1728 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 21, NO.

4, APRIL 2020

Augmenting Traffic Signal Control Systems for

Urban Road Networks With Connected Vehicles
Craig B. Rafter , Student Member, IEEE, Bani Anvari, Member, IEEE, Simon Box, and Tom Cherrett

Abstract— The increase in traffic volumes in urban areas TABLE I

makes network delay and capacity optimisation challenging. C OUNTRY-L EVEL S UMMARY OF THE E CONOMY-W IDE C OST OF
However, the introduction of connected vehicles in intelligent T RANSPORT D ELAY, AND D ELAY C OST F ORECASTS
transport systems presents unique opportunities for improving U NTIL 2030 (B ILLIONS USD) [1]
traffic flow and reducing delays in urban areas. This paper pro-
poses a novel traffic signal control algorithm called Multi-mode
Adaptive Traffic Signals (MATS) which combines position infor-
mation from connected vehicles with data obtained from existing
inductive loops and signal timing plans in the network to perform
decentralised traffic signal control at urban intersections. The
MATS algorithm is capable of adapting to scenarios with low
numbers of connected vehicles, an area where existing traffic
signal control strategies for connected environments are limited.
Additionally, a framework for testing connected traffic signal
controllers based on a large urban road network in the city
of Birmingham (UK) is presented. The MATS algorithm is
compared with MOVA on a single intersection, and a calibrated
TRANSYT plan on the proposed testing framework. The results
show that the MATS algorithm offers reductions in mean delay impacting on business efficiency. In 2017, INRIX estimated
up to 28% over MOVA, and reductions in mean delay and that traffic congestion cost the combined economies of the
mean numbers of stops of up to 96% and 33% respectively UK, Germany, and the USA $450 billion in lost time and
over TRANSYT, for networks with 0-100% connected vehicle
presence. The MATS algorithm is also shown to be robust under wasted energy [2]. Traffic delays are a significant problem in
non-ideal communication channel conditions, and when heavy developed countries.
traffic demand prevails on the road network. Connected Vehicles (CV) are those that use wireless com-
Index Terms— Intelligent transport systems, connected vehi- munications to share data with other vehicles and infrastruc-
cles, communication systems, traffic signal control, adaptive ture. CVs present unique opportunities to improve urban
signal control, V2I. traffic management systems’ effectiveness at reducing delay.
CVs have the advantage over inductive loops in that they
I. I NTRODUCTION
do not require intrusive roadworks to be undertaken to

I NCREASING traffic volumes in urban areas make network

delay and capacity optimisation challenging. The Centre
for Economics and Business Research (CEBR) estimated the
install infrastructure, such as inductive loops, to use their
data. However, their networking protocols are more com-
plex than those of unconnected vehicles, and they require
cost of traffic delays in 2013 for the UK, France, Germany, fleets to contain significant proportions of CVs before their
and the USA as 20.5, 22.5, 33.43, and 124.158 billion dollars applications become effective. Previous literature does not
respectively (see Table I) and these values are expected to adequately address the issue of traffic signal delay at CV
increase significantly over the next 15 years [1]. CEBR defines penetrations below 50%. The current literature also does
delay cost as the combination of the direct cost of fuel and not properly address the issue of imperfect communication
time wasted, along with the indirect cost resulting from delays channel conditions and testing traffic signal control algorithms
at increasing penetrations of CVs, and in realistic scenarios.
Manuscript received October 31, 2018; revised May 17, 2019,
September 18, 2019, and January 8, 2020; accepted January 24, 2020. Date This paper proposes a novel traffic signal control algorithm
of publication February 12, 2020; date of current version March 27, 2020. called Multi-mode Adaptive Traffic Signals (MATS) which
This work was supported by the Engineering and Physical Sciences Research combines position information from CVs with information
Council in partnership with the Transport Research Laboratory (TRL) under
Grant EP/L015382/1. The Associate Editor for this article was F. Viti. collected through existing inductive loops and fixed-time plans
(Corresponding author: Craig B. Rafter.) to perform decentralised intersection control in urban areas
Craig B. Rafter, Simon Box, and Tom Cherrett are with the Transportation to reduce overall traffic delay. In order to comprehensively
Research Group, Faculty of Engineering and the Environment, University of
Southampton, Southampton SO16 7QF, U.K. (e-mail: [email protected]; test the MATS algorithm, a microsimulation testing framework
[email protected]; [email protected]). for traffic signal controllers using CV data is presented. The
Bani Anvari is with the Centre for Transport Studies, Department of Civil, testing framework is based on a large urban road network in
Environmental and Geomatic Engineering, Faculty of Engineering, University
College London, London WC1E 6BT, U.K. (e-mail: [email protected]). the city of Birmingham (UK). It highlights how traffic signal
Digital Object Identifier 10.1109/TITS.2020.2971540 control algorithms should be evaluated under varying traffic
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
RAFTER et al.: AUGMENTING TRAFFIC SIGNAL CONTROL SYSTEMS FOR URBAN ROAD NETWORKS WITH CV 1729

