ROUNDABOUT CAPACITY AND COMPARATIVE

SOFTWARE ANALYSIS

by

Michael S. DeAmico

A thesis submitted in partial fulfillment of

the requirements for the degree of

Master of Science
(Civil and Environmental Engineering)

At the
UNIVERSITY OF WISCONSIN – MADISON
2012
 i

Abstract

Interest continues to increase in how roundabouts can most effectively be used in

the United States (U.S.) transportation system to achieve goals of safety, efficiency,
along with other benefits. Using the right software can enhance accuracy and
productivity in designing for these goals. However, a growing number of software
choices and evolving features provide the motivation for this research because
many questions exist about which software is the best option.
This research explores the popular models and software available for ana-
lyzing roundabout capacity in order to increase understanding of the similarities
and differences. Specifically, two roundabouts, one single-lane and one multi-lane,
that experienced congestion were compared using capacity models and results
from seven software packages: RODEL 1.9.7, ARCADY 7.1, RCAT 1.4, Kreisel 7.0,
Girabase 4.0, HCS 2010 6.1, and SIDRA 5.1.
Analysis showed that in calibrated scenarios, all models performed ac-
ceptably well in terms of root-mean-square error in comparison to the field
data. However, achieving a properly calibrated model is difficult because of data
collection and availability needs. Further, calibration can only be performed on
existing roundabouts, leaving analysis of future roundabouts to rely on projections
or uncalibrated models. From the sites analyzed, the uncalibrated U.S. based
model (NCHRP 572/HCM 2010) and German models fit the observed data better
than all other models. Certainly there are many considerations when choosing
a software package, but in terms of technical accuracy, a package that has the
capability of performing capacity analysis using U.S. based models is desirable
based upon the findings of this research.
In terms of software modeling, lane-by-lane methods were shown to be
ii

more accurate for capacity analysis. Because significant lane imbalance can occur,
variability in performance measures on the same approach is a concern for detailed
analysis. However, the analyst still has a responsibility for understanding how
drivers will utilize the available lanes because default assumptions in the software
packages were not always appropriate.
Ultimately, software is constantly evolving along with roundabout opera-
tions, and this research has come at the beginning of understanding roundabouts
in the U.S. Therefore, a definitive software recommendation is open-ended due
to the unique requirements of an agency and future research. The purpose,
potential growth, evolution, and flexibility of any software and parent company
must be carefully considered in order to make a smart investment for the future.
Further research needs include investigating secondary, advanced features (such
as integration with CAD software, corridor analysis, etc) as well as developing
design guidance and calibrated model parameters.
 iii

Acknowledgements

I would like to thank everyone at the University of Wisconsin Traffic Operations

and Safety (TOPS) Laboratory for allowing me to pursue my education and
helping me through this research. Without the guidance from Dr. David Noyce,
Madhav Chitturi, and Andi Bill, none of this work would have been possible.
Behind the scenes, I would like to thank the student assistants that helped
in the countless hours of data reduction, especially Julia McAdams and Caralee
Corcoran. To all my fellow graduate students, thank you for your advice and
support through the good and challenging times. To my colleagues that have also
been engaged in roundabout research, Brian Porter and Dongxi Zheng, you have
my sincere gratitude and respect.
For my loving wife Amanda, who probably counts vehicles circulating
through roundabouts in her sleep, thank you for sustaining my eternal happiness.
iv

Contents

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x

1 Introduction 1
1.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Objective, Contributions, and Scope of Work . . . . . . . . . . . . . . 3
1.3 Document Organization . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Literature Review 6
2.1 Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Congestion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Gap Acceptance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Capacity Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4.1 U.K. Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4.2 German Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4.3 French Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4.4 NCHRP 572 Model . . . . . . . . . . . . . . . . . . . . . . . . . 19
 v

2.4.5 Australian Method ARR 321 . . . . . . . . . . . . . . . . . . . 22

2.4.5.1 Calibration and Differences Between ARR 321 and
SIDRA . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.6 Summary and Notes on Capacity Model Parameters . . . . . 30
2.5 Other Recent Capacity Studies . . . . . . . . . . . . . . . . . . . . . . 31
2.6 Software Background . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

3 Study Design 35

4 Data Collection and Site Descriptions 41

4.1 Site Selection and Descriptions . . . . . . . . . . . . . . . . . . . . . . 41
4.2 Canal St Site . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 De Pere Site . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.4 Field Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.5 Field Data Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

5 Capacity Analysis 50
5.1 Observed Queuing Data . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.2 Observed Gap Acceptance Data . . . . . . . . . . . . . . . . . . . . . 55
5.3 Measured Site Geometry . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.4 Capacity Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.4.1 Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.4.2 Error Measurement . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.4.3 Model Characteristics . . . . . . . . . . . . . . . . . . . . . . . 63
5.4.4 Canal St Site Capacity Analysis . . . . . . . . . . . . . . . . . 65
5.4.5 De Pere Site Capacity Analysis . . . . . . . . . . . . . . . . . . 71
5.4.6 Capacity Data Analysis Summary . . . . . . . . . . . . . . . . 84
vi

5.5 Capacity Modeling Extensions . . . . . . . . . . . . . . . . . . . . . . 86

5.5.1 Adjusting for Effective Geometry . . . . . . . . . . . . . . . . 86
5.5.2 Combined De Pere EB and NB Data Analysis . . . . . . . . . . 88

6 Software Output and Usability 93

6.1 Canal St Site Software Results . . . . . . . . . . . . . . . . . . . . . . . 97
6.2 De Pere Site Software Results . . . . . . . . . . . . . . . . . . . . . . . 100
6.3 Technical Accuracy Summary . . . . . . . . . . . . . . . . . . . . . . . 104
6.4 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.5 Interface Usability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.5.1 RODEL 1.9.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.5.2 ARCADY 7.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6.5.3 RCAT 1.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.5.4 KREISEL 7.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.5.5 GIRABASE 4.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.5.6 HCS 2010 6.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.5.7 SIDRA 5.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.6 Feature and Interface Comparisons . . . . . . . . . . . . . . . . . . . 120
6.6.1 Input requirements . . . . . . . . . . . . . . . . . . . . . . . . 122
6.6.2 Feature Comparison . . . . . . . . . . . . . . . . . . . . . . . . 124
6.6.3 Installation and License Requirements . . . . . . . . . . . . . 127
6.6.4 Training Needs and Availability of Support . . . . . . . . . . 128
6.6.5 Considerations for Future Software Versions . . . . . . . . . . 129

7 Conclusions and Future Work 130

7.1 Capacity Model Comparison . . . . . . . . . . . . . . . . . . . . . . . 130
7.2 Software Modeling Comparisons . . . . . . . . . . . . . . . . . . . . . 131
 vii

7.3 Software Usability Comparison . . . . . . . . . . . . . . . . . . . . . . 132

7.4 Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
7.5 Other Considerations for Discussion . . . . . . . . . . . . . . . . . . . 134

References 136

Appendix A - De Pere AM Capacity Data 143

Appendix B - Software Output for all Approaches 149

viii

List of Figures

2.1 Example Entering-Circulating Capacity Graph Development . . . . . 7

2.2 Critical Gap Depiction . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Follow-up Headway Depection . . . . . . . . . . . . . . . . . . . . . . 11
2.4 U.K. Model Geometric Parameters . . . . . . . . . . . . . . . . . . . . 15
2.5 Effects of Calibration Parameters for the HCM 2010 Capacity Equa-
tions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.6 Comparison of SIDRA Capacity with Environment Factors and the
ARR 321 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3.1 Software Evaluation Methodology . . . . . . . . . . . . . . . . . . . . 36

4.1 Study Locations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

4.2 Canal St Site Configuration . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3 De Pere Site Configuration . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.4 Typical Camera Setup for Recording Traffic . . . . . . . . . . . . . . . 46
4.5 Sample Synchronized Video Screenshot used for Data Reduction . . 47
4.6 Time Stamp Extraction Software User Interface . . . . . . . . . . . . . 48

5.1 Canal St Site Peak Hour Volumes . . . . . . . . . . . . . . . . . . . . . 54

5.2 De Pere Site Peak Hour Volumes . . . . . . . . . . . . . . . . . . . . . 54
 ix

5.3 Canal St Field Capacity Data . . . . . . . . . . . . . . . . . . . . . . . . 66

5.4 Canal St Approach Based Capacity Comparison . . . . . . . . . . . . 68
5.5 Canal St Lane Based Capacity Comparison . . . . . . . . . . . . . . . 70
5.6 De Pere Northbound PM Field Capacity Data . . . . . . . . . . . . . . 72
5.7 De Pere Northbound Approach Based Capacity Comparison . . . . . 74
5.8 De Pere Northbound Lane Based Capacity Comparison . . . . . . . . 77
5.9 De Pere Eastbound PM Field Capacity Data . . . . . . . . . . . . . . . 79
5.10 De Pere Eastbound Approach Based Capacity Comparison . . . . . . 81
5.11 De Pere Eastbound Lane Based Capacity Comparison . . . . . . . . . 83
5.12 Root Mean Square Error Summary . . . . . . . . . . . . . . . . . . . . 85
5.13 U.K. Model Effective Geometry Comparison from the De Pere Site . 87
5.14 U.K. Model Effective Geometry Comparison from the Canal St Site . 88
5.15 Combined De Pere Capacity Data Comparison . . . . . . . . . . . . . 90

6.1 RODEL 1.9.7 Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

6.2 ARCADY 7.1 Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.3 RCAT 1.4 Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.4 Kreisel 7.0 Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.5 GIRABASE 4.0 Interface . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.6 HCS 2010 6.1 Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.7 SIDRA 5.1 Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
x

List of Tables

2.1 Summary of Capacity Models . . . . . . . . . . . . . . . . . . . . . . . 13

2.2 Summary of Model Equation Parameters . . . . . . . . . . . . . . . . 31

3.1 Evaluation Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

5.1 Summary of Observed PM Peak Minutes of Queuing . . . . . . . . . 52

5.2 Summary of Observed PM Peak Minutes of Queuing . . . . . . . . . 53
5.3 Canal St Site Gap Acceptance Parameters . . . . . . . . . . . . . . . . 55
5.4 De Pere Site Gap Acceptance Parameters . . . . . . . . . . . . . . . . 56
5.5 Canal St Site Geometry and Characteristics . . . . . . . . . . . . . . . 58
5.6 De Pere Site Geometry and Characteristics . . . . . . . . . . . . . . . 59
5.7 Average Entering and Circulating Flows Used for U.K. Model
Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.8 Canal St Field Data Regression Results . . . . . . . . . . . . . . . . . . 65
5.9 RMSE and Model Characteristics from the Canal St Approach Based
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.10 RMSE and Model Characteristics from the Canal St Lane Based
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.11 De Pere Northbound PM Field Data Regression Results . . . . . . . . 71
 xi

5.12 RMSE and Model Characteristics from the De Pere NB Approach

Based Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.13 RMSE and Model Characteristics from the De Pere NB Lane Based
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.14 De Pere Eastbound PM Field Data Regression Results . . . . . . . . . 78
5.15 RMSE and Model Characteristics from the De Pere EB Approach
Based Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.16 RMSE and Model Characteristics from the De Pere EB Lane Based
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.17 RMSE and Model Characteristics from the De Pere Effective Geom-
etry Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.18 RMSE and Model Characteristics from the Canal St Effective Geom-
etry Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.19 Model Characteristics from the Combined De Pere Capacity Data
Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.20 Combined De Pere Capacity Data Set Regression . . . . . . . . . . . . 91

6.1 Software Calibration Procedure Summary . . . . . . . . . . . . . . . . 95

6.2 Difference in Decimal Places Reported for Performance Measures . . 96
6.3 Control Delay Criteria for Level of Service . . . . . . . . . . . . . . . . 97
6.4 Canal St Site: Southbound 25th St Approach - Uncalibrated . . . . . . 98
6.5 Canal St Site: Southbound 25th St Approach - Calibrated . . . . . . . 99
6.6 De Pere Site: Eastbound Main Ave Approach - Uncalibrated . . . . . 101
6.7 De Pere Site: Eastbound Main Ave Approach - Calibrated . . . . . . . 102
6.8 De Pere Site: Northbound Broadway St Approach - Uncalibrated . . 103
6.9 De Pere Site: Northbound Broadway St Approach - Calibrated . . . . 104
xii

6.10 Graphical Rating Scale for Technical Accuracy . . . . . . . . . . . . . 105

6.11 Software and Model Technical Accuracy Summary . . . . . . . . . . . 106
6.12 Graphical Rating Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.13 Software Interface Evaluation . . . . . . . . . . . . . . . . . . . . . . . 122
6.14 Software Data Input Needs . . . . . . . . . . . . . . . . . . . . . . . . 123
6.15 Major Features for Software Comparison . . . . . . . . . . . . . . . . 124
6.16 Comparison of Advanced Secondary Features . . . . . . . . . . . . . 126
6.17 Installation And Licensing Requirements Comparison . . . . . . . . . 127
6.18 Software Training and Support Availability . . . . . . . . . . . . . . . 128
 1

1 Introduction

Roundabouts have been used worldwide due to their contribution to safety

and operational efficiency, as well as other side benefits such as speed control
and reduced operational costs. Some European countries such as the United
Kingdom (U.K.), Germany, and France, as well as Australia, have a long history
of roundabout research while experience in the United States is varied and
relatively recent. How well worldwide experience translates to U.S. roundabout
implementations is still a topic of active research because driver populations,
expectations, and roadway geometry vary significantly from country to country.
This research provides insight and comparisons of various models and software
developed around the world for operational analysis of roundabouts.
Successful design and implementation of roundabouts largely depends on
communication and quality engineering, but is also influenced by public opinion
and driver education. Recommending roundabouts for the correct reasons and
communicating to the user with good geometry, signs, and markings are all
responsibilities of the engineer in the design phase to achieve smooth traffic
operations and safety. Equally important for success in the design and planning
phases, are accurate estimations of intersection demand and capacity, which
critically provide the basis for all performance measures used in comparing
intersection control alternatives. Worldwide studies on roundabout capacity have
2

led to the development of many mathematical models, based on traffic flow theory,
empirical observations, or a combination of theory and observation, to predict
performance under congested conditions. Software development has grown in
parallel with capacity research to provide an interface to the models for analysts.
With a multitude of options available, understanding the underlying models and
principles of each software aids informed decision making in order to provide the
most value for a successful, safe, and efficient transportation system.

1.1 Problem Statement

Traffic volumes continue to grow while the available lane-miles remain relatively
steady, creating a need for ways to allow more vehicles in an already congested
transportation system. Fairly evaluating intersection alternatives hinges on ap-
propriate analysis methods. As roundabouts continue to be recommended as
an intersection alternative for safety and operational reasons, engineers need to
have confidence that they are analyzing roundabouts appropriately. Operationally,
engineers in the U.S. and particularly Wisconsin, have been using the roundabout
capacity model based on Kimber’s research from the 1970s and ’80s done in
the U.K for design and analysis. Here, referred simply as the U.K. capacity
model, Kimber showed that capacity can be predicted as a function of roundabout
geometry. Some concern has been expressed about the validity of applying the
U.K. model in the U.S. where driver behaviors, expectancies, and geometries
may differ. Such concerns have led many agencies to undergo evaluations of the
multitude of software packages to identify which best fit their needs.
In design, a desirable solution needs to balance constraints and should not
be over or under designed. Choosing unsuitable software, or misusing appro-
 3

priate software, could lead to poor estimations of capacity. Further, inaccurate

estimations of capacity could lead to undesirable consequences of decreased safety
or increased delay and queuing. As traffic volumes continue to increase system-
wide, research can provide insight about how to make the best use of roundabouts
as part of a diverse toolbox for transportation engineers.

1.2 Objective, Contributions, and Scope of Work

The objective of this thesis was to:

• Compare underlying capacity models within several roundabout analysis

software packages which included:

◦ RODEL version 1.9.7

◦ ARCADY version 7.1

◦ RCAT version 1.4

◦ Kreisel version 7.0

◦ Girabase version 4.0

◦ HCS 2010 version 6.1

◦ SIDRA version 5.1

• Compare the capacity prediction from each of the above software models to
field data as a means of rating technical accuracy;

• Compare the software in terms of usability and features; and

• Provide a summary of findings in an evaluation matrix consisting of major

categories that rate technical accuracy, usability, licensing type, and cost.
4

Upon completion of these objectives, the findings expanded the understanding

of capacity operations in the U.S. where few roundabouts have experienced
congestion. Concerns addressed within touched on topics of universal and
broad interest. Evaluating and comparing software packages helps researchers,
practitioners, and software developers around the world understand the current
state-of-the-art in analysis tools and better understand future needs.
The scope of this research was constrained to where congestion was
sufficient to cause direct measurements of capacity, which resulted in analyzing
two Wisconsin roundabouts. One multilane and one single lane roundabout
were used, and the focus was strictly on capacity measurements. Software
comparison and analysis was performed for only these two cases in order to make
direct comparisons. Because of the rapidly changing environment of software
development, only the most widely-used versions identified at the onset of this
thesis were used.

1.3 Document Organization

This thesis is organized into seven chapters that follow a progression from
identifying a purpose and need to findings and conclusions. Chapter 1 introduces
the topic and purpose of this research. Chapter 2 presents relevant definitions
and literature required to understand the analysis performed. Chapter 3 explains
the study design and process followed for achieving results. Chapter 4 describes
the field data collection procedures and introduces the study locations. Chapter
5 presents the analysis of the field data, which formed the basis for the majority
of the findings. Chapter 6 investigates the software analysis that was completed
for the studied locations, as well as a comparison of software usability and
 5

features. Chapter 7 highlights the conclusions and considerations resulting from

this research. References and appendices are provided thereafter.
6

2 Literature Review

Discussion of capacity models as well as background for each software used is

presented in the following literature. A few definitions are presented first to aid
understanding and clarify commonly used terminology.

2.1 Capacity

Capacity, in general, has been defined as the maximum sustainable number

of vehicles to traverse a location within a given time period under prevailing
conditions [1]. For roundabouts, this means that each approach has a capacity
for entering vehicles traversing the yield line. Capacity is dynamic in nature
due to continually varying traffic composition (heavy vehicles, motorcycles, and
bicycles), proportions of turning vehicles, driver population characteristics, and
weather conditions. For example, a roundabout that services nearly all heavy
vehicles at one time of the day could be expected to have a different capacity
during a time when only passenger cars are serviced. Varying conditions are
the reason that capacity must be thought of in terms of what flow rates can
be repeatedly observed during peak periods and not the maximum flow ever
observed [1]. Generally capacity data is based on minute-by-minute counts of
entering vehicles and conflicting (circulating) vehicles for a specific approach or
 7

lane as shown in Figure 2.1. The general trend is that fewer vehicles can enter the
roundabout as the number of circulating vehicles increases.

Figure 2.1. Example Entering-Circulating Capacity Graph Development

2.2 Congestion

Capacity models generally require making minute-by-minute observations during

congestion. These are time periods where the demand volume meets or exceeds
the currently available capacity, resulting in sustained queuing and delay. The
need for observing operations at congestion can be understood by making an
analogy to gathering saturation flow data, which relates to capacity, at a signal. A
signal has a predictable cycle that alternates green intervals which are analogous
to continuous gaps in the circulating traffic in a roundabout, and red intervals
which are analogous to an extended time without gaps in the circulating stream
of a roundabout. Collecting saturation flow data at a signal is relatively straight
forward because one only needs to wait for a queue to develop during a red
interval. At a roundabout, there is no signal to make the alternation of right-of-
way or queuing predictable, and therefore having queuing present throughout the
study period is critical. Such time periods of congestion allow direct observation
8

of at- or near-capacity data and also driver behavior.

