Market Penetration Rates." Transportation Research Procedia, 15, pp. 389402. Available online: https://doi.org/10.1016/j.trpro.2016.06.033.
Stanek, D., Huang, E., Milam, R.T., and Wang, Y. (2017). "Measuring Autonomous Vehicle Impacts on Congested Networks Using Simulation." Transportation Research Board 97th Annual Meeting, Washington, DC, January 711.
Sukennik, P. (2018). Micro-simulation Guide for Automated Vehicles. Deliverable 2.5 of CoEXist project, October. Available online: https://www.h2020-coexist.eu/wpcontent/uploads/2018/11/D2.5-Micro-simulation-guide-for-automated-vehicles.pdf.
154

Sukennik, P. and Kautzsch, L. (2018). Default Behavioral Parameter Sets for Automated Vehicles (AVs). Deliverable 2.5 of CoEXist project, February. Available online: https://www.h2020-coexist.eu/wp-content/uploads/2018/10/D2.3-default-behaviouralparameter-sets_final.pdf.
Sukennik, P., Lohmiller, J., and Schlaich, J. (2018). "Simulation-Based Forecasting the Impacts of Autonomous Driving." International Symposium of Transport Simulation (ISTS'18) and the International Workshop on Traffic Data Collection and its Standardization (IWTDCS'18), Transportation Research Procedia, Elsevier, Amsterdam, The Netherlands.
Tennant, C., Howard, S., Franks, B., and Bauer, M.W. (2016). Autonomous Vehicles Negotiating a Place on the Road: A Study on How Drivers Feel about Interacting with Autonomous Vehicles on the Road. LSE Consulting, London School of Economics and Political Science, City University of London, UK. Executive summary available online: https://www.lse.ac.uk/business/consulting/assets/documents/autonomous-vehiclesexecutive-summary.pdf.
Tian, R., Li, S., Kolmanovsky, I., Girard, A., and Yildiz, Y. (2018). "Adaptive GameTheoretic Decision Making for Autonomous Vehicle Control at Roundabouts." IEEE Conference on Decision and Control (CDC), 2018, pp. 321326. Available online: http://dx.doi.org/10.1109/CDC.2018.8619275.
Tiblijas, D.A., Giuffr, T., Surdonja, S., and Trubia, S. (2018). "Introduction of Autonomous Vehicles: Roundabouts Design and Safety Performance Evaluation." Sustainability, 10(4), 1060. Available online: http://dx.doi.org/10.3390/su10041060.
Toms, R.F., Fernandes, P., Macedo, E., Bandeira, J.M., and Coelho, M.C. (2019). "Assessing the Emission Impacts of Autonomous Vehicles on Metropolitan Freeways." Transportation Research Procedia, 47, pp. 617624. Available online: https://doi.org/10.1016/j.trpro.2020.03.139.
Toth, C. S.(2014) "Empirical Study of the Effect of Offramp Queues on Freeway Mainline Traffic Flow". Dissertation, Georgia Institute of Technology. Available online: https://www.semanticscholar.org/paper/Empirical-study-of-the-effect-of-offramp-queueson-Toth/fcfe4bb7b959776b3bc5c2619049d1e4b483af62
Treiber, M., Hennecke, A., and Helbing, D. (2000). "Congested Traffic States in Empirical Observations and Microscopic Simulations." Physical Review E, 62(2), pp. 18051824. Available online: https://arxiv.org/ct?url=https%3A%2F%2Fdx.doi.org%2F10.1103%2FPhysRevE.62.180 5&v=cfae6527.
155

Wagner, P. (2016). "Traffic Control and Traffic Management in a Transportation System with Autonomous Vehicles." In: Maurer, M., Gerdes, J., Lenz, B., Winner, H. (eds) Autonomous Driving, Springer, Berlin, Heidelberg. Available online: http://dx.doi.org/10.1007/978-3-662-48847-8_15.
Wang, Y. and Wang, L. (2017). "Autonomous Vehicles' Performance on Single Lane Road: A Simulation under VISSIM Environment." 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), October 1416. Available online: https://doi.org/10.1109/CISP-BMEI.2017.8302162.
Wei, J., Dolan, J.M., and Litkouhi, B. (2013). "Autonomous Vehicle Social Behavior for Highway Entrance Ramp Management." 2013 IEEE Intelligent Vehicles Symposium (IV), pp. 201207. Available online: http://dx.doi.org/10.1109/IVS.2013.6629471.
Xu, S. (2021). "How to Use Python to Separate Two Gaussian Curves?" stack overflow, https://stackoverflow.com/questions/51318981/how-to-use-python-to-separate-twogaussian-curves, last accessed December 13, 2021.
Yeo, H., Skabardonis, A., Halkias, J., Colyar, J., and Alexiadis, V. (2008). "Oversaturated Freeway Flow Algorithm for Use in Next Generation Simulation." Transportation Research Record: Journal of the Transportation Research Board, 2088(1), pp 6879. Available online: http://dx.doi.org/10.3141/2088-08.
Yu, H., Tak, S., Park, M., and Yeo, H. (2019). "Impact of Autonomous-Vehicle-Only Lanes in Mixed Traffic Conditions." Transportation Research Record: Journal of the Transportation Research Board, 2673(9), pp. 430439. Available online: https://doi.org/10.1177%2F0361198119847475.
Zhao, C., Wang, W., Li, S., and Gong, J. (2019). "Influence of Cut-In Maneuvers for an Autonomous Car on Surrounding Drivers: Experiment and Analysis." IEEE Transactions on Intelligent Transportation Systems, 21(6), pp. 22662276. Available online: http://dx.doi.org/10.1109/TITS.2019.2914795.
Zhou, M., Qu, X., and Jin, S. (2017). "On the Impact of Cooperative Autonomous Vehicles in Improving Freeway Merging: A Modified Intelligent Driver Model-Based Approach." IEEE Transactions on Intelligent Transportation Systems, 18(6), pp. 1422 1428. Available online: https://doi.org/10.1109/TITS.2016.2606492.
156