demands of mixed-mode traffic, varying levels of CV penetra- areas where traffic remains similar to the calibration state.
tion, and under imperfect communication channel conditions. Fixed-time plans do not adapt to live traffic conditions, so do
The MATS algorithm is tested against MOVA using the case not perform well where traffic demand varies significantly. The
study in [3]. In addition, a calibrated TRANSYT [4] plan is TRAffic Network StudY Tool (TRANSYT) [4], is one of the
used as a benchmark in the testing framework. The MATS most widely deployed fixed-time optimisation packages still
algorithm is capable of performing under various levels of in modern usage. TRANSYT uses historic flow measurements
connectivity and aims to reduce traffic delay. to generate optimum signal timing plans for both isolated and
The contributions of this paper are as follows: networked intersections. TRANSYT calculates the optimal
1) A new traffic signal control algorithm, MATS, is pro- signal timings for a given road network model by minimising
posed which combines information from existing a performance function consisting of delay, number of stops,
fixed-time plans and loop detectors, and position data and economic factors. TRANSYT has been shown to reduce
from CVs to perform decentralised control on signalised delay up to 24% over pre-existing signal timing plans in the
intersections. The MATS algorithm operates under simi- New England region of the USA [8].
lar principles to the state-of-the-art MOVA algorithm [5], Actuated signal control systems, employ data gathered from
that is to optimise network capacity for saturated condi- roadside sensors such as inductive loops or video cameras
tions and minimise stops and delays for undersaturated to extend the green time of a signal stage between a min-
conditions. The MATS algorithm is novel in that it does imum and maximum limit depending on the current traf-
not require entirely new infrastructure, high penetrations fic conditions. Microprocessor Optimised Vehicle Actuation
of CVs, or ideal data. Rather, the algorithm can be (MOVA) [5] is an example of an actuated signal control
deployed alongside legacy systems to augment them strategy that uses loop detector data to attempts to minimise
with data from CVs, even under non-ideal communica- delay and stops for the entire intersection. MOVA typically
tion channel conditions, an area where existing literature reduces delay at isolated intersections by 13% on average over
is limited. other actuated systems [9].
2) A testing framework for the microsimulation of traffic Adaptive signal controllers use data from infrastructure in
signal controllers that use CV data is proposed. The real-time to optimise an objective function to reduce traffic
new testing framework overcomes the limitations of delays and congestion. Adaptive strategies can work with
existing tests by implementing a large-scale, realistic both isolated (decentralised) intersections, and with groups
simulation case study, which accounts for mixed-mode of signal controllers to reduce traffic delays. Currently used
traffic, multiple levels of traffic demand, degraded loop adaptive signal controllers include: SCOOT [10], SCATS [11]