2.3 Gap Acceptance

Some capacity models use traffic flow theory related to gap acceptance, of which
two main parameters are critical gap (tc ) and follow-up headway (t f ). Some
research has also used the terms critical headway and follow-up time to represent
the same parameters, but definitions are consistent throughout the literature. Gap
acceptance models have been used for determining capacity at other unsignalized
intersections, such as two-way stop controlled or yield controlled intersections
[1]. Roundabouts and these other unsignalized intersection types share a common
traffic flow theory concept of a priority, or major, traffic stream conflicting a minor,
or entering, traffic stream. Capacity of the entering stream is then a function of how
time gaps between major stream vehicles are distributed and how well the minor
stream utilizes these gaps. The following definitions of critical gap and follow-up
headway further clarify the idea.
If major stream traffic was flowing bumper-to-bumper, with little time
between vehicles, the entering stream would not find any acceptable gaps to enter
the roundabout. Thus there must be some minimum acceptable gap in order to
provide any capacity to the minor stream. Critical gap is the minimum amount
of time between circulating vehicles that a driver would find acceptable in order
to safely enter the roundabout [1]. Figure 2.2 illustrates the concept of an entering
vehicle accepting a gap. Only the gaps accepted and rejected by a driver can be
observed; the smallest gap that a driver would accept cannot be directly seen but
can be estimated.
 9

Figure 2.2. Critical Gap Depiction

Several methods exist to estimate critical gap, including use of logit proce-
dures, the Raff method, and the Hewitt method [2]; however, the statistical proce-
dure of the maximum likelihood method was focused on for this research and has
been shown to be accurate and practical [3]. In the field of statistics, the maximum
likelihood procedure is used for estimating parameters of distributions, which can
be applied to finding the mean and variance of a critical gap distribution. For each
driver, two values need to be obtained to apply the procedure: first, the largest
rejected gap, and second, the actual gap accepted. Critical gap is estimated to have
a value between these two observations. Mean and variance for a sample of critical
gap estimates, assuming a log-normal distribution, is determined by numerically
solving the following equations [4]:
10

n
f ( x ) − f (y )
∑ F(yii ) − F(xii ) = 0
i =1

n
( xi − µ̂) f ( xi ) − (yi − µ̂) f (yi )
∑ F ( yi ) − F ( xi )
=0
i =1

where

n = Total drivers observed

xi = Log of the gap accepted by ith driver
yi = Log of the largest gap rejected by ith driver
f( ) = Probability density function assumed, with µand σ2
F( ) = Cumulative density function assumed, with µand σ2
µ̂ = Estimated mean of the critical gap distribution
µ = Mean of the critical gap distribution
σ2 = Variance of the critical gap distribution

Follow-up headway is the amount of time between entering vehicles that

are utilizing the same gap in circulating traffic [1]. Unlike critical gap, follow-up
headway can be directly measured in the field by taking a sample average and
standard deviation [5]. Figure 2.3 shows the concept where multiple entering
vehicles use the same gap. Each vehicle must have been queued in order to
qualify for a true follow-up headway measurement. The follow-up headway
process at a roundabout is analogous to the discharge of a queue of vehicles at
a signal and measuring saturation headway. If there were no conflicting vehicles
in the roundabout, capacity would only be a function of the maximum theoretical
frequently that vehicles could enter the roundabout.
 11

Figure 2.3. Follow-up Headway Depection

In general, as critical gap decreases, capacity increases because drivers

are hesitating less by rejecting fewer gaps, or accepting more risk, to enter the
roundabout. The same is true for follow-up headway because more than one
driver could use the same gap, which also increases capacity. These generalizations
hold true assuming that drivers are consistent in rejecting and accepting gaps of
the same length, and that all drivers behave the same in comparison to each other,
which may not always hold, but provide reasonable estimates [1].

2.4 Capacity Models

Understanding and modeling roundabout capacity has been researched world-

wide because capacity measurements form the basis for operationally comparing
intersection alternatives. Three categories of models have emerged in regards to
what parameters are included in a capacity model: gap acceptance, geometry,
or a hybrid of gap acceptance and geometry. Gap acceptance models predict
12

capacity as a function of critical gap and follow-up headway driver behavior

parameters. Geometric capacity models show that capacity is correlated to
roundabout geometry such as entry with and inscribed circle diameter. Hybrid
models combine elements of both methods to predict capacity. Another major
division between models is whether or not the model predicts capacity lane-by-
lane or if the prediction is aggregated for the entire approach.
Extensive research has gone into developing all model types, and a sum-
mary of each is presented in Table 2.1, along with the associated software package
used in this study. A brief discussion of the models follows after the table.
Noticeably, three software packages were selected to analyze the U.K. model due to
the fact that this model is the basis for the current standard required in Wisconsin
as described in the Wisconsin Facilities Development Manual (FDM) [6]. What has
been termed the ’WisDOT Adjusted’ U.K. model uses the U.K. model except that
the entry width parameter is restricted to a discrete range of values prescribed in
the FDM [6]. Essentially, this restriction is in the spirit of calibrating the model to
what Wisconsin drivers are expected to use as effective entry width.
 13

Table 2.1. Summary of Capacity Models

Software Capacity Model

Name Calibration Needs
Used Aggregation Parameters

Observe entering and

circulating flows during
U.K. Model [7] Approach Geometry saturated periods, then
RODEL 1.9.7, adjust only the intercept
ARCADY 7.1, of the linear model
RCAT 1.4
WisDOT
Same as U.K. Model, except entry width is limited to a
Adjusted' U.K.
discrete range of values
Model [6]

German Model Gap

Kreisel 7.0 Approach
[8, 9] Acceptance

Hybrid Gap
French Model
Girabase 4.0 Approach Acceptance Observe critical gap and
[10]
and Geometry follow-up headway and
substitute default
parameters with the
observed values
NCHRP 572 / HCS 2010 Gap
Lane-by-lane
HCM 2010 [1, 5] 6.1 Acceptance

Observe critical gap and

follow-up headway and
Hybrid Gap substitute default
SIDRA Standard SIDRA
Lane-by-lane Acceptance parameters with the
Model* [11, 26] 5.1
and Geometry observed values, or
establish calibrated
environment factors

* For capacity graphing, ARR 321 was used to approximate the SIDRA Standard Model
with an environment factor of 1.0

Capacity models for roundabouts have historically started with gap ac-
ceptance theory of unsignalized intersections, and countries worldwide have
undergone various changes of their recommended model. Specifically, the research
14

of Tanner in the 1960’s, Harders in the 1960’s and 1970’s, and Siegloch in the 1970’s,
as well as many others, has been used for the capacity of unsignalized intersections
[12] which has been applied to roundabout capacity models worldwide [1, 5-
11]. Much debate has occurred about the two primary techniques for developing
capacity models: gap acceptance or empirical regression [13-19], each with
advantages and disadvantages, similarities and differences, but more importantly
the focus should be on how to best learn from, and use the various capacity
models. Recently, Troutbeck has been mentioned regarding that “there are no
strong reasons for adopting either technique” [13]. Countries from around the
world have used various capacity modeling techniques at different times; just as
capacity is dynamic, so is the best modeling technique. Germany has used linear
and gap acceptance modeling techniques [8, 13, 20], as well as France [13], and
the U.K. has changed through estimates based on weaving, gap acceptance, and
linear regression [15]. Certainly roundabout capacity modeling has changed in the
U.S. as well, from linear form [21], to early gap acceptance techniques [22] which
continue to be refined as more is learned in the U.S. [5].

2.4.1 U.K. Model

R. L. Kimber with the Transport Research Laboratory in the mid 1970’s worked on
an extensive research project that led to the development of a capacity model for
roundabouts in the U.K. Observations from a multitude of sites provided the data
for developing a capacity model using linear regression. Prediction of entering
flow was found to correlate to circulating flow and six geometric parameters [7]:

• Entry width (e),

• Flare length (l 0 ),
 15

• Approach width (v),

• Entry radius (r),

• Entry angle (PHI, φ), and

• Inscribed circle diameter (D).

Depictions of the six geometric parameters can be seen in Figure 2.4 [6].

Figure 2.4. U.K. Model Geometric Parameters

(Based on WisDOT FDM 11-26-20 Attachment 20.1)

No parameters related to critical gap, follow-up headway, or other driver

behavior parameters are used in the model. This is because the U.K. model linear
regression approach does not explicitly try to recreate a model of driver behavior,
but rather capture the end result of all of the complex factors that affect capacity.
Capacity prediction from the U.K. model is aggregated for the entire approach and
uses the six geometric parameters listed above. The resulting capacity equation,
with geometry parameters as previously described, is:
16



k ( F − f c Qc )
 if f c Qc ≤ F
c=

0
 if f c Qc > F

where

c = Approach capacity (pcu/h)

 
1
k = 1 − 0.00347(φ − 30) − 0.978 r − 0.05
F = 303x2
fc = 0.210t D (1 + 0.2x2 )

0.5
tD = 1+ D −60
1+e 10

e−v
x2 = v+
1 + 2S
1.6(e − v)
S =
l0

The model is very sensitive to the entry width parameter, which like all
other variables is a continuous variable [7]. What has been termed the ‘WisDOT
adjusted’ U.K. model limits the entry width parameter to discrete values to better
predict the actual amount of capacity expected. For software modeling purposes,
single lane roundabout entries are limited to using widths of 4.0 to 4.3 m, two-
lane entries 6.7 to 8.0, and three-lane entries 9.75 to 12.0 m when using the U.K.
model with WisDOT adjustments [6]. Making such adjustments is what is referred
to as using effective geometry that drivers actually use, regardless of the exact
field measured dimensions. Recommended calibration procedures require at least
three periods of 20 minutes each with sustained queues of at least five vehicles
on the approach being calibrated, the details to calibrate the capacity equation are
 17

summarized as follows [7]:

• During the study period, the amount of entering and circulating flow is
gathered on a minute-by-minute basis;

• From the gathered data, average entering flow, Q̄e , and average circulating
flow, Q̄c , are computed;

• Slope computation remains the same from the uncalibrated equation, with
variables as defined previously:

slope = k f c

• A calibrated, local, intercept for the capacity equation is calculated by:

F l = Qe + k f c Q̄c

• The calibrated local capacity, Q̄le ,equation becomes:

Qle = F l + k f c Qc

Notably, following this capacity calibration procedure results in a linear

capacity prediction that is forced through the average entering flow of the field
data, which can be seen in the Qle equation by setting Qc = Q̄c .

2.4.2 German Model

The model presented in the German Highway Capacity Manual (HBS 2001) was
used for this research [8, 9]. This model uses gap acceptance theory with critical
gap and follow-up headway as the main parameters. Capacity prediction is
aggregated at the approach level; however, the number of lanes is an input in
18

the model to allow for the higher capacities seen at multilane roundabouts. The
resulting capacity equation is:

 nc
ne − qc (tc − t f −tmin )

t qc
c = 3600 1 − min e 3600 2
3600nc tf

where

c = Entry capacity for approach (pcu/h)

qc = Circulating flow (pcu/h)
nc = Number of circulating lanes
ne = Number of entry lanes
tf = Follow-up headway, 2.9 s
tc = Critical gap, 4.1 s
tmin = Minimum gap between succeeding circulating vehicles, 2.1 s

No calibration procedure for the capacity equation has been specifically

identified. Instead, on the basis of how gap models in general are calibrated, this
research used the critical gap and follow-up headway values observed from the
field study to represent a calibrated German model. Just as the U.S. Highway
Capacity Manual has been updated to the year 2010 edition, an update for the
German HBS 2001 will be available soon [23].

2.4.3 French Model

Original research from France obtained for this study was only published in
the French language [10], which presented some difficulties. Other literature
[14, 24] has presented French capacity equations; however, these appeared to be
based on Girabase released in 1994. The current version from 1999 used in this
 19

evaluation included new inputs, and consequently the results did not match the
magnitude of capacity predictions from the older formulas. An English version of
the Girabase manual revealed that the capacity model is of the hybrid type with
basis in gap acceptance theory originally from Siegloch, modified to be sensitive
to geometric parameters including entry width, splitter island width, circulating
width, and radius of the central island [25]. Additionally, the model is sensitive to
the environment: urban, rural, or suburban based on the inputs of the Girabase
software package. Capacity predictions are aggregated per approach. Beyond
choosing the environment type, no specific calibration parameters were identified.

2.4.4 NCHRP 572 Model

Published in 2007, NCHRP Report 572 represents the most recent and extensive
evaluation of roundabout capacity in the U.S. Eighteen single-lane and seven
two-lane sites were used to analyze relationships between various parameters
and capacity. The analysis showed that driver behavior appeared to be a more
significant factor in capacity compared to detailed geometric measurements [1, 5].
Regression of field data led to parameters for input into a simple lane based model,
based on the gap acceptance theory of Siegloch’s formula [3]. Two models resulted:
one for a single lane entry, and one for the dominant lane of a two-lane entry. These
equations are:
for a single-lane roundabout

−3 q
c = 1130e−1.0×10 c
20

for the critical lane of a multilane roundabout

−3 q
c = 1130e−0.7×10 c

where

c = Lane capacity (pcu/h)

qc = Circulating flow (pcu/h)

Essentially, the HCM 2010 models can be thought of as interchangeable with

the findings in the NCHRP Report 572 with a few minor distinctions; the model
names are used synonymously throughout this report. The HCM 2010 adopted
the NCHRP Report 572 single lane model, and stated the two-lane entry model as
follows:

−3 q
c R = 1130e−0.7×10 c

−3 q
c L = 1130e−0.75×10 c

where

cR = Capacity for the right entry lane (pcu/h)

cL = Capacity for the left entry lane (pcu/h)

Differences between the right lane and left lane equations are small. Both
have the same intercept of 1130 pcu/h, but not the same slope. Resulting
differences reach a maximum of about 28 pcu/h lower capacity in the left lane
compared to the right lane equation for circulating flows in the range of about 1100
to 1800 pcu/h. Calibration of the capacity formulas can be achieved by entering
 21

custom critical gap parameters into the equations, which affects the “A” and “B”
terms related to both the intercept and slope of the model [1]. The generic capacity
formula for calibration is:
c = Ae− Bqc

where

c = Capacity for an entry lane (pcu/h)

3600
A =
tf
t
(tc − 2f )
B =
3600
qc = Conflicting flow (pcu/h)
tf = Follow-up headway (s)
tc = Critical gap (s)

Calibrating the A and B terms results in changing either or both the intercept
and slope (shape) of the capacity function as seen in the three cases demonstrated
in Figure 2.5. Case i shows the default right-lane capacity equation for a two-lane
roundabout. Lower critical gap values alter the B term and result in flatter slopes
with more capacity throughout the range of data, as in Case ii, while increased
critical gap values result in steeper slopes, as in Case iii. Adjusting follow-up
headway primarily affects the intercept but also slightly changes the slope, again
demonstrated in Case iii.
22

Figure 2.5. Effects of Calibration Parameters for the HCM 2010 Capacity
Equations

2.4.5 Australian Method ARR 321

Based on research from the Australian Research Board, including work from
Akcelik, Troutbeck, and others, the Australian capacity model has evolved from
ongoing studies of many roundabouts. Sensitivity to traffic and geometric
parameters have resulted in a complex lane-by-lane model, but each piece of the
model can be understood through gap acceptance theory. The current version of
SIDRA uses several proprietary functions that are not openly published. As such,
 23

this research uses the capacity formulas by the Australian Road Research Board
report 321 (ARR 321) which is a comprehensive report published that SIDRA has
since expanded upon [11].
Equations for the ARR 321 method can best be understood by making a compari-
son that capacity at a roundabout is analogous to capacity at a signal. At a signal,
capacity (c) is proportional to the saturation flow rate (s) and the ratio between
effective green time ( g) and cycle length (C ), resulting in the equation: c = s( g/C ).
For roundabouts, the g/C ratio is analogous to the effective unblocked time where
vehicles could enter the roundabout, and saturation flow rate is analogous to the
maximum amount of entering flow possible with no conflicting vehicles (3600/t f ).
The ARR 321 capacity equation used in this research is:

 
c = max Qm , f od Q g

with

Qm = min [qe , 60nm ]

3600  qc q c  −λ(α−∆c )
Qg = 1 − ∆c + 0.5βϕc e
β 3600 3600

f od = 1 − f qc ( pqd pcd )

where

c = Capacity for an entry lane (veh/h)

Qm = Minimum capacity of the opposed stream (veh/h)
Qg = Capacity estimate using gap-acceptance method (veh/h)
qe = Entry lane arrival flow (veh/h)
qc = Conflicting flow for entry lane (veh/h)
24

nm = Minimum entering vehicles (veh/min)

β = Follow-up headway (s)
α = Critical gap (s)
nc = Number of conflicting lanes
∆c = Circulating stream intra-bunch headway (s)
2.0 s for nc = 1, 1.2 s for nc = 2, 1.0 s for nc > 2
ϕc = Proportion of unbunched (free) circulating vehicles
λ = Exponential arrival headway distribution model parameter
f od = Origin-destination (O-D) pattern adjustment factor
f qc = A calibration parameter for the O-D pattern effect
pqd pcd = Proportion of the total circulating stream flow that originated from,
and were queued on, the dominant approach (≈ 0.5 to 0.8)

Additionally, several parameters are calculated as follows:

for the proportion of unbunched (free) circulating vehicles:

ϕc = e−2.5∆c qc − δϕc

where

δϕc = Extra bunching (platooning) in the circulating stream,

calculated by SIDRA, using extra bunching data
specified for the approaches
 25

for the exponential arrival headway distribution model parameter:


 ϕc qc /3600 qc 0.98
 for ≤
 1 − ∆c qc /3600 ∆c

3600


λ=

49ϕc qc 0.98


for >


∆c ∆c

3600

for the O-D adjustment calibration parameter:

with a single lane circulating flow



0.04 + 0.00015qc

 for qc < 600



f qc = 0.0007qc − 0.29 for 600 ≤ qc < 1800





0.55
 for qc > 1800

with a single lane circulating flow






 0.04 + 0.00015qc for qc < 600


f qc = 0.00035qc − 0.08 for 600 ≤ qc < 1800





0.55
 for qc > 1800

Further, follow-up headway and critical gap are calculated by:

for the dominant lane, follow-up headway
26


h i
 β0 − qe /qc qc
β0d − β om −


d (qe /qc )max qcm ( β Lm − β om ) for qc ≤ qcm
βd =
 β0


d for qc > qcm

subject to β Lm ≥ β om and qe /qc ≤ (qe /qc )max

where

βd = Dominant lane follow-up headway, adjusted for the ratio

of entering to circulating flows (s)
β0d = Unadjusted dominant lane follow-up headway (s)
= β0o − 3.94 × 10−4 qc
subject to β min ≤ β0d ≤ β max where β min and β max are set
minimum and maximum values of follow-up headway, 1.2
s and 4.0 s respectively
β0o = Follow-up headway at zero circulating flow (s)
= 3.37 − 0.0208Di + 0.889 × 10−4 Di2 − 0.395ne + 0.388nc
subject to 20 ≤ Di ≤ 80
where Di = inscribed diameter (m)
ne = number of entering lanes, excluding
bypasses
nc = number of circulating lanes
β om = Minimum β d for zero circulating flow (β om = 1.8 s used)
subject to β om ≥ β min
qcm = Limit on circulating flow rate above which the follow-up
headway is not adjusted, (qcm = 900 pcu/h used)
 27

β Lm = Value of follow-up headway when qc = qcm in the

unadjusted β0d equation
qe /qc = Ratio of entry flow to circulating flow
(qe /qc )max = Upper limit on the qe /qc ratio (3.0 used)

for the subdominant lane follow-up headway

β s =2.149 + (0.5135β d − 0.8735)rds

subject to β d ≤ β s ≤ β max

where

rds = Ratio of entry lane flows

= qd /qs
where qd = dominant lane entry flow rate
qs = subdominant lane entry flow rate

for the dominant or subdominant lane critical gap


(3.6135 − 3.137 × 10−4 qc − 0.339w L − 0.2775nc ) β

 for qc ≤ 1200
α=

(3.2371 − 0.339w L − 0.2775nc ) β
 for qc > 1200
α
subject to 1.1 ≤ ≤ 3.0 and αmin ≤ α ≤ αmax
β

where

α = Dominant or subdominant lane critical gap, ad or αs

respectively (s)
28

β = Dominant or subdominant lane follow-up headwayβ d

or β s
wL = Average entry lane width (m)
αmin = 2.2 s
αmax = 8.0 s

One aspect seen in the ARR 321 method is the sensitivity to the ratio of entry
flow to circulating flow. This sensitivity helps the model to avoid underpredicting
capacity at low circulating flow rates. With low circulating flow large gaps can
occur that many entering vehicles could take advantage of, which in turn increases
the amount of capacity. The effect of increasing capacity based on the entry
to circulating flow ratio is constrained by limiting the minimum and maximum
values that can be used for the ratio. When performing a capacity study on an
existing roundabout, the ratio can be calculated directly from field data, however
for future roundabouts, an assumption must be made about the extent of the
adjustment.

2.4.5.1 Calibration and Differences Between ARR 321 and SIDRA

Several significant differences between the methods implemented in the current

version of SIDRA and ARR 321 are noted in the SIDRA User guide, some of which
include [26]:

• An introduction of an environment factor as a calibration parameter;

• Revision of follow-up headway, critical gap, and circulating vehicle headway

parameters;
 29

• New functions for the origin-destination pattern effect on capacity as well as

the proportion of bunched (platooned) vehicles;

• Addition of capacity sensitivity to entry radius and entry angle parameters;

and

• Changes for many parameters in terms of default values, boundary condi-

tions, and other modifications to the original equations.

These differences cause the capacity results calculated from ARR 321 to be notably,
but situationally dependent, from output from SIDRA, especially regarding the
environment factor. The environment factor (EF) represents a calibration term
to accommodate local conditions where capacity may be higher or lower than
observed elsewhere, and the choice of an EF value can have a large influence on
capacity. Higher EF values result in lower capacity estimates. Using an uncal-
ibrated scenario with EF = 1.0 reflects typical conditions observed in Australia,
while an EF of 1.2 is recommended in the SIDRA User Guide to better reflect
capacity recently seen in the U.S [26]. Even though the exact capacity function
for SIDRA could not be obtained, Figure 2.6, adapted from the SIDRA User Guide,
allows an approximate comparison between environment factors and the ARR 321
method. Depending on the amount of conflicting flow, capacity can be about 100
to 200 pcu/hr lower for EF = 1.2 compared to EF = 1.0. The ARR 321 method can be
seen as an estimated representation to the uncalibrated SIDRA scenario with EF =
1.0, and should be thought as such when interpreting capacity graphs throughout
this research. For all calibrated scenarios using ARR 321, default critical gap
and follow-up headway parameters were substituted by field values rather than
attempting to approximate an appropriate SIDRA environment factor.
30

Figure 2.6. Comparison of SIDRA Capacity with Environment Factors and the
ARR 321 Method (Adapted from the SIDRA User Guide, Figure 19.4.2)

2.4.6 Summary and Notes on Capacity Model Parameters

Table 2.2 summarizes the parameters in each model equation to give a sense of the
similarities and differences. All models use the amount of circulating flow as an
input, and thus has been omitted from the table. Parameters listed in Table 2.2 are
not necessarily the same as what a user would need to enter in a software package
implementing the models, especially regarding the ARR 321 method, because
some parameters are automatically calculated from broader inputs such as traffic
volumes. Others parameters have default values that only need to be changed
for calibration purposes. Further, some software packages require additional
inputs such as lane lengths and lane configurations. Section 6.6.1 discusses actual
 31

software inputs for implementing each model.

Table 2.2. Summary of Model Equation Parameters

German French NCHRP 572 /

UK (HBS 2001) (Girabase) HCM 2010 ARR 321
 Entry width  Critical gap  Critical gap  Critical gap  Critical gap
 Approach  Follow-up  Follow-up  Follow-up  Follow-up headway
width headway headway headway  Minimum circulating
 Flare length  Minimum  Area type headway
 Entry radius circulating (urban, rural,  Minimum entering flow
 Entry angle headway suburban)  Ratio of entry flow to
 Inscribed  Number of  Entry width circulating flow
circle lanes  Splitter island  Origin-destination affect
diameter conflicting width adjustment factor
 Number of  Circulating  Arrival headway
lanes width distribution factor
entering  Radius of  Proportion of unbunched
central island conflicting vehicles
(platooning effect)
 Inscribed diameter
 Entry lane width

2.5 Other Recent Capacity Studies

Two recent studies, one with data collected from Michigan and one with data
collected from Indiana, have also looked at roundabout capacity [27, 28]. Each
study was presented in the year 2011 at the 3rd International Conference on
Roundabouts. Data from these studies will be compared to this research in Section
5.2.
The first study of Michigan roundabouts focused on two sites, each with
triple-lane entries, as an exploration of how well existing models extend to larger
roundabouts. Conclusions showed that the triple-lane roundabouts analyzed had
significant lane imbalance with the innermost lane servicing the highest amount of
32

traffic, likely due to a downstream lane drop. Extending the HCM 2010 two-lane
equation to these three-lane roundabouts tended to overestimate capacity due to
more conservative driver behavior. Calibration improved the model prediction
[27].
Three single lane entry sites in Indiana were analyzed in terms of capacity
in the second study. In summary, field observations revealed considerably smaller
critical gap and follow-up headway values compared to the default HCM 2010
values, suggested to be the result of potential driver familiarity with roundabouts
in the Carmel, Indiana area. Consequently, the default HCM 2010 model was
found to be conservative and underpredicted capacity [28].

2.6 Software Background

Other transportation software comparison studies have been completed [29-

33], but none for comprehensively evaluating roundabout software, in terms
of usability and performance, were found to be published to date. The most
similar report to this research was on comparing signalized intersection software,
which compared seven software packages, including SIDRA [32]. For roundabout
analysis, this research is timely as other agencies have been investigating the
various models and software packages available [33]. A brief description of each
software package used in this study follows.

RODEL Originally released in 1992, RODEL (ROundabout DELay) was devel-

oped in England by Barry Crown as a way to analyze roundabouts
with the U.K. model [34, 35]. Many state DOTs, including Wisconsin,
currently uses this program as the standard for roundabout analysis as
outlined in the FDM [6]. Version 1.9.7, which uses a DOS interface, was
 33

identified as the most widely used and released version identified at

the beginning of this research.

ARCADY Developed by the U.K.’s Transport Research Laboratory (TRL), AR-

CADY (Assessment of Roundabout Capacity and Delay) also imple-
ments the U.K. model. ARCADY provides many features beyond
capacity modeling as well. With about 30 years of development,
TRL is now on version 7.1 of ARCADY which has been used for this
evaluation [36].

RCAT Roundabout Capacity Analysis Tool (RCAT), copyrighted in 2009 by

Diodos Software, uses Microsoft Excel to implement analysis using the
U.K. model, similar to RODEL and ARCADY [37].

Kreisel Many capacity models from around the world, including some not
discussed here such as the Swedish and Israeli methods, can be
evaluated within Kreisel. but this evaluation focused on using the
German Highway Capacity Manual (HBS 2001) method. An English
interface for version 7.0 of the software was used throughout the study.

Girabase The Center for Studies on Networks, Transport, Urban Planning and
Public Buildings (CERTU) in France, published Girabase software to
implement the French model for roundabout capacity. CERTU was
formed in 1994 from the distillation of two prior French agencies.
Version 4 of this software, released in 1999 with a French interface, was
used for the evaluation [38].

HCS The Center for Microcomputers in Transportation (McTrans) was formed

in 1986 and has ties to the University of Florida as well as the Federal
34

Highway Administration (FHWA) [39]. Many types of transportation

facilities, including roundabouts, can be evaluated in their Highway
Capacity Software (HCS) product. HCS 2010 fully implements the
analysis methods described in the recent release of the HCM 2010.

SIDRA Formally called SIDRA INTERSECTION, this software has evolved

over 30 years of research in signalized and unsignalized intersections
under the guidance of the Australian Research Board and Akcelik &
Associates [26, 40]. Version 5.1 of SIDRA was used for this eval-
uation. In terms of capacity analysis, Some proprietary functions
and parameters based on recent research have been implemented in
SIDRA and are not reflected in the source material used during this
research. Section 2.4.5.1 includes specific details and implications of
the differences between the SIDRA model and the ARR 321 method
used within.

In summary of the literature review, many roundabout capacity models and

software packages are available for analysis purposes. Research and development
from around the world has led to models that successfully incorporate either or
both gap acceptance and geometric parameters. Many countries, including the
U.S. have undergone various model revisions and will likely continue to do so in
the future as research continues to refine understanding. Software has evolved
along with each model and, like most technology, versions can change quickly.
Therefore, any commentary on software is only relevant to the version evaluated.
 35

3 Study Design

A systematic approach was maintained for this research in order to remain as

unbiased as possible. Also, a philosophy of thinking about models before software,
was kept in mind to ensure analyses and comparisons were based on scientific
evidence and engineering judgment. Communication and input from the vendors,
however, occurred throughout the research as a way to build relationships and
gain insight to features and correct use of the software that may have been
otherwise overlooked. Figure 3.1 shows a high level overview of the methodology
used for the evaluation, with each of the seven milestone steps further described in
this section. Key to this research, Steps 2 and 3, drove the science behind the major
findings and fulfilled the objectives of comparing capacity models. Steps 4 and 5
satisfied the objective of comparing software usability. Step 6 fulfilled the objective
of summarizing findings in an evaluation matrix.
36

① Steering Committee Development

Field Capacity Data ② Compare Models to Literature

from Previous Study Field Data Review

Software Calibrated
Vendor Input Scenarios
③ Software Analysis
Uncalibrated
Scenarios

④ Software Usability and

Feature Comparison

⑤ Cost, IT, and Training

Needs Comparison

⑥ Draft Evaluation Matrix

⑦ Steering Committee Feedback

Figure 3.1. Software Evaluation Methodology

Step 1: A diverse steering committee was formed to provide a comprehen-

sive range of perspectives during the evaluation. Members consisted of WisDOT
staff from each region, the central office, and Information Technologies (IT)
department. Transportation consultants were also part of the steering committee,
which allowed needs to be expressed from both the public and private sectors.
The large group also covered the range of a project lifecycle from different
 37

members having expertise in planning, design, and operations. After forming the
committee, a set of needs and outcomes were identified to form the basis for the
evaluation as listed in Table 3.1. Meetings were held at intermediate milestones to
discuss preliminary results and identify future investigations.

Table 3.1. Evaluation Criteria

Category Criteria
Technical Accuracy Difference between software prediction and field observations

Usability List of inputs to reflect data collection and software input intensity
Ease of software use
Listing of advanced features
Training availability and technical support
Other miscellaneous features as needed

IT Requirements Installation Requirements

Licensing Type

Cost Licensing Cost

Step 2: Entering and circulating flow data, which form the basis for capacity
analysis, were compared to theoretical capacity models to help in evaluating
technical accuracy. Microsoft Excel spreadsheets were developed and based
on research papers from the primary sources of each capacity model. Both
uncalibrated and calibrated models were used in the comparison. Calibration
of the capacity model spreadsheets is discussed in Section ??. Graphing entering
flow versus circulating flow from the field data and fitting each theoretical model
allowed root-mean-square error (RMSE) computations for a quantitative estimate
to compare the differences between each model. Section ?? further discusses error
measurement.
Step 3: Software analysis consisted of evaluating default, uncalibrated
scenarios as well as calibrated scenarios. In each case, data input was kept to
38

the essential minimum in order to simulate future analysis where the analyst
would have projected estimates and limited field data. Analyzing from a future
perspective was more useful because much of the practical work in the roundabout
field is focused on design. Essential data input included:

• Peak hour volumes and peak hour factor,

• Percent trucks, and

• Geometric information, if needed for the particular model.

Software calibration was kept consistent with the calibration and comparisons
performed in Step 2, with further details of calibration procedures described in
Chapter 6. Because field data were only collected for certain approaches, only the
studied approaches received calibrated parameters in the software. All other ap-
proaches retained their default parameters. Further, only the capacity model was
calibrated, either by observed gap acceptance parameters or entering-circulating
flow relationships where applicable. No other parameters were changed in the
calibrated scenarios. Software output was recorded for each approach but the
emphasis was placed on the field study approaches. Not all software allowed
for calibration of the models being evaluated, including RODEL 1.9.7, RCAT
1.4, Kreisel 7.0, and Girabase 4.0; these packages were excluded from calibrated
comparisons.
Step 4: While a user can become accustomed to any interface and limi-
tations, the software should not present a barrier to quality analysis. Software
usability was evaluated holistically based upon the experiences during Step 3.
A major complication of evaluating usability are first defining usability and
second dealing with subjective biases. Usability was defined and evaluated by
considering the following points:
 39

• Initial learning curve,

• Logical layout and data input ease,

• User feedback and error prevention features, as well as

• Long-term memory load for infrequent users.

Subjective biases in terms of the presented usability definition would be best

minimized by having multiple evaluators. However, logistical and budget needs
presented a barrier to conducting large scale usability testing. As such, the
usability results should be treated with caution as the results are based on opinion.
Step 5: Each software vendor was given the same list of questions in order to
compare the IT needs and licensing costs. Three questions were asked to complete
this step:

1. What are the installation requirements for the software (CPU, RAM, Operat-
ing System, etc)?

2. What type of licensing requirements and options are available (Standalone,

Network Based, etc) and what is the cost associated with the licensing?

3. What is the availability and cost of support services? Is any training available
and at what cost?

Step 6: As a result of all of the previous steps, findings were summarized in

tables, called evaluation matrices, to allow side-by-side comparisons. The matrices
present preliminary findings to the steering committee. Responsibility was placed
on the steering committee to assign weights to the importance of each feature for
refinements and future decision making. At the time of this writing, weighting
criteria was beyond the scope of this thesis and subject to future research.
40

Step 7: Presenting the draft evaluation matrix to the steering committee

gave an opportunity for all public and private sector stakeholders to review the
work, ask questions, and identify issues that need to be investigated further before
decision making. Again, at the time of this writing, final decisions of the steering
committee were beyond the scope of this thesis and subject to future research.
These seven steps represented a highly exploratory process where certainly
more questions arose throughout the journey. Upon completion of these seven
steps, however, valuable insights were gained in regards to how capacity models
compare, what current software packages provide, and an understanding of what
future steps may be necessary.
 41

4 Data Collection and Site

Descriptions

Data collection proceeded by selecting locations for study, gathering field opera-
tional data, and finally reducing data. Gathering operational data was one part of
a larger comprehensive evaluation of roundabouts, which has formed the basis for
other studies [41]. This section describes only the data collection procedures that
were relevant to gathering the operational parameters for this particular research.

4.1 Site Selection and Descriptions

Roundabout locations were chosen primarily based upon the potential to observe
queued operations. Both multilane (maximum of two entering lanes) and single-
lane roundabouts were considered. Based on the goals of this research, several
sites were identified for inclusion in data collection. Once the field data was
collected, however, only two sites experienced enough queuing for capacity data
analysis and will be described in this section. The two sites will be distinguished
by referring to them as the "Canal St" and "De Pere" sites. Relative locations can be
seen in Figure 4.1. Within each site, the local street names will be used to reference
each roadway approach.
42

De Pere Site

Canal St Site

Figure 4.1. Study Locations

4.2 Canal St Site

Located in Milwaukee, WI the Canal St roundabout is in an urban, industrialized

location, just southwest of the major interchange between I-94 and I-43. Average
daily traffic on the east-west Canal St is over 6000 veh/day [42]. Figure 4.2 shows
the layout of the t-intersection roundabout. Canal St is the major road with two-
lane approaches in the east-west direction. The southbound 25th St is the minor
approach with one entering lane conflicted by two lanes. Extra entry width is
provided for heavy vehicles. Complete intersection geometry details are presented
in Section 5.3. At the time of data collection, lane striping was minimal in the
circulatory roadway, unlike other roundabouts in Wisconsin. Notably, the 25th St
 43

approach experienced the most queuing due to heavy through-movement traffic

on Canal St and was the approach used for this study.

Basemap from DIME, De Pere Interactive Maps

North

Studied
Approach:
SB 25th St
WB Canal St

EB Canal St
26° Entry
172 ft Inscribed
Angle
Diameter
36 ft Splitter
68 ft
Island
Radius

23 ft Entry
SB 25th St Detail

32 ft Circulating
Width

Figure 4.2. Canal St Site Configuration

4.3 De Pere Site

The De Pere site, named in part because of its connection to the De Pere bridge
(formally the Claude Allouez Bridge), is located in the downtown of De Pere,
WI along the Fox River, near Green Bay. De Pere is home to many roundabouts,
and this particular location is the hub of major routes including STH 32 and STH
44

57, as well as CTH G and CTH X for Brown County. There are five bridges to
the north of this roundabout, the nearest being two miles north where STH 172,
which connects I-43 and US-41, crosses the river. To the south, there are no other
bridges crossing the Fox River for 11 miles. Average daily traffic counts from
2009 for the eastbound bridge approach were almost 30,000 veh/day and 21,000
veh/day for the northbound approach [42]. Construction along STH 172 caused
extra traffic to detour through the De Pere roundabout, creating the congestion
necessary for capacity measurements. Every approach has two lanes entering
(assigned through-left and through-right) with two lanes conflicting, as shown in
Figure 4.3. Studied approaches included northbound Broadway St and eastbound
Main Ave (from the bridge). Complete intersection geometry details are presented
in Section 5.3.
 45

Basemap from DIME, De Pere Interactive Maps

North SB Broadway St

174 ft Inscribed
Diameter

Studied Approach: 30 ft circulating width

EB Main Ave (bridge) WB Wisconsin St

Studied
Approach:
NB Broadway
St
28 ft Entry Width
28 ft Entry Width
65 ft Radius
65 ft Radius
23° Entry Angle
25° Entry Angle

EB Main Ave (bridge) NB Broadway St

Figure 4.3. De Pere Site Configuration

4.4 Field Data Collection

Time periods for collection were chosen based upon collecting traffic operations
representative of normal conditions:

• Dry weather conditions,

• Daytime traffic operations, and

• Typical weekday peak operations.

46

At the chosen sites, only the approaches with the most queuing were further
analyzed. Field operational data was obtained by means of video recordings,
typical of the setup shown in Figure 4.4. High definition cameras were set up
to observe the studied approach and corresponding exit of the major movement.
Camera placement was as close to the roadway as possible, typically 200 to 400 ft
away. A MiovisionTM proprietary fish-eye camera was set up to observe the central
island and circulating traffic. Mounting and location of each camera was done in
such a way to minimize disruption to traffic and not cause a distraction to drivers
during recording.

Video Camera Video Camera

~20 ft high ~20 ft high
mounted to steel pole mounted to steel pole
secured to sign or utility pole secured to sign or utility pole

200 - 400 ft 200 - 400 ft

MiovisionTM Unit
~25 ft high
Tripod mounted

Figure 4.4. Typical Camera Setup for Recording Traffic

(Adapted from Reference 41)

All video recording was done during good weather and daytime hours,
typically between 12:00 PM and 6:00 PM for PM peak observations and 6:30 AM
to 9:00 AM for AM peak observations. Videos from each camera were combined
 47

in post-processing using Sony Vegas 9.0 to obtain a single synchronized video for
data reduction as shown in Figure 4.5.

Exit Video Camera

Entrance Video Camera MiovisionTM Video Camera

Figure 4.5. Sample Synchronized Video Screenshot used for Data Reduction

(Adapted from Reference 41)

Geometry measurements for each roundabout were obtained from as-built

construction plans or from scaled aerial photos imported into CAD software, in
accordance with the original documentation for each model. Measurements for all
approaches were recorded in order to provide the necessary inputs for operational
modeling in the various software packages. Complete geometry details for each
intersection are presented in Section 5.3.