detector coverage, and a full 24-hour simulation period. and InSync [12]. In the UK, SCOOT has been shown to reduce
3) A communication channel noise characterisation system traffic delays by 12% on average, but up to 33% compared with
for the testing framework is proposed. In this paper, TRANSYT, and 26% on average but up to 48% compared with
non-ideal communication channel conditions are more an isolated vehicle actuation scheme [13].
comprehensively treated than in previous simulation There has been a great deal of success in reducing delays
studies. Typically, only channel latency is considered. with adaptive traffic signal control strategies that use data from
Here, GPS measurement error and packet loss are also infrastructure to best respond to varying traffic demand on
included in the communication channel error model and the road network. The introduction of CVs and the concept
is shown to have a significant impact on the connected of Connected Intelligent Transport Systems (C-ITS) present
signal control strategy. exciting opportunities for innovation in traffic signal control.
Section I-A reviews traffic signal control literature while CV systems are inherently well suited to mitigate delay,
Section II explains the proposed MATS algorithm. The algo- as CVs are an abundant source of data for Adaptive Traffic
rithm is tested using the proposed microsimulation testing Signal Control Systems (ATSCS). ATSCSs are more beneficial
framework based on a large-scale urban road network in than traditional traffic control strategies at reducing traffic
the city of Birmingham (UK), at varying levels of traffic delay [14], especially in urban areas with fluctuating traffic
demand, at CV penetrations from 0-100%, and under imper- demands. This paper focuses on using data sent from CVs to
fect communication channel conditions in Section III. The infrastructure to improve decentralised adaptive traffic signal
microsimulation results comparing the MATS algorithm with control in urban areas.
MOVA, and a calibrated TRANSYT plan on two case studies Recent research has developed traffic signal controllers
are presented in Section IV. Finally, the conclusions and for C-ITS environments (see [15] and [16] for reviews)
avenues for future work are drawn in Section V. which assume perfect communication between vehicles and
infrastructure, or require all of the vehicles in the network
A. Related Work to be connected as in the slot-based reservation system of
Effective traffic signal control strategies for urban road Fajardo et al. [17]. As CVs are only anticipated to be intro-
networks have been well-studied [6], [7]. There are three duced into the road network from 2020 onward, it will take
approaches to traffic signal control that are currently used: time for the vehicle fleet to become fully connected [18], and
fixed-time, actuated, and adaptive. hence there is a need to develop signal control strategies that
Fixed-time traffic control systems create optimised support this transition period. Other traffic signal controllers
signal-timing plans from historical data and are suitable in for CVs rely entirely on CV data, with limited consideration
1730 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 21, NO. 4, APRIL 2020