4.5 Field Data Reduction

Reduction of the video data occurred through the use of software developed at
the University of Wisconsin Traffic Operations and Safety (TOPS) Laboratory [43].
48

Figure 4.6 shows a screenshot of the software interface.

Figure 4.6. Time Stamp Extraction Software User Interface

Essentially, the software allows the user to manually record timestamps

while watching a synchronized video by typing a letter correspond to specific
events for each vehicle. The letters shown in each lane of Figure 4.6 coincide with
the following events [41]:

• Upstream approach time (letters A/B) – a time stamp was recorded when
vehicles passed a point of known distance upstream of the roundabout;

• Back of queue time (letters C/ D) – a time stamp was recorded when vehicles
joined the back of a queue; First in server time (letters E/F) – a time stamp
was recorded when vehicles arrived at the first in server or first in queue
position;
 49

• Entry time (letters G/H) – a time stamp was recorded when vehicles began
their entrance in to the roundabout;

• Exit time (letters I/J, K/L, M/N) – a time stamp was recorded when vehicles
exited the roundabout;

• Downstream exit time (letters O/P) – a time stamp was recorded when
vehicles making a through movement passed a point of known distance
downstream of the roundabout;

• Conflicting vehicle time (letters Q/R) – a time stamp was recorded when
circulating vehicles passed a point perpendicular to the yield line;

• Exiting vehicle time (letters S/T) – a time stamp was recorded when circulat-
ing vehicles exited the roundabout at the approach in question.

Collecting timestamps was done to the maximum extent possible, limited by

the camera field of view. For tracking queuing, times were noted when the
queue extended beyond the camera, which was especially prevalent at the De
Pere site. From these timestamps, gap acceptance parameters, entry flow, and
conflicting flow data could be identified. Critical gap and follow-up headway
were derived in a manner similar to that of "Method 2" used in NCHRP 572,
where vehicles needed to have rejected a gap in order for inclusion in the data
set [5]. Additionally, turning movement counts were obtained by means of the
MiovisionTM data reduction service.
50

5 Capacity Analysis

This chapter presents the data and analysis for the capacity study which forms the
basis for model comparisons.

5.1 Observed Queuing Data

Reduction of the video data collection resulted in a number of one-minute intervals

that were fully queued in order to evaluate capacity. Queues had to be at least five
vehicles long throughout the entire minute in order to qualify as fully queued to
be compatible with previous research [7]. Table 5.1 summarizes the number of
observations made during the PM peak studies. Times of day shown reflect the
hours observed. Also an approximate number of the total subset of minutes used
for capacity analysis within the observed hours is listed. The Canal St site was
characterized by steady but sometimes sporadic queues during the study period,
and resulted in one data set for the studied single-lane approach. Three different
data sets for each studied approach resulted for the two-lane De Pere site:

• A data set for whenever the left lane was queued (queuing may or may not
have been present in the right lane);
 51

• A data set for whenever the right lane was queued (queuing may or may not
have been present in the left lane); and

• A data set for when both the left lane and right lane were queued.

The De Pere site had more consistent queues in the PM peak period compared
to the Canal St site. Especially the northbound Broadway approach which
experienced queues in the left lane for 268 minutes of the 300 minutes of video
data. Due to limited resources, data for an AM peak period was only collected
for the De Pere site and is shown in Table 5.2. No queuing was observed
on the eastbound (bridge) Main Ave approach during the AM peak, but the
northbound approach still exhibited queuing. Unless otherwise specified, all
entering-circulating graphs, gap acceptance parameters, and other comparisons
are based on PM peak data. AM peak data followed similar trends and is
summarized in Appendix A. An exploration of combining AM and PM peak data
is presented in Section 5.5.2.
52

Table 5.1. Summary of Observed PM Peak Minutes of Queuing

PM Peak
Number of One-
Minute Queuing
Studied Approach Intervals
Canal St Site
Thursday April 15, 2010
Between 1:30 pm to 6:00 pm
SB 25th St
71
[Out of approx. 250 min]

De Pere Site
Wednesday May 19, 2010
Between 11:30 am to 6:30 pm
EB Main Ave (bridge)
[Out of approx. 200 min]
Left Lane 82
Right Lane 125
Both Lanes 66
NB Broadway St
[Out of approx. 300 min]
Left Lane 268
Right Lane 77
Both Lanes 76
 53

Table 5.2. Summary of Observed PM Peak Minutes of Queuing

AM Peak

Number of One-
Minute Queuing
Studied Approach Intervals
De Pere Site
Thursday May 20, 2010
Between 6:30 am to 8:50 am
EB Main Ave (bridge)
[Out of approx. 120 min]
Left Lane 1
Right Lane 2
Both Lanes 1
NB Broadway St
[Out of approx. 120 min]
Left Lane 84
Right Lane 27
Both Lanes 24

In addition to one-minute queued intervals, peak hour turning movement

counts and percentages of heavy vehicles were obtained from the MiovisionTM
video and data reduction. Consecutive one-minute turning movement counts
for the entire data collection period were aggregated into 15-minute counts, of
which the four highest consecutive fifteen minute periods was considered the peak
hour. Heavy vehicle percentages were determined for each roadway approach
by dividing the number of heavy vehicles counted in the peak hour by the total
number of vehicles counted for the hour. Figure 5.1 shows the resulting vehicle
counts for the Canal St site with the peak hour occurring between 3:30 PM and
4:30 PM. Noticeably, the eastbound and westbound through movements were
dominant and resulted in the queuing on the southbound approach used for this
research.
54

Figure 5.1. Canal St Site Peak Hour Volumes

Figure 5.2 shows the resulting vehicle counts for the De Pere site for the peak
hour between 5:00 PM and 6:00 PM. Heavy northbound and eastbound volumes
were causal factors for the queuing and lane utilization patterns observed.

Figure 5.2. De Pere Site Peak Hour Volumes

 55

5.2 Observed Gap Acceptance Data

Gap acceptance parameters were obtained from the data collected during the
PM study periods in a manner consistent with the NCHRP 572 study using the
maximum likelihood method and only considering vehicles that rejected at least
one gap before accepting a gap in order to be included in the data set. Table 5.3
and 5.4 show the obtained critical gap and follow-up headway parameters as well
as comparisons to the field data from the NCHRP 572 sites. "Method 2" gap
acceptance values from the NCHRP 572 study, where vehicles needed to have
rejected gap, were used for comparison due to the alike data collection used in
this research. Both the single lane and multilane NCHRP data are shown on
Southbound 25th St approach because the roundabout is a single-lane site but
has characteristics of a multilane site due to the presence of two conflicting lanes.
Standard deviations for all measurements are shown in parenthesis.

Table 5.3. Canal St Site Gap Acceptance Parameters

SB NCHRP 572 NCHRP 572

25th St Single-lane Multilane
Critical gap (s) 5.5 (2.0) 5.0 (1.2) 4.5 (1.7)
Follow-up
2.6 (1.4) 3.2 (1.1) 3.1 (1.1)
headway (s)
56

Table 5.4. De Pere Site Gap Acceptance Parameters

NB EB
Critical Gap (s) Broadway St Main Ave (Bridge) NCHRP 572
Left lane 4.1 (1.0)* 4.4 (1.6) 4.8 (2.1)
Right lane 3.4 (1.0) 4.3 (1.4)* 4.3 (1.5)
Approach 3.8 (1.1) 4.3 (1.5) 4.5 (1.7)

Follow-up NB EB
Headway (s) Broadway St Main Ave (Bridge) NCHRP 572
Left lane 3.1 (1.3)* 2.8 (1.2) 3.2 (1.1)
Right lane 3.0 (1.2) 2.8 (1.1)* 3.0 (1.2)
Approach 3.0 (1.2) 2.8 (1.1) 3.1 (1.1)
*Data point from dominant lane on the subject approach

Data from the Canal St site showed higher critical gap values than either
the NCHRP single or multilane sites, but follow-up headway was about 0.5 s less
than the NCHRP sites. The recent single-lane roundabout study in Indiana found
gap acceptance values much lower than the Canal St site, with critical gap ranging
from 3.39 to 3.79 s and follow-up headway ranging from 2.10 to 2.43 s [28].
Data from the De Pere site showed lower or equal critical gap and follow-
up headway, ranging from 0.0 to 0.7 s lower in all cases compared to the NCHRP
sites. Standard deviation from all sites in this study ranged from 1.0 s to 2.0 s
which was similar to the range 1.1 s to 2.1 s in the NCHRP study. A reoccurring
trend from two-lane entry roundabout studies was also observed at the De Pere
site: the right lane tends to have a lower critical gap value than the left lane,
perhaps because drivers making a right turn maneuver feel somewhat protected
by vehicles entering the roundabout in the left lane.
Comparing to the three-lane entry roundabout study in Michigan, gap
acceptance parameters for the dominant lane at the Michigan sites were higher
with their critical gap reported as 4.66 s and follow-up headway as 3.34 s [27].
 57

5.3 Measured Site Geometry

Geometry from each site was found from construction plans or scaled aerial
photos. The results used for the spreadsheet capacity modeling, as well as software
analysis are shown in Table 5.5 for the Canal St site and Table 5.6 for the De Pere
site. Metric measurements are required for model input and are shown along with
U.S. customary units. The WisDOT adjusted U.K. model used 4.3 m for any single
lane entry and 8.0 m for any two-lane entry, overriding the actual measured widths
shown. One assumption was made for the De Pere site on the eastbound approach:
no flare was observed because the bridge was the same or slightly wider than the
entry width. For modeling, this meant that the half-width parameter needed to
be modified to 8.0 m to match the entry width to satisfy the requirement that half-
width cannot exceed entry width.
58

Table 5.5. Canal St Site Geometry and Characteristics

Parameter SB 25th* EB Canal WB Canal

E - Entry width
7.01 m (23 ft) 8.53 m (28 ft) 8.53 m (28 ft)
(actual width measured)

E - WisDOT Adjusted 4.3 m (14 ft) 8.0 m (26 ft) 8.0 m (26 ft)

l' - Effective flare length 15.85 m (52 ft) 21.95 m (72 ft) 39.32 m (129 ft)

V - Approach road half-width 4.27 m (14 ft) 7.32 m (24 ft) 7.32 m (24 ft)

R - Entry radius 20.73 m (68 ft) 28.35 m (93 ft) 22.25 m (73 ft)

PHI - Entry Angle 26° 39° 10.5°

D - Inscribed circle diameter 52.43 m (172 ft) 52.43 m (172 ft) 52.43 m (172 ft)

Splitter Island Width 10.9 m (36 ft) 9.7 m (32 ft) 9.7 m (32 ft)

Approach Speed 30 mph 30 mph 30 mph

*denotes studied approach

 59

Table 5.6. De Pere Site Geometry and Characteristics

Approach

EB Main NB
Parameter (Bridge)* Broadway* WB Wisconsin SB Broadway
E - Entry width
(actual width 8.53 m (28 ft) 8.53 m (28 ft) 9.14 m (30 ft) 8.53 m (28 ft)
measured)

E - WisDOT
8.0 m (26 ft) 8.0 m (26 ft) 8.0 m (26 ft) 8.0 m (26 ft)
Adjusted

l' - Effective flare 23.01 m (75.5

0 m (0 ft) 12.19 m (40 ft) 32.31 m (106 ft)
length ft)

V - Approach road
8.53 m (28 ft) 7.32 m (24 ft) 7.01 m (23 ft) 7.01 m (23 ft)
half-width

R - Entry radius 19.81 m (65 ft) 19.81 m (65 ft) 29.87 m (98 ft) 19.81 m (65 ft)

PHI - Entry Angle

25° 23° 24° 21°
(deg)

D - Inscribed circle 53.04 m 53.04 m 53.04 m 53.04 m

diameter (174 ft) (174 ft) (174 ft) (174 ft)

Splitter Island
7.3 m (24 ft) 7.3 m (24 ft) 15.6 m (51 ft) 7.3 m (24 ft)
Width

Approach Speed 25 mph 25 mph 25 mph 25 mph

*denotes studied approach

5.4 Capacity Data Analysis

In order to compare collected field data to capacity models, spreadsheets were

developed based on the original literature explaining each model. Such compar-
isons give insight into explaining software results. The difference between capacity
spreadsheets and software output is that software only analyzes one entering-
circulating data point at a time based on traffic volume input and any assumed
60

interaction between traffic volumes on other approaches. Capacity spreadsheets

on the other hand, analyze multiple entering-circulating data points at a time and
compare them directly to field observations which already reflect any interaction
between the approaches.
All spreadsheets used equations directly from the original research in each
model. Four other important items related to the capacity spreadsheet analysis are:

1. Because of the difficulties in obtaining detailed English documentation

for the capacity model currently used in Girabse as mentioned in Section
2.4.3, a spreadsheet could not be adequately developed to replicate results
compatible with the Girabase software. Therefore further discussion of the
French model has been omitted from all applicable capacity graphs. 2.4

2. Results from the U.K. model reflect the ’WisDOT adjusted’ entry width
unless otherwise stated.

3. The ARR 321 method is sensitive to the ratio between the entry flow and
circulating flow. Because data was collected for both these flows, the ratio
could be calculated precisely. However, to be consistent with all other models
and the fact that SIDRA defaults to a "medium" level of adjustment, this
parameter was estimated at an average value of 1.5, based on the maximum
and minimum allowable range, rather than computed from field data. These
assumptions allow the uncalibrated ARR 321 method to approximate SIDRA
with an environment factor of 1.0 as discussed in Section 2.4.5.1.

4. One-minute entering and circulating counts were converted to passenger car

equivalents, with heavy vehicles counting as 2.0 passenger cars, and motor-
cycles and bicycles counted as 0.5 passenger cars. Equivalent passenger car
 61

hourly flows (pcu/hr) were obtained by multiplying one-minute counts by

60 minutes/hour.

5.4.1 Calibration

For the spreadsheet analysis, models were calibrated based on the collected gap
acceptance and entering-circulating field data. Each roadway approach studied
was calibrated separately. Gap acceptance model equations shown in Section
2.4 including the German, NCHRP, and ARR 321 methods, were calibrated by
replacing the default critical gap and follow-up headway values with those
obtained by the field data from Table 5.3 and 5.4. As a feature of the ARR 321
method, there is no environmental factor as part of the equations and therefore
no approximations were possible to establish what environmental factor would be
most appropriate for SIDRA software analysis.
Calibration of the linear U.K. model followed the procedure outlined in
Section 2.4.1 where only the intercept was changed based on average entering-
circulating field data. Observed averages are shown in Table 5.7. Resulting
calibrated and uncalibrated intercepts follow in the discussions contained in
Section 5.4.4 for the Canal Site and Section 5.4.5 for the De Pere site.
62

Table 5.7. Average Entering and Circulating Flows Used for U.K. Model
Calibration

Average Approach Average Approach

Studied Roadway Entering Flow Circulating Flow
Approach (pcu/hr) (pcu/hr)
Canal St Site:
SB 25th St 538 812
PM Peak

De Pere Site:
EB Broadway 1900 366
PM Peak

De Pere Site:
NB Broadway 788 1334
PM Peak

De Pere Site:
NB Broadway 1099 963
AM Peak

As a starting point for calibration of the U.K. model, field measured

geometry was used to remain independent of any ‘WisDOT adjusted’ parameters
not specified in the original model documentation. Doing so resulted in slightly
larger slopes, which improved the fit to field data, but had negligible impact on
overall capacity results. A comparison of the different slopes and intercepts from
using field measured versus effective geometry is presented in Section 5.5.1.

5.4.2 Error Measurement

One common method of comparing statistical models is by computing the root-

mean square-error for each model. RMSE is an estimate of precision and represents
the average difference between the model prediction and observed data. To make
a fair comparison between approach based models and lane based models, RMSE
was divided by the number of lanes, resulting in estimates of RMSE per lane, with
 63

units of pcu/hr/ln. Formulaically, RMSE was determined by:

p
(Model Prediction − Field Observation)2 /n
RMSE =
Number of Lanes

where

Model Prediction = entering flow predicted for a given circulating

flow
Field Observation = entering flow observed from a one-minute
entering-circulating capacity datapoint
n = Number of observations

Lower errors indicate a better fit of the model to the data. There is no rule
about what a "good" RMSE value is, goodness depends on how precise the model
needs to be. Typically, a model with lower RMSE is chosen as the best model,
but understanding why the model shows a fits to the data and any underlying
assumptions need to be considered in choosing a model. A model could show
a good prediction of capacity but use parameters that do not actually have any
causal effect on capacity; that is to say correlation does not imply causation.

5.4.3 Model Characteristics

Trends in the field entering versus circulating data were highlighted by performing
simple linear regression. Linear regression models reveal characteristics such
as the maximum capacity and how rapidly capacity decreases with increasing
circulating flow. Regression was performed not in an attempt to develop a new
capacity model, but rather for comparative purposes only. Two models were fit,
64

with the first being:

Qe = A + BQc

where

Qe = Entering flow (pcu/hr)

Qc = Circulating flow (pcu/hr)
A = Intercept constant
B = Slope constant

and the second linear model:

ln( Qe ) = A + BQc

which can be re-expressed as an exponential relationship in the form of:

Qe = Ae BQc

where

Qe = Entering flow (pcu/hr)

Qc = Circulating flow (pcu/hr)
A = Intercept constant
B = Slope constant

These two models were chosen to be fit to the field data collected because
of their similarity to the forms of existing capacity models, for example the U.K.
model exhibits the first linear form shown above, and the NCHRP 572 model is in
the form of the second model shown above. Further, such regression techniques
 65

were used in the development of both the U.K. and NCHRP 572 models [5, 7].
Slope and intercept terms in the linear models describe the maximum
capacity and rate of decrease toward minimum capacity, respectively. The larger
the slope, the less conflicting flow is needed to reach minimum capacity. For
the exponential relationship in gap acceptance models, the slope is constantly
changing, but the slope can be represented by the constant term within the
exponent. The larger the exponent constant, the larger the rate of change in the
capacity prediction. As a general rule, lower intercepts and higher slopes are an
indication of lower capacity predictions. Slope is not readily determinable for the
ARR 321 method due to the piecewise nature of the capacity function, resulting in
slopes shown as "N/A" in the following model characteristics tables.

5.4.4 Canal St Site Capacity Analysis

The Canal St site analyzed the single lane approach of southbound 25th St. Most
of the observations were during periods of medium conflicting flow of about 600
to 1000 pcu/h. Queues were typical on this minor street approach because of the
heavier through movements on the major street. Field data, along with linear and
exponential regressions, are shown in Table 5.8 and Figure 5.3.

Table 5.8. Canal St Field Data Regression Results

Intercept RMSE
Regression (pcu/h) Slope n R2 (pcu/h/ln)

Linear 829 0.359 71 0.367 67

Exponential 869 6.35×10-4 71 0.314 69

 66
Field Data Field Data
2400 Linear Regression 2400 Linear Regression

2000 2000

Entering Flow (pcu/h)

Entering Flow (pcu/h)

1600 1600

1200 1200

800 800

400 400

0 0
0 400 800 1200 1600 2000 2400 0 400 800 1200 1600 2000 2400
Conflicting Flow (pcu/h) Conflicting Flow (pcu/h)
(a) Field Data with Linear Regression (b) Field Data with Exponential Regression

Figure 5.3. Canal St Field Capacity Data

 67

Table 5.9 and Figure 5.4 compare the approach based models for the Canal
St site. The WisDOT adjusted U.K. model predicted a higher capacity than the
average observed data with a RMSE value of 381 pcu/h. Capacity prediction
from the default German model was closer to the observed data and slightly on
the upper end, with an RMSE value of 193 pcu/h/ln. Calibrating both the U.K.
and German models provided similar fits to the data, with the German model
predicting a higher intercept.

Table 5.9. RMSE and Model Characteristics from the Canal St Approach Based
Analysis

Uncalibrated Calibrated

Intercept RMSE Intercept RMSE

Model (pcu/h) Slope n (pcu/h/ln) (pcu/h) Slope n (pcu/h/ln)
U.K. 1323 0.532 71 381 1055 0.631 71 156

German 1241 1.53×10-4 71 193 1385 5.83×10-4 71 183

Overprediction from the U.K. model may be due to the periodic, but
steady, queuing that was observed rather than having longer sustained queuing
indicative of at-capacity operation that the original model was formulated from.
Intercept values from the regression analysis were also smaller than those seen in
the NCHRP 572 research, suggesting that the site may not have been operating
continuously under capacity conditions.
 68
# UK (WisDOT Adj.) Calibrated UK
2400 Default German 2400 Calibrated German
Field Data Field Data

2000 2000

Approach Entering Flow (pcu/h)

Approach Entering Flow (pcu/h)

1600 1600

#
1200 # 1200
##
##
##
##
##
##
800 ## 800
#
##
#
#
400 400

0 0
0 400 800 1200 1600 2000 2400 0 400 800 1200 1600 2000 2400
Conflicting Flow (pcu/h) Conflicting Flow (pcu/h)
(a) Uncalibrated Approach Based Models (b) Calibrated Approach Based Models

Figure 5.4. Canal St Approach Based Capacity Comparison

 69

Lane based models for the Canal St site are shown in Figure 5.5 and
Table 5.10. For both the ARR 321 and NCHRP 572 results, discrepancies appear to
be exaggerated for low circulating flows due to lack of congested observations. The
default ARR 321 method predicted capacity at the upper end of the observed data,
while the default NCHRP 572 model was closer to the observation averages with
RMSE of 310 and 153 pcu/h/ln, respectively. Calibration resulted in the models
being nearly indistinguishable in terms of slope and intercept, which gave similar
RMSE values of approximately 175 pcu/h/ln.

Table 5.10. RMSE and Model Characteristics from the Canal St Lane Based
Analysis

Uncalibrated Calibrated

Intercept RMSE Intercept RMSE

Model (pcu/h) Slope n (pcu/h/ln) (pcu/h) Slope n (pcu/h/ln)
ARR 321 1494 N/A 71 310 1351 N/A 71 173

NCHRP 1130 1.00×10-3 71 153 1165 1.17×10-3 71 178

 70
Default ARR 321 Calibrated ARR 321
2400 Default NCHRP 2400 Calibrated NCHRP
Field Data Field Data

2000 2000

Entering Flow (pcu/h)

Entering Flow (pcu/h)

1600 1600

1200 1200

800 800

400 400

0 0
0 400 800 1200 1600 2000 2400 0 400 800 1200 1600 2000 2400
Conflicting Flow (pcu/h) Conflicting Flow (pcu/h)
(a) Uncalibrated Lane Based Models (b) Calibrated Lane Based Models

Figure 5.5. Canal St Lane Based Capacity Comparison

 71

5.4.5 De Pere Site Capacity Analysis

Two roadway approaches at the De Pere site were analyzed during AM and PM
peak periods. Data from the PM analysis is presented in this section, with AM
results having similar trends which are shown in Appendix A. Queuing was
consistent on both of these approaches in the PM peak (268 of 300 min queued
for the northbound critical lane, 125 of 200 min queued for the critical eastbound
lane) allowing for capacity observations. For lane based models, only the critical
lane was analyzed due to the interest in analyzing high volume operations. First,
the PM peak data from northbound Broadway St was analyzed which showed
characteristics of high circulating flows between 1100 and 1800 pcu/h. Figure 5.6
and Table 5.11 show the field capacity data along with regressions of the data in
linear and exponential form. Data at low circulating flow rates is more indicative
of an actual intercept for a capacity model. Because no such data was observed on
the northbound approach, the actual intercept may differ from the value obtained
by regressing only data from the higher circulating flow rates.

Table 5.11. De Pere Northbound PM Field Data Regression Results

Intercept RMSE
Lane Regression (pcu/h) Slope n R2 (pcu/h/ln)
Linear 832 0.286 77 0.306 100
Right -4
Exponential 1031 6.47×10 77 0.281 101
Linear 895 0.471 268 0.53 109
Left -3
Exponential 1303 1.07×10 268 0.475 111
Approach Linear 1689 0.676 76 0.494 80
(Both
Lanes) Exponential 2349 8.50×10-4 76 0.454 79
 72
R Right Lane Data R Right Lane Data
2400 L Left Lane Data 2400 L Left Lane Data

B Both Lane Data B Both Lane Data

Right Lane Linear Regression Right Lane Exp. Regression

2000 2000
Left Lane Linear Regression Left Lane Exp. Regression

Both Lane Linear Regression Both Lane Exp. Regression

1600 1600

Entering Flow (pcu/h)

Entering Flow (pcu/h)

B B

B B

B B
1200 1200
B B B B B B
B B B B
B B B B B B
B B B B B B
B B B B B B B B
L L B B B B L L B B B B
L B B L B B
B B B B B B B B
800 L
L L L L L BB BBB
B
B 800 L
L L L L L BB BBB
B
B
B
B
L L L RL R L B B B B B B L L L RL R L B B B B B B
B B B B
R
L L L LR R R B B B B R
L L L LR R R B B B B
L R R
B L R R
B
L L L L L LL L L R R RB B B B L L L L L LL L L R R RB B B B
L L L L L L L L L L
L L R
L LRL L B L R
L R RL L BL B L L R
L LRL L B L R
L R RL L BL B
L L R R R R B L L R R R R B
L R
L L L L L R B
L R R
L LR
L B R
B B B R B L R
L L L L L R B
L R R
L LR
L B R
B B B R B
L R L L R L
L L L L LRR R
LLL R
L L RLRR
L LR
L R R
B B
400 L
L L L L LRR
L
R
LLL R
L L R
L L RLRR
L L L LR
L LR
R R L R L
L R
L
L LR
L
R
R
L RR
R
B B
L RR
400 L L L L R
R R L R L
L L L LR
L R
L LR
L R
L RRL RR
L L
L L L L L LRL R
L RL R
L B
LLL RL R R
L L L L L L LRL R
L RL R
L B
LLL RL R R
L
L L R
L L L L L LRL L R
L R
L L L L L R
L L L L L LRL L R
L R
L L L
L L L L
L L L L L R L L L L L L L R L L
L L L L L L L L L L L L L L L L L L

0 0
0 400 800 1200 1600 2000 2400 0 400 800 1200 1600 2000 2400
Conflicting Flow (pcu/h) Conflicting Flow (pcu/h)
(a) Field Data with Linear Regression (b) Field Data with Exponential Regression

Figure 5.6. De Pere Northbound PM Field Capacity Data

 73

Approach models for northbound Broadway St are shown in Figure 5.7.

Model error and characteristics and are shown in Table 5.12. The WisDOT adjusted
U.K. model predicted a higher capacity than the observed data with a RMSE of
347 pcu/h/ln. German capacity model results showed predictions nearer to the
observed data with a low RMSE of 82 pcu/h/ln. Calibration of both models
lowered the capacity estimates compared to their respective default predictions
and resulted in RMSE averaging at 83 pcu/h/ln.

Table 5.12. RMSE and Model Characteristics from the De Pere NB Approach
Based Analysis

Uncalibrated Calibrated

Intercept RMSE Intercept RMSE

Model (pcu/h) Slope n (pcu/h/ln) (pcu/h) Slope n (pcu/h/ln)
U.K. 2450 0.740 76 347 1801 0.759 76 80

German 2483 1.53×10-4 76 82 2400 5.56×10-5 76 86

 74
# UK (WisDOT Adj.) Calibrated UK
2400 Default German 2400 Calibrated German
Field Data Field Data

2000 # 2000

#
Approach Entering Flow (pcu/h)

Approach Entering Flow (pcu/h)

##
##
##
1600 # 1600
#
#
##
#
#
#
#
##
1200 ## 1200
#
#

800 800

400 400

0 0
0 400 800 1200 1600 2000 2400 0 400 800 1200 1600 2000 2400
Conflicting Flow (pcu/h) Conflicting Flow (pcu/h)
(a) Uncalibrated Approach Based Models (b) Calibrated Approach Based Models

Figure 5.7. De Pere Northbound Approach Based Capacity Comparison

 75

A possible explanation for the overprediction in the U.K. model is a

difference in driver behavior under high circulating flows. The U.K. model was
based on the regression of capacity data that included periods of priority reversal,
where circulating drivers yield to entering vehicles, which would increase capacity
at high circulating flows. Priority reversal was not observed during the data
collection in this research. Another possible explanation for the overprediction
is that the U.K model predicts capacity for all lanes at the entry, but most of the
demand traffic was in the left lane during the times observed. If the right lane
did not have sufficient demand to reflect capacity operations, the approach-based
model would overpredict field data from a partially saturated entry. Exploration
of using the U.K. model on lane-by-lane basis is presented in Section 5.5.2.
Comparatively, a close fit from the German model is likely due to the default
values of critical gap and follow-up headway being near the observed field values.
Figure 5.8 and Table 5.13 show results and comparisons from lane based
models for the critical left lane of northbound Broadway St. The uncalibrated
ARR 321 method predicted capacity above the observed data with a RMSE of 376
pcu/h/ln. Results from the default NCHRP 572 model were lower than the ARR
321 method but near the higher capacity observations. Calibration brought the
ARR 321 method in line with the field data with a RMSE value of 110 pcu/h/ln.
Little changed between the uncalibrated and calibrated versions of the NCHRP
572 model because the gap acceptance values observed were near the default
uncalibrated values.
76

Table 5.13. RMSE and Model Characteristics from the De Pere NB Lane Based
Analysis

Uncalibrated Calibrated

Intercept RMSE Intercept RMSE

Model (pcu/h) Slope n (pcu/h/ln) (pcu/h) Slope n (pcu/h/ln)
ARR 321 1633 N/A 268 376 1133 N/A 268 110

NCHRP 1130 7.00×10-4 268 149 1165 7.38×10-4 268 145

 Default ARR 321 Calibrated ARR 321
2400 Default NCHRP 2400 Calibrated NCHRP
Field Data Field Data

2000 2000
Left (Critical) Lane Entering Flow (pcu/h)

Left (Critical) Lane Entering Flow (pcu/h)

1600 1600

1200 1200

800 800

400 400

0 0
0 400 800 1200 1600 2000 2400 0 400 800 1200 1600 2000 2400
Conflicting Flow (pcu/h) Conflicting Flow (pcu/h)
(a) Uncalibrated Lane Based Models (b) Calibrated Lane Based Models

Figure 5.8. De Pere Northbound Lane Based Capacity Comparison

77
78

Turning to the eastbound (bridge) approach at the De Pere site, Main Ave
experienced low circulating flows between 200 and 600 pcu/h in contrast to the
high circulating flows of the northbound approach. Figure 5.9 and Table 5.14 show
the field data and regressions. Because of the low circulating flow, the intercept
of the regressions may be more reliable than that observed for the northbound
Broadway St approach; however, slope may not due to lack of observations
throughout the entire range of circulating flows (low and high). Intercepts from
the regression of the eastbound approach were slightly higher than that observed
by the NCHRP 572 research.

Table 5.14. De Pere Eastbound PM Field Data Regression Results

Intercept RMSE
Lane Regression (pcu/h) Slope n R2 (pcu/h/ln)
Linear 1193 0.546 125 0.318 126
Right
Exponential 1198 5.44×10-4 125 0.309 126
Linear 1171 0.704 82 0.514 108
Left
Exponential 1207 8.03×10-4 82 0.511 107
Approach Linear 2344 1.212 66 0.472 99
(Both
-4
Lanes) Exponential 2386 6.51×10 66 0.456 98
 R Right Lane Data R Right Lane Data
B B
2400 L Left Lane Data 2400 L Left Lane Data
B B
B B
B B B B B Both Lane Data B B B B B Both Lane Data
B B
B B
B B B Right Lane Linear Regression B B B
Right Lane Exp. Regression
2000 B B B B B B B 2000 B B B B B B B
B B B B Left Lane Linear Regression B B B B Left Lane Exp. Regression
B B B B B B B B
B B
B B B
B B
B
Both Lane Linear Regression B B
B B B
B B
B
Both Lane Exp. Regression
B B B B B B
Entering Flow (pcu/h)

Entering Flow (pcu/h)

B B B B
1600 B B B
1600 B B B
R R
B B
R B B B R B B B
R R B R R B
R L R R
L R L R R
L
1200 R R
L R
L R R 1200 R R
L R
L R R
R R
L R
L R R R R
L R
L R R
R R R
L R
L R
L R R R R R R R
L R
L R
L R R R R
R
L R
L L R
L R
L R R R R
L R
L L R
L R
L R R R
R R R
L R
L R
L R
L R
L R
L R
L R R R R
L R
L R
L R
L R
L R
L R
L R
L R
L R
L R
L R
L R
L L R
L R
L R
L R
L R
L
R L R
L R
L R
L R L R
L R
L R
L
800 L R
L R
L L R
L 800 L R
L R
L L R
L
L R R
L R
L R L R R
L R
L R
R L L L R L L L
L R
L L L R
L L
L L

400 400

0 0
0 400 800 1200 1600 2000 2400 0 400 800 1200 1600 2000 2400
Conflicting Flow (pcu/h) Conflicting Flow (pcu/h)
(a) Field Data with Linear Regression (b) Field Data with Exponential Regression

Figure 5.9. De Pere Eastbound PM Field Capacity Data

79
80

Figure 5.10 and Table 5.15 show the approach based model results. While
some lane imbalance was present, the approach based aggregation from the
WisDOT adjusted U.K. model was only slightly above the average capacity
observations near the intercept, but diverged with increasing conflicting flow.
German model results were comparable to the field data throughout the range
of data. One noticeable difference between these two models was the shape of the
capacity curve in this particular range of low circulating flow data. The linear U.K.
model showed a more gradual decline in capacity with increasing conflicting flow,
while the German exponential model predicted a steeper decline more similar to
the observed data.

Table 5.15. RMSE and Model Characteristics from the De Pere EB Approach
Based Analysis

Uncalibrated Calibrated

Intercept RMSE Intercept RMSE

Model (pcu/h) Slope n (pcu/h/ln) (pcu/h) Slope n (pcu/h/ln)
U.K. 2465 0.74 66 181 2182 0.771 66 105

German 2483 1.53×10-4 66 101 2571 2.22×10-4 66 103

 # UK (WisDOT Adj.) Calibrated UK
2400 # Default German 2400 Calibrated German
#
#
# Field Data Field Data
#
#
#
#
2000 # 2000
#
#
Approach Entering Flow (pcu/h)

Approach Entering Flow (pcu/h)

1600 1600

1200 1200

800 800

400 400

0 0
0 400 800 1200 1600 2000 2400 0 400 800 1200 1600 2000 2400
Conflicting Flow (pcu/h) Conflicting Flow (pcu/h)
(a) Uncalibrated Approach Based Models (b) Calibrated Approach Based Models

Figure 5.10. De Pere Eastbound Approach Based Capacity Comparison

81
82

Lane based model results for the eastbound approach are presented in
Figure 5.11 with error and model characteristics in Table 5.16. The right lane was
critical in this case due to the high volume of right turning vehicles. Capacity
predictions from the uncalibrated ARR 321 method were near the maximum
capacity observations. On the other hand, the NCHRP 572 model showed results
on the lower end of the observed data. Calibration brought the capacity results
from ARR 321 method down to a RMSE of 144 from 316 pcu/h/ln. Capacity
prediction also improved from the default scenario after calibrating the NCHRP
572 model to a RMSE of 132 pcu/h/ln, resulting in higher capacity predictions per
conflicting flow compared to the uncalibrated scenario.

Table 5.16. RMSE and Model Characteristics from the De Pere EB Lane Based
Analysis

Uncalibrated Calibrated

Intercept RMSE Intercept RMSE

Model (pcu/h) Slope n (pcu/h/ln) (pcu/h) Slope n (pcu/h/ln)
ARR 321 1633 N/A 125 316 1255 N/A 125 144

NCHRP 1130 7.00×10-4 125 167 1165 7.38×10-4 125 132

 Default ARR 321 Calibrated ARR 321
2400 Default NCHRP 2400 Calibrated NCHRP
Field Data Field Data

2000 2000

Right (Critical) Lane Entering Flow (pcu/h)

Right (Critical) Lane Entering Flow (pcu/h)

1600 1600

1200 1200

800 800

400 400

0 0
0 400 800 1200 1600 2000 2400 0 400 800 1200 1600 2000 2400
Conflicting Flow (pcu/h) Conflicting Flow (pcu/h)
(a) Uncalibrated Lane Based Models (b) Calibrated Lane Based Models

Figure 5.11. De Pere Eastbound Lane Based Capacity Comparison

83
84

5.4.6 Capacity Data Analysis Summary

To summarize the above capacity comparisons, all of the root mean square error
results are presented in the column chart of Figure 5.12. Importantly, all of the
models performed well when calibrated, as is expected because calibration fits the
model to the data. However, calibration can only be performed in retrospect on
existing roundabouts, and therefore default models provide a starting point for
analyzing future situations. Some default models consistently performed better
than others such as the German and NCHRP models. The fact that the German
model showed a good fit is likely due to the fact that the default gap acceptance
parameters in the model were similar to those found in this research. Error
values may be overrepresented for the Canal St site because the site experienced
less demand volume and congestion compared to the De Pere site. Because of
differences between the SIDRA Standard Model and the ARR 321 method, RMSE
values shown here may not reflect error from SIDRA software analysis. For
example, instead of calibrating by using gap acceptance parameters, using an
environment factor of 1.2 would lower the RMSE in the uncalibrated scenarios by
approximately 100 pcu/hr/ln using estimates of the differences between models
discussed in Section 2.4.5.1.
 85

Uncalibrated Calibrated

6
38

37
7
34
350 350

6
0

31
31
300 300
RMSE (pcu/hr/ln)

RMSE (pcu/hr/ln)
250 3
19 250
1

200 200

3
18

8
18

17
7

17
6
16
3
9

15

2
15
14

11

13
150 150

4
14

5
5
1

14
10
10

10
100 100

86
82

80
50 50

0 0

21
an
P

UK
*
an

P
UK

21
HR

HR

R3
rm
rm

R3
NC

NC
Ge
Ge

AR
AR

Canal Site NB De Pere EB De Pere

* Uncalibrated ARR 321 approximates SIDRA with an environment factor, EF = 1.0.

Using EF = 1.2 would lower uncalibrated RMSE by approximatedly 100 pcu/hr/ln in these scenarios.

Figure 5.12. Root Mean Square Error Summary

Error between any model and observed data could be due to numerous
factors. Sampling error due to observing limited regions, sites, approaches,
time periods, driver populations, geometric configurations, etc, all contribute to
variance within the data.
86

5.5 Capacity Modeling Extensions

Two additional analyses were performed with the capacity data, which were: look-
ing at the effects of using field measured geometry versus using a reduced effective
geometry measurement in the U.K. method , and using the U.K. approached based
method for a lane-by-lane analysis.

5.5.1 Adjusting for Effective Geometry

Geometric inputs require careful consideration in the U.K. model because these
are the only parameters to which capacity is sensitive based on inspection of the
capacity equation. Further, each geometric parameter is treated as a continuous
variable allowing for minute changes to affect capacity. Figure 5.13 and 5.14 show
the difference in the capacity predictions for the studied approaches when field
measured entry widths are used in lieu of ’effective’ entry widths specified by the
FDM. Table 5.17 and 5.18 show the respective comparisons of model parameters.
Using effective geometry improves the capacity prediction dramatically for the
Canal St site, which emphasizes the sensitivity of the entry width parameter.
Capacity prediction only improves slightly for the De Pere site. If the influence
or interactions between input parameters and output capacity are not fully
understood, erroneous predictions could easily be obtained from extrapolating the
model to situations beyond the original model scope.
 87

# WisDOT Adj. UK # WisDOT Adj. UK

2400 # Field Measured UK 2400 Field Measured UK
#
#
# Field Data Field Data
#
#
#
#
2000 # 2000
# #
#
#

Approach Entering Flow (pcu/h)

Approach Entering Flow (pcu/h)

##
##
##
1600 1600 #
#
#
##
#
#
#
#
##
1200 1200 ##
#
#

800 800

400 400

0 0
0 400 800 1200 1600 2000 2400 0 400 800 1200 1600 2000 2400
Conflicting Flow (pcu/h) Conflicting Flow (pcu/h)

(a) EB De Pere Approach (b) NB De Pere Approach

Figure 5.13. U.K. Model Effective Geometry Comparison from the De Pere Site

Table 5.17. RMSE and Model Characteristics from the De Pere Effective Geometry
Comparison

EB - PM Main Ave NB - PM Main Ave

Intercept RMSE Intercept RMSE

Model (pcu/h) Slope n (pcu/h/ln) (pcu/h) Slope n (pcu/h/ln)

WisDOT Adj.
2465 0.740 66 181 2450 0.740 76 347
U.K.

Field Measured
2628 0.771 66 246 2555 0.759 76 386
U.K.

Calibrated
U.K. 2182 0.771 66 105 1801 0.759 76 80
(not depicted)
88

# WisDOT Adj. UK
2400 Field Measured UK
Field Data

2000

Approach Entering Flow (pcu/h)

1600

#
1200 #
##
##
##
##
##
##
800 ##
#
##
#
#
400

0
0 400 800 1200 1600 2000 2400
Conflicting Flow (pcu/h)

Figure 5.14. U.K. Model Effective Geometry Comparison from the Canal St Site

Table 5.18. RMSE and Model Characteristics from the Canal St Effective
Geometry Comparison

Intercept RMSE
Model (pcu/h) Slope n (pcu/h/ln)

WisDOT Adj. U.K. 1323 0.532 71 381

Field Measured U.K. 1856 0.631 71 821

Calibrated U.K.
1055 0.631 71 156
(not depicted)

5.5.2 Combined De Pere Eastbound and Northbound Data

Analysis

Both studied approaches from the De Pere site had similar geometric parameters
for the U.K. model and therefore would have a similar capacities based on the
 89

model assumptions. The following assumptions were made when combining the
data:

• Data from different approaches with similar geometries are combinable;

• Lane data can be combined from differing critical lane positions (northbound
left critical lane, eastbound right critical lane); and

• Differing time periods can be combined (AM northbound data was in-
cluded).

The resulting combined entering-circulating data is shown in Figure 5.15, with (a)
combining approach based data and (b) combining critical lane data. Circulating
flows observed from the northbound AM peak fell between the eastbound PM
peak and northbound PM peak, with corresponding entering flows also between
the other data sets. Figure 5.15 also overlays the appropriate uncalibrated
approach and lane based models. Calibration was not performed due to lack of
appropriate means to combine the gap acceptance data for lane based models.
A "half capacity" U.K. model was applied to the critical lane data, which is a
technique available in RODEL and ARCADY to examine a single lane from a
multilane site with an approached based model. The assumption is that 50 percent
of the approach capacity will be dedicated to each lane. A "half-German" model is
not shown for the lane-based data, but does exhibit a similar fit as in the approach
based prediction.
 90
# UK (WisDOT Adj.) UK - Half Capacity
2400 # Default German 2400 Default ARR 321
#
# NB - AM Default NCHRP
#
#
#
# NB - PM NB - AM
#
2000 #
# EB - PM 2000 NB - PM
#
#

Critical Lane Entering Flow (pcu/h)

# EB - PM
Approach Entering Flow (pcu/h)

##
##
##
#
1600 #
#
1600
#
##
#
#
##
#
##
1200 ##
#
1200

800 800

400 400

0 0
0 400 800 1200 1600 2000 2400 0 400 800 1200 1600 2000 2400
Conflicting Flow (pcu/h) Conflicting Flow (pcu/h)
(a) Uncalibrated Approach Based Models (b) Uncalibrated Lane Based Models

Figure 5.15. Combined De Pere Capacity Data Comparison

 91

Table 5.19 contains the characteristics of each model used in comparison to

the combined capacity data. An average of the slope and intercept values from the
northbound and eastbound approaches was used to represent the U.K. model. For
comparison, Table 5.20 shows the results of regression analysis of the combined
data. Because a wide range of circulating flow data was observed, from 0 to 1920
pcu/h in the case of the combined critical lane data, regression slope and intercept
parameters are more representative of the data compared to regressing subsets of
circulating flows. Intercepts from the regression analysis showed higher results
than the NCHRP model and lower results than the U.K., ARR 321, and German
models. Regression slopes were steeper than the NCHRP, U.K., and German
models.

Table 5.19. Model Characteristics from the Combined De Pere Capacity Data
Comparison

Approach Based Critical Lane Based

Intercept RMSE Intercept RMSE

Model (pcu/h) Slope n (pcu/h/ln) Model (pcu/h) Slope n (pcu/h/ln)
U.K. -
U.K. 2458 0.74 166 292 1229 0.74 477 348
Half Cap.

German 2483 1.53×10-4 166 90 NCHRP 1130 7.00×10-4 477 151

ARR 321 1633 N/A 477 362

Table 5.20. Combined De Pere Capacity Data Set Regression

Intercept RMSE
Data Set Regression (pcu/h) Slope n R2 (pcu/h/ln)
Critical Linear 1158 0.632 477 0.826 128
Lane -3
Exponential 1404 1.13×10 477 0.773 123
Linear 2269 1.110 166 0.889 97
Approach -4
Exponential 2625 9.35×10 166 0.866 91
92

Trends in the combined scenario are logically the same as the trends ob-
served when analyzing data from each approach separately. At lower circulating
flow rates the U.K. model is closer to the average observed data, but diverges at
high circulating flow rates, leading to the relatively large RMSE values of 292 and
348 pcu/h/ln for the approach and lane based models, respectively. The German
approach based model fits the data well, with a RMSE value of 90 pcu/h/ln. For
lane based models, the NCHRP model tended to underpredict capacity at low
circulating flows and overpredict at high circulating flows, with the second lowest
RMSE value of 151 pcu/h/ln. The ARR 321 method showed overprediction with
an RMSE of 348 pcu/h/ln, but the slope visually appears to follow the general
trend of the data.
At higher circulating flow rates, the capacity prediction becomes increas-
ingly important. High circulating flow rates means that drivers will have fewer
gaps to choose from and have the potential to experience more delay. Linear
models predict a constant decrease in capacity toward a distinct x-axis intercept,
beyond which capacity is predicted at zero entering vehicles. Exponential models
have a more horizontal relationship at high circulating flow rates, converging
quicker to an asymptote above zero entering vehicles, similar to how left turn
lanes at signals can experience vehicles sneaking into the intersection on the yellow
indication which adds capacity. For roundabouts, extra capacity could be gained
by aggressive drivers forcing gaps or having periods of priority reversal. These
aggressive characteristics would explain why the U.K. model has a lower slope.
Exponential gap acceptance models would need a more complex relationship, such
as the ARR 321 method, or consider a range of critical gap and follow-up headway
values for different circulating flows to account for such behaviors.
 93

6 Software Output and Usability

All software modeling was conducted with a philosophy as if the roundabouts

studied did not exist yet. This philosophy led to comparing essentially ’default’
situations where if little is known about future conditions, an analyst may rely
on uncalibrated software based default values. This default scenario resulted
in the most basic analysis and the following assumptions were further used for
comparing software results:

• Turning movement counts were used to determine the peak hour volume,
percentage of trucks, and peak hour factor. While turning movement counts
are a measure of departure volume, the counts were input in place of true
demand volumes;

• Even with entering departure volumes, no restriction was placed on volume

to capacity (v/c) ratios. Departure volumes should never allow a v/c ratio
greater than 1.0, but the software was allowed to compute and report any
v/c ratio, including values exceeding 1.0. This allowed for identification of
conservative model results;

• Default lane utilization was assumed, allowing for the software to identified
lane imbalance, if any;
94

• Exact queues from the field data were not known and therefore only an
approximation could be determined if software queue results were too low
based on the field of view in the video data collection; and

• SIDRA delay and level of service was setup to maximize compatibility to the
HCM 2010 while still using the SIDRA Standard capacity model. As such,
the default environment factor of 1.2 for U.S. conditions was used to better
represent typical software analysis.

• Calibration was performed by adjusting gap acceptance parameters only.

SIDRA software defines gap acceptance parameters by-movement rather
than the by-lane method used in this research. By-movement values present
a difficulty in assigning values for through movements where more than one
through lane exists. An assumption was made to assign the gap acceptance
values for the through movement based on the gap acceptance values from
the lane most utilized by the through movements.

Calibrated scenarios then expanded upon the default scenario by making the same
model calibration adjustments performed in the spreadsheet analysis, described
in Section 5.4.1, on the field studied approaches. No additional parameters were
changed in the calibrated scenarios. Approaches not studied were untouched and
default values were used in the calibrated scenarios. Table 6.1 presents a summary
of the calibration procedures used in each software package, with noticeably only
ARCADY, HCS, and SIDRA allowing for calibration.
 95

Table 6.1. Software Calibration Procedure Summary

Software
Software
Allows Software Calibration Method
Package
Calibration?

RODEL
No N/A
1.9.7

Replace default slope and intercept with calibrated

ARCADY
Yes values based on field observed average entering
7.1
and circulating flow

RCAT
No N/A
1.4

KREISEL
No N/A
7

GIRABASE
No N/A
4

HCS 2010 Replace default critical gap and follow-up headway

Yes
6.1 values with field observed values

Disable software calculated critical gap and follow-

SIDRA
Yes up headway values. Enter user defined, field
5.1
observed values.

Some software packages use slightly different definitions or calculations

for capacity, delay and queuing. Consistency was sought to make sure that all
results were equally comparable. Some notable exceptions were: All packages
implementing the U.K. model were based on maximum queue length compared
to 95th percentile queue length from Kreisel, HCS, and SIDRA. Girabase queue
length was based on an average and maximum; RODEL and Girabase did not
output LOS for each approach; and Girabase capacity was based on "reserve
capacity" from which capacity comparable to other software was derived by hand.
The number of significant digits reported for each performance measure varied by
96

software package. The values shown in Sections 6.1 and 6.2 reflect exactly what
the user would see within the software for comparison purposes. Table 6.2 shows
the resulting number of decimal places reported by each software. In some cases,
rounding makes a large difference, especially regarding delay where rounding to
the nearest second or half second can alter the level of service reported.

Table 6.2. Difference in Decimal Places Reported for Performance Measures

# of Decimal Places
v/c Ratio Delay Queue
RODEL
2 1 1
1.9.7
ARCADY
2 2 2
7.1
RCAT
2 2 1
1.4
KREISEL
2 0 0
7
GIRABASE
2 0 0
4
HCS 2010
2 1 1
6.1
SIDRA
3 1 1
5.1

All level of service (LOS) scores were based on the definition from the
Highway Capacity Manual 2010, which assigns LOS F to any lane with volume
to capacity ratio (v/c) greater than 1.0, and with other LOS assignments based on
unsignalized intersection delay as shown in Table 6.3 [1]. A full list of differences
between performance measure variations between software packages is presented
with the software output from all approaches in Appendix B.
 97

Table 6.3. Control Delay Criteria for Level of Service

Control
Level of
Delay
Service
(s/veh)
0 - 10 A
> 10 - 15 B
> 15 - 25 C
> 25 - 35 D
> 35 - 50 E
> 50 F

6.1 Canal St Site Software Results

Uncalibrated results for the studied southbound 25th St approach at the Canal St
site are shown in Table 6.4. Capacity results ranged from the mid 500 vph from
Kreisel and HCS to over 1300 vph from Girabase, resulting in a wide range of
other performance measure results.
98

Table 6.4. Canal St Site: Southbound 25th St Approach - Uncalibrated

Capacity Delay Queue

(veh/h) v/c Ratio (s/veh) LOS (veh)
RODEL
832 0.67 14.4 B 3.3*
1.9.7

ARCADY
784 0.78 20.47 C 3.34*
7.1
RCAT
824 0.67 14.94 B 3.5*
1.4

KREISEL
547 1.08 227 F 49
7
GIRABASE
971 0.61 5 A 1 to 5*
4

HCS 2010
569 1.04 74.2 F 16.2
6.1

SIDRA
680 0.866 18.7 C 8.5
5.1
*queue length low based on video evidence

Packages implementing the U.K. model conservatively showed v/c ratios of

0.67 to 0.78, corresponding to delays of about 15 to 20 s, respectively. Interestingly,
RCAT reports delay of 14.94 s as LOS B, while 15.0 s would be classified as C
despite the negligible difference in operations. Queue lengths of about three
vehicles were considered low based on video evidence.
Kreisel and Girabase showed opposite results. While Kreisel predicted
slightly over saturated conditions with a 1.08 v/c ratio and long delay of 227 s,
Girabase showed conditions as less than half saturated and negligible delay. A
queue length of 49 vehicles from Kreisel seems long but cannot be confirmed or
rejected from video data collection.
HCS results also showed slight oversaturation with a 1.04 v/c ratio, but less
extreme delay and queuing than Kreisel. SIDRA results were in between the U.K.
 99

model packages and HCS. A delay of 19 seconds was similar to U.K. packages but
queue length was more reasonable at nearly nine vehicles.
Calibration was possible in ARCADY, HCS, and SIDRA with results shown
in Table 6.5. In each case, capacity results were similar ranging from about 420
to 460 vph. All showed oversaturated conditions with long delays and queues.
ARCADY seemed particularly sensitive to high v/c ratios by showing over 660
s of delay and a queue of 105 vehicles. Calibrated parameters based on the data
collection were possibly too conservative in this case, as a v/c ratio near 1.0 was
expected from the capacity conditions observed.

Table 6.5. Canal St Site: Southbound 25th St Approach - Calibrated

Capacity Delay Queue

(veh/h) v/c Ratio (s/veh) LOS (veh)
RODEL
― ― ― ― ―
1.9.7

ARCADY
416 1.47 661.32 F 105.13
7.1
RCAT
― ― ― ― ―
1.4

KREISEL
― ― ― ― ―
7
GIRABASE
― ― ― ― ―
4

HCS 2010
445 1.34 191.8 F 26.8
6.1
SIDRA
461 1.279 161.2 F 54.9
5.1
― software does not allow for calibration
100

6.2 De Pere Site Software Results

Software results from the studied eastbound Main Ave approach at the De Pere site
are shown for the uncalibrated scenario in Table 6.6. All software packages, except
Girabase, showed LOS F and corresponding high v/c ratios, delays, and queues.
Video data collection was not able to confirm the maximum back of queue, but
anecdotal evidence from the Northeast Region DOT suggested about 70 vehicles
for the right lane and 35 vehicles for the left lane as reasonable estimates. Of the
two lane-based software packages, only HCS hinted at some lane imbalance with
a queue of 44 vehicles in the right lane and 32 vehicles in left lane. Lane imbalance
can result in significant differences in performance measures for different turning
movements on the same approach, which could make lane-by-lane models more
useful in such situations. Delay from Kreisel was extreme at over 500 seconds,
ARCADY and HCS showed the next highest delay at over 100 seconds, and both
RODEL and SIDRA showed about 60 seconds of delay.
 101

Table 6.6. De Pere Site: Eastbound Main Ave Approach - Uncalibrated

Capacity Delay Queue

(veh/h) v/c Ratio (s/veh) LOS (veh)
RODEL
2206 0.95 58.3 F 80.1
1.9.7

ARCADY
2182 1.06 110.72 F 80.07
7.1
RCAT
2222 0.95 52.74 F 71.7
1.4

KREISEL
1897 1.15 530 F 167
7
GIRABASE
2609 0.84 2 A 0 to 2*
4

HCS 2010 L: 871 L: 1.18 L: 112.0 L: F L: 31.8

6.1 R: 855 R: 1.31 R: 164.1 R: F R: 44.1

SIDRA L: 1027 L: 1.044 L: 60.1 L: F L: 42.5

5.1 R: 1069 R: 1.044 R: 59.1 R: F R: 43.5
*queue length low based on video evidence

Calibration available in ARCADY, HCS, and SIDRA also showed LOS F

as can be seen in Table 6.7. Queue length from ARCADY nearly tripled to over
200 vehicles queued on the approach corresponding to over 400 s of delay, which
are likely too large of estimates. Queue and delay results from HCS decreased
by about one third. Lane imbalance was still present, although not to the extent
estimated in the field. SIDRA showed reasonable queuing for the right lane, but
did not show smaller queues in the left lane to reflect any lane imbalance.
102

Table 6.7. De Pere Site: Eastbound Main Ave Approach - Calibrated

Capacity Delay Queue

(veh/h) v/c Ratio (s/veh) LOS (veh)
RODEL
― ― ― ― ―
1.9.7

ARCADY
1891 1.22 411.17 F 227.25
7.1

RCAT
― ― ― ― ―
1.4
KREISEL
― ― ― ― ―
7

GIRABASE
― ― ― ― ―
4
HCS 2010 L: 963 L: 1.07 L: 69.2 L: F L: 24.2*
6.1 R: 972 R: 1.19 R: 114.3 R: F R: 35.6
SIDRA L: 899 L: 1.207 L: 121.5 L: F L: 81.1
5.1 R: 913 R: 1.207 R: 121.2 R: F R: 82.0
*queue length low based on video evidence
― software does not allow for calibration

Table 6.8 contains results from the Northbound Broadway St Approach

in the uncalibrated scenario. Results for capacity were mixed, with approach
capacity ranging from about 780 vph from Kreisel to about 1500 vph from RODEL,
ARCADY, and RCAT. All software packages showed low values for queuing,
contrary to what was observed in the field. Extent of the video recording showed at
least 10 vehicles consistently in queue in the left lane and likely the back of queue
extended another 10 to 20 vehicles. Right lane queues were more sporadic and
did not appear to exceed 10 vehicles which showed the significant variability in
performance measures that can occur within lanes on the same approach. Both the
lane based modeling packages, HCS and SIDRA, identified lane imbalance skewed
toward higher queuing in the left lane, a similar trend to field observations.
 103

Table 6.8. De Pere Site: Northbound Broadway St Approach - Uncalibrated

Capacity Delay Queue

(veh/h) v/c Ratio (s/veh) LOS (veh)
RODEL
1512 0.43 4.1 A 0.9*
1.9.7

ARCADY
1476 0.48 4.71 A 0.93*
7.1

RCAT
1497 0.43 4.31 A 1.0*
1.4

KREISEL
772 0.87 32 D 16*
7
GIRABASE
1186 0.57 3 A 0 to 3*
4
HCS 2010 L: 421 L: 0.91 L: 55.4 L: F L: 9.9*
6.1 R: 450 R: 0.65 R: 24.7 R: C R: 4.5
SIDRA L: 468 L: 0.819 L: 38.1 L: E L: 8.4*
5.1 R: 395 R: 0.735 R: 34.2 R: D R: 5.9
*queue length low based on video evidence

Table 6.9 shows calibrated results where available. ARCADY capacity was
lowered by about 38 percent after calibration. For the lane based software, SIDRA
capacity was lowered by about 25 percent for the left lane and about six percent
for the right lane. HCS results were increased by about nine percent for the left
lane and 36 percent for the right lane. Queue length estimates still remained low
from ARCADY and HCS. SIDRA, however, showed more reasonable queuing and
better reflected the lane imbalance with 22 vehicles queued in the left lane versus
five in the right lane.
104

Table 6.9. De Pere Site: Northbound Broadway St Approach - Calibrated

Capacity Delay Queue

(veh/h) v/c Ratio (s/veh) LOS (veh)
RODEL
― ― ― ― ―
1.9.7

ARCADY
934 0.76 16.12 C 3.09*
7.1

RCAT
0― 0. ―00 0― 0― ―0
1.4
KREISEL
0― 0― 0― 0― ―0
7

GIRABASE
0― 0. ―00 0― ―0 ―0
4
HCS 2010 L: 458 L: 0.85 L: 42.5 L: E L: 8.4*
6.1 R: 614 R: 0.48 R: 13.6 R: B R: 2.6
SIDRA L: 350 L: 1.095 L: 110.0 L: F L: 22.4
5.1 R: 417 R: 0.697 R: 29.8 R: D R: 4.8
*queue length low based on video evidence
― software does not allow for calibration

6.3 Technical Accuracy Summary

In light of the model comparisons to the field data and RMSE calculations as well
as the software output, the relative technical accuracy of each model and software
could be compared based on capacity output. Due to the complex and numerous
amount of data involved a graphical rating scale was developed to allow quick
comparisons of the relative technical accuracy shown in Table 6.10. Table 6.11
shows the resulting comparisons.
 105

Table 6.10. Graphical Rating Scale for Technical Accuracy

Poor - Model and software did not match field data

Fair - Model and software match field data

reasonably when used with some caution

Good - Model and software match field data with

some exceptions

Very Good - Model and software consistently

matched field data

Excellent - Model and software results clearly

matched field data in all cases
106

Table 6.11. Software and Model Technical Accuracy Summary

Technical Accuracy
(Model prediction vs Field Data)
Ratings are Based on Consistency of Capacity Prediction
Software
Default Model Calibrated Model
‘WisDOT Adjusted’
(using field measured (using field collected
U.K. Model
geometry) data)
‘WisDOT Calibrated’ U.K. Model U.K. Model
U.K. Model
RODEL *RODEL did not feature
1.9.7 capacity calibration

‘WisDOT Calibrated’ U.K. Model U.K. Model

U.K. Model
ARCADY
7.1

‘WisDOT Calibrated’ U.K. Model U.K. Model

U.K. Model
RCAT *RCAT did not feature
1.4 capacity calibration

- German HBS 2001 German HBS 2001

Model Model
KREISEL
*KREISEL did not feature
7
capacity calibration for the
HBS 2001 Model

- French Model French Model

GIRABASE *GIRABSE did not feature

4 capacity calibration

- HCM 2010 Model HCM 2010 Model

HCS 2010
6.1

- ARR 321* ARR 321

SIDRA
5.1

* Uncalibrated ARR 321 approximates the SIDRA Standard Model with an

environment factor of 1.0
 107

6.4 Limitations

Several limitations in this research warrant discussion, but even with these
limitations valuable insights can still be gained. First, only three approaches
from two roundabouts were considered. Even with observing a small number of
locations, the total sample size from each site was relatively large. For instance,
the 268 queued minute observations from the PM peak of the left lane on the
northbound Broadway St approach alone was about two-thirds the size of the
entire multi-lane data set, 414 observations, in the NCHRP 572 research [14].
Having a large sample from one site allows for a good representation of a specific
scenario, useful for calibration for one site, but lacks the between-site variation
needed for broad capacity model development to minimize sampling error.
Second, software analysis was limited to the turning movement data collected
which was representative of the traffic volume serviced and not necessarily the
traffic volume demand. If the traffic volume serviced is used and it is less than
the true demand, queues and delays will be underrepresented. Software packages
need accurate demand traffic volumes for queue and delay prediction. However,
queue and delay models use the volume to capacity ratio and are thus also
dependent on capacity estimates. Therefore, identifying the best queuing and
delay models may not be possible but trends may still be identified. Models that
tend to overpredict capacity would potentially have lower v/c ratios leading to
the possibility of underpredicted queuing and delay, and vice versa for models
that underpredict capacity.
108

6.5 Interface Usability

Each software package was evaluated in general terms of how user-friendly the
program operates, recalling from Chapter 3, that usability was to be evaluated
by initial learning curve, ease of data input, feedback and error prevention, and
long-term memory load for infrequent users. Because of the subjective nature
of these criteria, the comments reflected here represent the best consensus of the
researchers and steering committee. Further, as software continually changes,
these comments are based on the versions of the software provided at the onset
of this research.

6.5.1 RODEL 1.9.7

RODEL 1.9.7 is a Microsoft-DOS application that uses a single window to display

all input and output information at the same time as shown in Figure 6.1.
Presenting all input and output information at once was seen as both a positive
and negative aspect of the software. Seeing everything at once provides a concise
overview while too much information can be distracting when trying to focus on
one specific task.
Benefits of the interface include:

• Easy to learn;

• Simple input of parameters;

• Inputs can be quickly changed to see the potential effects on performance

measures; and

• Scenarios can be easily be copied between files to create different sets of

scenarios, although these cannot be compared side-by-side.
 109

Considerations for the interface include:

• A younger generation of users may not be comfortable with DOS interface.

Control of the program is by keyboard only;

• Red text on a black background can be straining on the eyes;

• Lack of formatted output creates the need to use screenshots (or retyping all
output) as the common reporting mechanism.

• The color scheme is not conducive to efficient use of ink during printing of
screenshots; The lack of an in-context help system makes abbreviations and
other terms potentially difficult to remember;

• Some commands are hidden or difficult to remember, such as Ctrl+F2 to view

slope and intercept parameters; and

• Lack of labeling input and output columns forces the user to mentally rotate
information, which can lead to data entry or reporting errors.
110

Figure 6.1. RODEL 1.9.7 Interface

6.5.2 ARCADY 7.1

ARCADY 7.1 uses a multiple document interface style, shown in Figure 6.2. While
the screenshot looks cluttered, users have full control over what information
display at any given time; this is an extreme example to show many different types
of features. Toolbars across the top and left side of the main window organize the
analysis workflow. Four major types of dialogs are commonly used in ARCADY
and are labeled in the figure. Labeled dialog "1" shows a tree interface that contains
the different scenarios and sites, as well as the geometry and capacity information
for each approach. Dialog "2" shows a list style interface for entering data or
viewing results. These style dialogs can also be viewed in a grid format to show
relevant information from multiple approaches. Dialog "3" shows a schematic
of the roundabout, which can be used to overlay information and highlight the
approach to which selected data applies. Finally dialog "4" shows an example
 111

graphs that can be used to analyze, compare, and apply different scenarios.

 


Figure 6.2. ARCADY 7.1 Interface

Benefits of the interface include:

• Information is organized into discrete areas allowing the user to view only
the most relevant information at any given time;

• A dialog can be displayed that informs the user of any errors or warnings;
Built-in glossary to quickly define any terms or acronyms;

• Easy side-by-side comparisons through customizable tables;

• Formatted output reports; and Data entry in tables can be copied and pasted
to or from other applications (Excel).

Considerations for the interface include:

112

• Initially difficult to learn;

• Overwhelming number of options can be intimidating especially if the

analyst is only interested in a subset of the available features and needs to
filter out unwanted choices;

• Lack of strict step-by-step workflow can lead to data entry error, although
the warnings dialog helps (if it is open);

• Hard to remember where options or inputs are located within the program
because of the deep tree structure and not always intuitive location;

• The above points lead to a large long term memory demand, which is not as
desirable for casual users; and

• Can be difficult for someone to quickly double-check all inputs.

6.5.3 RCAT 1.4

Microsoft Excel provides the basic interface for RCAT 1.4 and the spreadsheet is
organized into four areas, 3 for input and one for output. All areas have similar
design, and an example of the output area is shown in Figure 6.3. The four
areas are: traffic demand turning movement counts, traffic flow profile, geometry,
performance measure output. Navigation buttons in the upper right corner of each
area allow quickly switching between the different input and output areas.
 113

Figure 6.3. RCAT 1.4 Interface

Benefits of the interface include:

• Familiar tools for users already comfortable with Excel; and

• Highly organized workflow for efficient data entry and performance mea-
sure output.

Considerations for the interface include:

• Lack of error or warning messages other than the standard Excel errors
within cells; and
114

• Worksheet format is rigid and locked by a 3rd party application, making cus-
tomization not possible within RCAT. Other spreadsheets could potentially
be developed to link to RCAT input and output.

6.5.4 KREISEL 7.0

Kreisel 7.0 uses a multiple document interface to display input and output
information as shown in Figure 6.4. Typically only one dialog is open at a time
at the users’ discretion. A toolbar across the top of the main window guides users
through the analysis workflow.

Figure 6.4. Kreisel 7.0 Interface

Benefits of the interface include:

• Moderate learning curve;

 115

• Familiar tools for users already comfortable with Excel;

• Highly organized workflow for efficient data entry and performance mea-
sure output; and

• A schematic, while primitive, does help visually reinforce the general round-
about shape.

Considerations for the interface include:

• Many modeling options exist for each type of performance measure (capacity
models, delay models, etc) and the user must be careful to choose the correct
options; and

• Data entry is performed through a grid interface where most input options
are labeled by abbreviations. Having the user guide handy helps to reassure
term definitions.

6.5.5 GIRABASE 4.0

Simple and effective best define the interface for GIRABASE 4.0 which is shown
in Figure 6.5. A version with a French interface was used for this research, but an
English interface is available based on information from the vendor.
116

Figure 6.5. GIRABASE 4.0 Interface

Benefits of the interface include:

• Quick learning curve;

• Simple tabs guide user through the analysis workflow and is easy to
remember; and

• Errors and warnings are displayed through highlighting bad input values,
status bar messages, or pop-up dialogs.

Considerations for the interface include:

• Input and output cannot be easily displayed in a format for checking and
reporting, although an option exists for printing all output; and

• Performance measures are atypical of other software. For example, reserve

capacity is reported instead of capacity directly and level of service is not
 117

reported. These limitations require the user to post-process output in order

to make comparisons to other software.

6.5.6 HCS 2010 6.1

HCS 2010 uses a form-like interface for data entry and output display as shown in
Figure 6.6.

Figure 6.6. HCS 2010 6.1 Interface

118

Benefits of the interface include:

• Quick learning curve;

• Simple scrolling interface provides logical workflow;

• A single page formatted report provides quick access for checking and
reporting analysis; and

• The interface is familiar to users of other HCS modules (stop controlled,

freeways, etc).

Considerations for the interface include:

• Column labels are not always visible when scrolling, so data could mistak-
enly be entered into the wrong column; and

• Entering lane configurations for each approach can be confusing.

6.5.7 SIDRA 5.1

Figure 6.7 shows SIDRA 5.1 which uses a tabbed interface with a tree structure on
the right side to organize different scenarios.
 119

Figure 6.7. SIDRA 5.1 Interface

Benefits of the interface include:

• Moderately easy to learn, with an extensive help available from the user
guide;

• Highly organized workflow with logical tree hierarchy for input and output
which is easy to remember;

• Dialogs are supplemented with graphics to help visualize changes and data
input;
120

• Formatted reports and summaries allow for reporting and analysis checks
throughout the process; and

• Scenarios are easily cloned to analyze different geometries, volumes, etc.

Considerations for the interface include:

• Some options can be applied per approach or for the entire intersection; the
user must be careful to apply changes to the appropriate scope; and

• Multiple scenarios can only be compared by toggling between separate tabs,

which is not as easy as a side-by-side table.

6.6 Feature and Interface Comparisons

A graphical rating scale, explained in Table 6.12 was developed to allow quick
comparisons to summarize the usability and features evaluated in each software.
 121

Table 6.12. Graphical Rating Scale

Poor - Feature did not perform well or was absent

Fair - Feature performed reasonably when used with

some caution

Good - Feature performed well but showed some

limitations

Very Good - Feature performed consistently and

accurately

Excellent - Feature completely implemented, no

issues discovered

Usability was holistically evaluated as summarized in Table 6.13. Again,

these are qualitative, subjective results that were from a perspective that empha-
sized usage for DOT staff that may be more infrequent, rather than specialist,
roundabout analysts and are based on the version of the software listed.
122

Table 6.13. Software Interface Evaluation

Software Ease of User Interface

MS-DOS command line may be uncomfortable for younger generation

RODEL
users. No copy/paste or printing functions exist, so taking screen shots
1.9.7
to demonstrate results is less than ideal.

Uses a multiple document interface with tree outline of inputs and

scenarios. Can be overwhelming with knowing where to start data
ARCADY entry due to large number of options. Easy to compare multiple
7.1 scenarios side-by-side and explore relationships with graphing. Can be
hard to remember where certain inputs are located in the long term
making the software less usable for infrequent users.

Uses an Excel interface that is easy to use for those familiar with
RCAT
Microsoft Office. The interface is cleanly organized into three areas for
1.4
input and one area for output.

A logical toolbar layout with separate windows for each input type
KREISEL provides an orderly workflow. Need to exercise caution when choosing
7 from the multitude of model options. A grid-like interface for data input
makes most data entry simple.

Tabbed interface with logical ordering of input. Easy to remember how

GIRABASE
to use the software. Output screen shows entering versus circulating
4
flow graphs to aid in understanding capacity relationship.

Single window interface with inputs organized in a large grid of rows

HCS 2010
and columns. Column headings are not always visible when scrolling
6.1
through the long grid, leading to potential data entry errors.

Tabbed interface with tree outline of inputs and outputs. Logical

ordering of data input is easy to remember. Graphical output and a
SIDRA
variety of formatted reports can easily be printed or transferred to other
5.1
documents. User needs to check frequently to assure inputs apply to a
specific leg or the entire intersection.

6.6.1 Input requirements

Input data needs were compared in order to quantify the data input intensity of
each software package as shown in Table 6.14.
 123

Table 6.14. Software Data Input Needs

List of Inputs for Basic Roundabout Analysis

Software
Traffic Data Geometry Data Other Data
RODEL  Entry Width -
1.9.7  Half Width
 Traffic Volumes
ARCADY  Flare Length
 % Trucks -
7.1  Entry Radius
 Traffic Demand Profile  Phi, Entry Angle
RCAT
 Inscribed Diameter -
1.4
 Traffic Volume  Number of Lanes Entering
KREISEL converted to pcu/h or  Number of Lanes Conflicting per  Approximate
7.0 Traffic Volume by approach Exit Capacity
vehicle type  Inscribed Diameter
 Central Island Diameter
 Truck Apron Width
 Traffic Volume  Circulating Width
 Environment:
GIRABASE converted to pcu/h or  Approach Angle
Urban, Rural,
4.0 Traffic Volume by  Approach Grade
 Entry Width at 4 and 15 m Suburban
vehicle type
 Splitter Island Width
 Exit Width
 Number of Lanes Entering
 Traffic Volumes
HCS 2010  Number of Lanes Conflicting per
 Peak Hour Factor -
6.1 approach
 % Trucks  Bypass lanes, if any
 Number of Lanes Entering and
exiting
 Lane Disciplines/Configuration
 Number of Lanes Conflicting per
Approach
 Traffic Volumes  Approach and
SIDRA  Approach and Exit Short Lane
 Peak Hour Factor Exit Cruise
5.1 Lengths
 % Trucks  Lane widths and lengths Speeds
 Central Island Diameter
 Circulating Width
 Entry Angle
 Approach Grade
- : no other data was required for the scenarios evaluated
124

6.6.2 Feature Comparison

Major features that were of primary importance for the software analysis are
summarized in Table 6.15. Where analyzing multiple models within one software
is possible, Kreisel 7.0 can return results from about 30 different roundabout
capacity models, and SIDRA 5.1 was found to have implementations of the HCM
2010 and limited results from the German model.

Table 6.15. Major Features for Software Comparison

Can the Allow analyzing

software Analysis Allow multiple
return results by Maximum calibration of scenarios within
from multiple approach number of model the same file,
Software models? or by-lane approaches parameters side-by-side
RODEL
1.9.7
No Approach 6 No No

ARCADY
7.1
No Approach 20 Yes Yes

RCAT
1.4
No Approach 4 No No

Approach
KREISEL
7
Yes Based 8 No No
(HBS 2001)

GIRABASE Approach
4
No
Based
8 No No

Lane
HCS 2010
6.1
No Based, up 4 Yes No
to 2 lanes
Lane
SIDRA Based, can
5.1
Yes
do more 8 Yes Yes
than 2
lanes

Because software is continually changing, some secondary and desirable

features were identified, but not formally evaluated, and are shown in Table 6.16.
 125

Several points explaining the fractional ratings (partial circles) of the versions of
the software evaluated were worth noting:

• Bypass lanes are modeled by removing right turns in the U.K. model
packages;

• HCS modeling of other intersection types (signals, stop controlled, etc)

requires the data to be re-entered by hand, whereas SIDRA allows scenarios
to be copied and pasted; ARCADY corridor modeling is for roundabouts
only;

• HCS allows for modeling other intersection types besides roundabouts, but
requires the user to retype common information, such as traffic volumes;

• Formatted reports in Girabase are available by printing only;

• Kreisel and Girabase schematics are more limited than other software pack-
ages; HCS requires CORSIM for visualization;

• ARCADY safety analysis is based on U.K. research;

• Girabase graphing is limited to showing only entering versus circulating

flow relationships; and

• SIDRA includes graphical sensitivity analysis for major inputs and outputs,
but cannot graph any variable like ARCADY.
 126
Table 6.16. Comparison of Advanced Secondary Features

Allow
Allow Includes Includes
modeling Includes Includes
Allow bypass modeling schematic or graphing
Software other formatted safety
lanes linked sites other analysis
intersection report output analysis
(corridors) visualization capabilities
types

RODEL
1.9.7

ARCADY
7.1

RCAT
1.4

KREISEL
7

GIRABASE
4

HCS
6.1

SIDRA
5.1
 127

6.6.3 Installation and License Requirements

Installation and licensing requirements for the version of each software evaluated
were identified and are summarized in Table 6.17.

Table 6.17. Installation And Licensing Requirements Comparison

Software Installation Requirements Licensing Type & Cost

Windows XP or Older (not Windows 7). Must be
RODEL Response from vendor was based on a
able to run DOS command line programs. No
1.9.7 future version of the software.
specific hardware requirements.

Windows (XP, Vista, 7) Any modern PC will run

ARCADY, for fast performance, recommended Network ($9000 for 4 concurrent users)
ARCADY
parameters are:2+ GHz Processor, 1 GB RAM, or Standalone ($2500). Additional seats
7.1
50 MB hard drive space, hardware accelerated available with discounted prices.
OpenGL capability.

Requires Microsoft Excel. Has been tested in

Cost for one standalone license is $195.
RCAT Excel 2007 on Windows XP and 7. (RCAT was
For 5 or more licenses, a quote will be
1.4 successfully used in Excel 2010 for this project as
determined.
well).

Cost is on case-by-case basis for large

state agencies. Future versions will have
Windows (XP, 7). No specific RAM, CPU, or hard
licensing authorization over the internet.
KREISEL drive requirements; it runs on old PCs/Laptops as
Upgrades come at 7% of the purchase
7 well as modern ones. Installation directory uses
price. A single user can purchase the
23 MB of hard drive space.
program for 1285 Euro. (about $1900
U.S.)

Windows 95,98,NT,2000,XP. Software will run

on Windows 7 but has not been tested
GIRABASE Single Workstation Licenses Only at 990
extensively. No specific RAM, CPU, or hard drive
4 Euro (about $1420)
requirements. Installation directory uses 2 MB of
hard drive space.

Windows (2000,XP,Vista,7-32 bit, 7-64 bit).

HCS 2010 Some modules require .NET Framework An agency license is $12,000+, does not
6.1 (roundabout module does not) 512 MB RAM, 750 use a network based license server
MB Hard Drive Space

Windows XP (SP2), Vista, 7, 32 bit and 64 bit.

Standalone (2350 AUD, about $2500
Internet Explorer 7 or later. Requires .NET
U.S.) or Network Based License (16000
SIDRA Framework 3.5 (SP 1) and Microsoft SQL
AUD, or about $16900 U.S., for 10
5.1 Compact edition 3.5 (SP 1) 1 GB RAM,
Network Seats). Enterprise licensing
Installation directory uses 50 MB of hard drive
available with case-by-case costs.
space.
Exchange Rates Used (June 28th, 2011)
1 EUR = 1.43663 USD, 1 AUD = 1.05377 USD
128

6.6.4 Training Needs and Availability of Support

Training and support for the version of each software package evaluated are
summarized in Table 6.18.

Table 6.18. Software Training and Support Availability

Software Training Needs Availability of Support

RODEL Response from vendor was based on Response from vendor was based on a
1.9.7 a future version of the software future version of the software

Software and design training is

available in the U.S. from TRL or
ARCADY Included for the 1st year, 15% of the initial
authorized training centers. Costs
7.1 purchase price each year thereafter
are quoted based on class size and
type.
RCAT
No specific training required. Email support and updates free 1 year.
1.4
No specific training required.
Courses available in lengths of 1 or
KREISEL
more days. Costs have been from Telephone support (in German) at no cost.
7
400 to 2000 Euros depending on the
attendees and duration.

Training is available for an additional

GIRABASE
fee that is not included in the Software guaranteed stable for 12 months.
4
purchase price.

1 to 3 day workshops and courses

HCS 2010
available. Flexible in teaching at 1st year included, $400/year thereafter
6.1
various skill levels.

A U.S. representative is available for

special training courses that can be
SIDRA arranged on demand. Fees range One year free tech support and upgrades
5.1 from $500 to $1000 for a 2-day included in purchase price.
course, depending on the number of
trainees.
 129

6.6.5 Considerations for Future Software Versions

Like most technology, software versions can change at a rapid pace, making
comparisons difficult when the user knows the latest-and-greatest features are
coming soon. One emerging feature has been the integration of design and
analysis software packages. Developers from ARCADY and SIDRA are working
on incorporating real-time links between their analysis software and CAD design
programs. ARCADY 7.1 currently interfaces with the CAD package AutoTrack
9 Junctions developed by Savoy Computing Services Ltd [44]. Future versions
of SIDRA were demonstrated to link with the CAD software TORUS developed
by Transoft Solutions [45]. RODEL has also been undergoing revisions with the
introduction of RODEL V1-Win as an interim beta software before the release of
RODEL V2.
130

7 Conclusions and Future Work

This research has analyzed current roundabout software in terms of a scientific

perspective of technical accuracy, as well as from a managerial standpoint of
usability and features. Certainly the choice about the future direction for round-
about analysis is a complex decision with widespread impacts that cannot be
taken lightly. The following concluding comments summarize the major trends
identified within this research.

7.1 Capacity Model Comparison

Within the scope of this research, all models were shown to perform well if
properly calibrated as RMSE values were relatively similar ranging between 80
and 183 pcu/h/ln, depending on the scenario. However, calibration using field
data is difficult for future, non-existent conditions when designing a proposed
roundabout, which emphasizes the importance of proper use of uncalibrated
models. Some uncalibrated models, specifically the NCHRP 572 and German mod-
els, showed consistently lower error than others, between 82 and 193 pcu/h/ln.
Caution needs to be used in applying any model, especially when extrapolating
foreign models to the U.S., because scenarios in the U.S. may differ from those used
in the model development. Based on the scenarios evaluated, incorporating the
 131

U.S. based capacity research into roundabout analysis would be desirable. Equally,
situations beyond the scope of the NCHRP 572 research may require alternative
analysis tools as recognized by the HCM 2010.
The U.K. model and ARR 321 method overpredicted capacity in the default
scenarios, although only slightly in some cases, and SIDRA would show lower
capacity from the ARR 321 method used. Specifically at the De Pere site, the U.K.
model overpredicted capacity in cases of high circulating flows where capacity
estimates are most critical for performance measures. This overprediction may
be due to the more conservative driving observed compared to the aggressive
behaviors that are reflected in the U.K. model. Combining data from the two
approaches studied at the De Pere site showed that extending the approach based
U.K. model to a lane-by-lane analysis needs careful consideration because each
lane may not have equal capacity if lane utilization is not balanced. Importantly,
models that overpredict capacity now may not overpredict capacity in the future
as drivers gain more experience in the U.S.

7.2 Software Modeling Comparisons

Software modeling logically followed similar trends to the capacity modeling

analysis. Where capacity models were shown to overpredict capacity, the software
also showed high predictions of capacity, and thus less queuing than what was
observed in the field. The French model implemented in Girabase did not seem
to return results consistent with field observations. Calibration was only possible
in three of the seven software packages: ARCADY, HCS and SIDRA. Calibration
resulted in lower capacity predictions, which was consistent with the capacity data
analysis. However software calibration had varied success, likely due to some of
132

the limitations of the study and software analysis. In the case of the Canal St site,
queues were longer than expected from all models. In the case of the U.K. model,
this likely occurred because of the more sporadic congested time periods rather
than the extended congestion that was observed at the De Pere site.
Lane-by-lane modeling was shown to be a more desirable method for
capacity analysis because significant lane imbalance can result in variability in
performance measures on the same approach, as was observed at the De Pere site.
However, for eastbound De Pere, no software showed lane imbalance in queuing
to the estimated extent observed in the field. This shows that the analyst still has
responsibility for understanding how drivers will utilize the available lanes. On
the northbound approach, queue estimates were low from all software packages in
both the calibrated and uncalibrated scenarios, however, HCS and SIDRA correctly
identified lane imbalance that is not otherwise detectable with an approach based
method.

7.3 Software Usability Comparison

Experience in the U.S. appears to be at the cusp of change in terms of capacity

model development as well as software packages. This research has come at the
beginning of major capacity studies and certainly does not mark a definitive end as
much remains to be learned about the future of roundabout operations. Ultimately
software is constantly evolving and this research considered the most widespread
versions of the popular analysis packages; new versions, or entire packages may
emerge in the future. Therefore, the purpose of the software as well as the potential
growth, evolution, and flexibility of any software and parent company must be
considered to make a smart investment for the future. Likely there may not be
 133

one perfect solution, so a variety of the most useful software tools that fulfill
specific roles should be considered. Realistically an analyst could become used
to any software after a sufficient amount of experience, which makes comparing
usability difficult. However, taking the perspective of an occasional user, usability
varied from simple but less feature rich packages, like RCAT and GIRABASE, to
complex packages and feature rich packages like ARCADY. Larger companies,
such as those that produce ARCADY and SIDRA, seemed to offer more frequent
updates, support, and features although at a greater cost.

7.4 Future Research

During this study, numerous questions and areas for future research were iden-
tified, further emphasizing that current practice is only at the beginning of fully
understanding roundabout operations. Some potential future research areas are:

• More intensive research into delay and queue models. Even long queues
tended to roll leading to questions regarding the definition of queuing and
associated models during congestion;

• Study roundabout operational parameters and performance measures to test

sensitivity to location specific factors, such as urban versus rural, regional
differences, etc;

• Understanding how roundabouts operate within a corridor of other round-

abouts or a corridor with mixed intersection types;

• Expand the analyses and comparisons to include microsimulation; and

134

• Use the lessons learned from this research toward future studies when more
roundabouts approach capacity operations to refine calibration of model
parameters.

7.5 Other Considerations for Discussion

Beyond specific future research ideas, many other pertinent questions, that are not
easily answered, should be considered:

• What is the purpose of the model and software output? Is the purpose just to
obtain LOS or is it needed to determine geometric design parameters? More
than one type of software may be appropriate.

• How will driver behavior change in the future? Will gap acceptance param-
eters change or will more aggression be observed? Given the unpredictable
nature of the future, should capacity estimates be based on a range of values
rather than one average value?

• How appropriate is it to use foreign models in the U.S. when geometric

design and driver behavior may differ on fundamental levels?

• Are there any concerns over choosing a model with proprietary, unpublished
functions? Understandably there is need to protect intellectual property, but
there is also a need for the analyst to be able to check and make sure the
underlying research and models apply to site specific situations to avoid
unsubstantiated software output.

• The relative merits between default models that work right “out-of-the
box” versus models that require adjustment or calibration deserves careful
 135

consideration as development of proper adjustments for any model requires

extensive data collection and reduction.

• How will the software fit into the workflow for roundabout design and anal-
ysis? Better understanding how the software is intended to be used within
a larger roundabout design and analysis workflow could help refine the
choices. The full impact of choosing an analysis tool should be investigated.
Extensive "what-if" testing to see how past decisions may have changed with
new analyses may be useful.

• Should the software be used as a compliment to established guidelines

for good roundabout design? Enhanced guidelines may lead to more
consistency between roundabouts which could reinforce driver expectation
and understanding and may in turn increase safety as well as capacity.

Clearly, there are many aspects that require careful thought when evaluating
roundabout, or any other type of, analysis software. Approaching problems from
a scientific and open minded perspective helps in making informed decisions to
provide the most value for investing in the transportation system.
136

References

[1] Highway Capacity Manual. TRB, National Research Council, Washington,

D.C., 2010.

[2] Miller, A. Nine Estimators of Gap-Acceptance Parameters. In Traffic Flow

and Transportation, 1972, pp. 215-236.

[3] Brilon, W., R. Koenig, and R. J. Troutbeck. Useful Estimation Procedures for
Critical Gaps. In Transportation Research Part A: Policy and Practice, Vol. 33,
No. 3-4, Elsevier, 1999, pp. 161-186.

[4] Troutbeck, R. J. Estimating the Critical Acceptance Gap from Traffic

Movements. In Physical Infrastructure Centre Research Report, Queensland
University of Technology. Physical Infrastructure Center, Queensland, 1992,
pp. 1-23.

[5] Rodegerdts, L., M. Blogg, E. Wemple, E. Myers, M. Kyte, M. Dixon, G. List,

A. Flannery, R. J. Troutbeck, W. Brilon, and Others. NCHRP Report 572:
Roundabouts in the United States, Transportation Research Board of the
National Academies, Washington, D.C., 2007.

[6] Wisconsin Department of Transportation Facilities Development Manual.

Wisconsin Department of Transportation, Madison, Wisconsin.
 137

http://roadwaystandards.dot.wi.gov/standards/fdm/. Accessed July

2010.

[7] Kimber, R. M. The Traffic Capacity of Roundabouts. Laboratory Report 942.

Transport and Road Research Laboratory, Crowthorne, Berkshire, United
Kingdom, 1980.

[8] Brilon, W. Roundabouts: A State of the Art in Germany. Presented at the

National Roundabout Conference. Vail, Colorado. May 22-25, 2005.

[9] FGSV. Handbuch für die Bemessung von Straßenverkehrsanlagen (German

Highway Capacity Manual). Forschungsgesellschaft für Straßen- and
Verkehrswesen (Hrsg.), No. 299. FGSV Verlag GmbH, Koln, Germany, 2001.

[10] Louah, G. Panorama Critique des Modeles Francais de Capacite des

Carrefours Giratoires. Proc., Roundabouts 92, Nantes, France, October 1992.
SETRA, Bagneux, France, February 1993.

[11] Akcelik, R., E. Chung, and M. Besley. Roundabouts: Capacity and

Performance Analysis. Research Report ARR No. 321, 2nd ed. ARRB
Transport Research Ltd, Australia, 1999.

[12] Wu, N. A Universal Procedure for Capacity Determination at Unsignalized

(priority-controlled) Intersections. Transportation Research Part B, No. 35,
Issue 3. Elsevier Science Ltd, New York, Tokyo, Oxford, 2001.

[13] Rodegerdts, L. NCHRP 3-65: Applying Roundabouts in the United States:

Status Report to the Committee on Highway Capacity and Quality of
Service. Transportation Research Board of the National Academies,
Washington, D.C., 2004.
138

[14] Rodegerdts, L., M. Blogg, E. Wemple, E. Myers, M. Kyte, M. Dixon, G. List,

A. Flannery, R. J. Troutbeck, W. Brilon, and Others. NCHRP Web-Only
Document 94: Appendixes to Report 572: Roundabouts in the United States,
Transportation Research Board of the National Academies, Washington,
D.C., 2007.
http://onlinepubs.trb.org/onlinepubs/nchrp/nchrp_w94.pdf. Accessed
October 2011.

[15] Waddell, E. Evolution of Roundabout Technology: A History-Based

Literature Review. www.k-state.edu/roundabouts/research/Waddell.pdf.
Accessed October 2011.

[16] Akcelik, R. Roundabouts-Comments on the SIDRA Intersection Model and

the UK TRL Linear Regression Model.http://www.sidrasolutions.com/
documents/SIDRA_and_UK_Roundabout_Models.pdf. Akcelik & Associates
Party, Ltd. April 2011. Accessed October 2011.

[17] Akcelik, R. Some Common and Differing Aspects of Alternative Models for
Roundabout Capacity and Performance Estimation. Paper presented at the
3rd International Conference on Roundabouts. Carmel, IN, 2011.

[18] Kimber, R. M. Gap-Acceptance and Empiricism in Capacity Prediction.

Transportation Science. May 1989, Vol. 23, No. 2, pp. 100-111.

[19] Transport Research Laboratory (TRL) Software. Roundabout Design for

Capacity and Safety: The UK Empirical Methodology.
https://www.trlsoftware.co.UK/products/junction_signal_design/
arcady/roundabout_design_for_capacity_and_safety. Accessed October
2011.
 139

[20] Jacquemart, G. NCHRP Synthesis of Highway Practice 264: Modern

Roundabout Practice in the United States. Transportation Research Board of
the National Academies, Washington, D.C., 1998.

[21] Robinson, B. W., L. Rodegerdts, W. Scarbrough, W. Kittelson, R. Troutbeck,

W., Brilon, L. Bondzio, K. Courage, M. Kyte, J. Mason, A. Flannery, E. Myers,
J. Bunker, and G. Jacquemart. Roundabouts: An Informational Guide. Report
FHWA-RD-00-067. FHWA, U. S. Department of Transportation, June 2000.

[22] Highway Capacity Manual. TRB, National Research Council, Washington,

DC, 2000.

[23] Brilon, W. Studies on Roundabouts in Germany: Lessons Learned. Paper

submission to the 3rd International Conference on Roundabouts, Carmel, IN,
2011.

[24] Mauro, R. Calculation of Roundabouts Capacity, Waiting Phenomena and

Reliability. Springer-Verlag, New York, 2010.

[25] Girabase User Guide, version 1.4. Centre for Studies on Urban Planning,
Transportation and Public Facilities (CERTU). Lyon, France. 2006.

[26] SIDRA Intersection 5.1 User Guide. Akcelik and Associates Pty Ltd.,
Melbourne, Australia, 2011.

[27] Schroeder, B., et al. Empirical Evaluation of Lane-by-Lane Capacities of

Congested Triple-Lane Roundabout Approaches. Presented at 3rd
International Conference on Roundabouts, Carmel, IN, 2011.
140

[28] Wei, T., et al. Developing Capacity Models for Local Roundabouts: A
Streamlined Process. Presented at 3rd International Conference on
Roundabouts, Carmel, IN, 2011.

[29] Jones, S. L., et al. Traffic Simulation Software Comparison Study. Publication
02217. University Transportation Center for Alabama, June 2004.

[30] Collura J., et al. Simulation Models for Assessment of the Impacts of
Strategies for Highway Work Zones. In Transportation Research Record:
Journal of the Transportation Research Board, No. 2169, Transportation
Research Board of the National Academies, Washington, D.C., 2010, pp.
62-69.

[31] Meschner P. J., et al. Collision Diagram Software Evaluation. The Center for
Transportation Research and Education Iowa State University, June 1997.

[32] Click, S.M. and N. Rouphail. Field Assessment of the Performance of

Computer-Based Signal Timing Models at Individual Intersections in North
Carolina. Final Report for the North Carolina Department of Transportation.
Center for Transportation Engineering Studies, Department of Civil
Engineering, North Carolina State University, 2000.

[33] Lenters, M. Criteria for Choosing a Capacity Model. Personal

communication prepared concurrently for the Georgia Department of
Transportation and the Ministry of Transportation, Ontario Canada, 2010.

[34] RODEL Manual. RODEL Software Ltd.

http://www.rodel-interactive.com/uploads/Rodel_Manual_1-9-9.pdf.
Accessed July 2011.
 141

[35] RODEL Interactive Roundabout Design.

http://www.rodel-interactive.com/news.php. Accessed July 2011.

[36] Assessment of Roundabout Capacity and Delay (ARCADY). TRL Software.

https://www.trlsoftware.co.UK/ProductContent.aspx?ID=1. Accessed
July 2011.

[37] Roundabout Capacity Analysis Tool (RCAT). Diodos Transportation, LLC.

http:
//www.wix.com/ornekm/diodos/page-0#!__page-0/roundabout-analyzer.
Accessed July 2011.

[38] Center for Studies on Networks, Transport, Urban Planning and Public
Buildings (CERTU).
http://www.certu.fr/catalogue/i5/About_Certu/information.html.
Accessed July 2011.

[39] Center for Microcomputers in Transportation (McTrans)

http://mctrans.ce.ufl.edu/aboutMctrans/. Accessed July 2011.

[40] SIDRA Solutions. Akcelik and Associates.

http://www.sidrasolutions.com/akcelik_company_about_us.aspx.
Accessed July 2011.

[41] Porter, B. An Evaluation of Multilane Roundabout Capacity and Delay

Models Using Field Data from Wisconsin. Masters Thesis. University of
Wisconsin Madison, 2010.

[42] Wisconsin Department of Transportation Traffic Count Maps by County.

Wisconsin Department of Transportation, Madison, Wisconsin.
142

http://www.dot.wisconsin.gov/travel/counts/maps.htm. Accessed March

2012.

[43] Santiago, K. Video Event Recorder. University Of Wisconsin – Madison

Traffic Operations and Safety Laboratory, Madison, Wisconsin, 2010.

[44] AutoTrack 9 Junctions with Real Time Data Link to ARCADY 7.1. Savoy
Computing Press Release February 22, 2011.
http://savoy.co.uk/News/PR/pr_AutoTrack_9.pdf. Accessed October 2011.

[45] Seminar Demonstration of SIDRA. 3rd International Conference on

Roundabouts, Carmel, IN, 2011.
 143

Appendix A

De Pere AM Capacity Data - Northbound Approach

 145

De Pere Northbound AM Field Data

(a) (b)
Field Data with Linear Regression Field Data with Exponential Regression
Figure B-1. De Pere Northbound AM Field Capacity Data

Table B-1. De Pere Northbound AM Field Data Regression Results

Intercept RMSE
2
Lane Regression (pcu/h) Slope n R (pcu/h/ln)
Linear 1169 0.595 27 0.706 104
Right
-3
Exponential 1640 1.12×10 27 0.688 110

Linear 1091 0.622 84 0.707 102

Left
-3
Exponential 1603 1.33×10 84 0.656 114

Linear 2255 1.201 24 0.807 74

Approach
(Both Lanes) -3
Exponential 3550 1.28×10 24 0.777 92
146

De Pere Northbound AM – Approach Based Model Comparison

(a) (b)
Field Data with Linear Regression Field Data with Exponential Regression

Figure B-2. De Pere Northbound AM Approach Based Capacity Comparison

Table B-2. RMSE and Model Characteristics from the De Pere NB AM Approach
Analysis

Uncalibrated Calibrated

Intercept RMSE Intercept RMSE

Model (pcu/h) Slope n (pcu/h/ln) (pcu/h) Slope n (pcu/h/ln)
U.K. 2450 0.740 24 333 1829 0.759 24 93
-4 -5
German 2483 1.53×10 24 82 2400 5.56×10 24 94
 147

De Pere Northbound AM – Lane Based Model Comparison

(a) (b)
Uncalibrated Lane Based Models Calibrated Lane Based Models
Figure B-3. De Pere Northbound AM Lane Based Capacity Comparison

Table B-3. RMSE and Model Characteristics from the De Pere NB AM Lane Based
Analysis

Uncalibrated Calibrated

Intercept RMSE Intercept RMSE

Model (pcu/h) Slope n (pcu/h/ln) (pcu/h) Slope n (pcu/h/ln)
ARR 321 1633 N/A 84 381 1133 N/A 84 107
-4 -4
NCHRP 1130 7.00×10 84 149 1165 7.38×10 84 145
 149

Appendix B

Software Output for all Approaches

151
152
153
154
155
156
157