for unconnected vehicles such as in forecast based depar- the MATS algorithm reduces delays in undersaturated condi-
ture strategy optimisation [19]. More recent traffic signal tions, and increases capacity in saturated conditions. In addi-
control algorithms have begun to consider both connected tion, the MATS algorithm uses speed, position, and heading
and unconnected data sources. Beak et al. [20] used stop data from CVs in combination with fixed-time plans and data
bar detectors to supplement an adaptive phase optimisation from inductive loop sensors to actuate signal timings, to detect
strategy using CV data at CV penetrations as low as 25% blocking back, and to estimate queue lengths.
using a perfect communication system. Ilgin Guler et al. [21] The MATS algorithm bridges the gap between existing
proposed an algorithm to enumerate and optimise discharge and future technologies for traffic management by offering
sequences to reduce delay and tested it at CV penetrations multiple modes of operation based on what sources of data
from 0-100%. However, they only test an intersection with are available. At its lowest level of operation, it operates a
two one-way streets at two demand levels, and with perfect fixed-time plan in the absence of data from CVs or roadside
communications. Yang et al. [22], went further to incorporate infrastructure. As data from loop detectors and CVs becomes
loop detectors and trajectory data into their signal control available, the MATS algorithm adapts its mode of operation
optimisation strategy where varying levels of CV penetration to actuate signal timings using the gathered data. Furthermore,
and GPS errors were considered, across a single intersection it can respond to traffic demand in real-time and preserves
at two static demand levels. Common limitations in both the driver privacy as it does not require individual drivers to be
algorithms and testing frameworks from these studies were: tracked through the network. Also, it builds on established
1) only cars were simulated, 2) the road network size was small traffic management techniques and uses optimisation/heuristic
(4 intersections on average), 3) CV data were perfect or only procedures that are clearly defined, making the algorithm intu-
had one source of error, 4) the parameter space was limited, itive for transport planners to deploy. To increase capacity and
and 5) the impacts of roadside and connected infrastructure reduce delays in the network, the MATS algorithm maintains
degradation are ignored. a cyclic stage pattern and reduces the load on the downstream
The research gaps identified in the literature are addressed intersection rather than modifying its stage to serve stages with
in this paper in two ways: 1) the MATS control algorithm high demand (i.e. in back-pressure routing). By synthesising
that has been developed, and 2) a realistic testing framework fixed-time plans, loop detector data, and CV data into a single
for comprehensively testing traffic signal control strategies for algorithm, delays and stops are minimised for road users.
connected environments. The MATS algorithm addresses the In the following subsections, data acquisition and
issue of reducing delay at existing traffic signal control sites intersection control in the MATS algorithm are detailed. Data
in environments with increasing numbers of CVs and where acquisition considers the management of available data from
existing infrastructure is degraded. The MATS algorithm does connected sources. Intersection control considers integrat-
this in a novel way that combines 3 data sources (fixed-time ing the gathered data into traffic signal control and timing
plans, loop detectors, and CVs) rather than two, as is typical decisions.
in the literature. This paper completes the concepts introduced
in [23], [24] by modifying the algorithm to be robust in B. Vehicle Data Acquisition
real-world networks, addressing mode-switching issues, and Vehicle data acquisition determines which data originate
performing simulations under a comprehensive testing frame- from vehicles in the junction’s control region, determining the
work. The proposed microsimulation testing framework is queue length on routes that are not inactive, and the locations
unique in that it combines data from the Birmingham and West and speeds of the vehicles on the active lanes. A junctions’
Midlands traffic data portal [25] with OpenStreetMap (OSM) control region is the area surrounding the junction in which
data [26] to create a large-scale, current, and realistic sim- wireless communications are possible. If another junction
ulation case study. The testing framework overcomes the exists inside the control region, the boundary is cropped to the
limitations identified in the literature by allowing traffic signal conflicting junction’s nearest stop line. The boundary reduction
control algorithms to be tested with increasing levels of CV covers the widest possible control region while allowing data
penetration, mixed-mode traffic, and multiple traffic demands from vehicles associated with other junctions to be ignored.
over a 24-hour period. Furthermore, the three main issues The junction controller receives data from all vehicles inside
that create imperfect communication channel conditions (GPS its control region, ignoring those that are not.
measurement error, channel latency, and packet loss) are The junction controller monitors time-dependent data
addressed, which has been lacking in the literature. regarding the vehicles’ positions, headings, and speeds. The
junction controller has knowledge of its own layout/map and
II. T HE M ULTI -M ODE A DAPTIVE T RAFFIC can determine the headings that correspond to an approach
S IGNALS C ONTROL A LGORITHM on each of its lanes. Vehicles in range of the junction
and travelling with headings matching one of the known
A. Concept approaches (± a tolerance to allow for GPS positioning error)
The MATS algorithm builds upon the principles for man- are considered to be approaching the junction.
aging oversaturated and undersaturated flows from the state-
of-the-art vehicle actuated control algorithm – MOVA [5], C. Intersection Control
and extends those principles with blocking back detection and 1) Initial Stage Time: The initial stage time is defined based
queue length estimation using CV data. Similar to MOVA, on the length of the queue in inactive lanes. Control strategies
RAFTER et al.: AUGMENTING TRAFFIC SIGNAL CONTROL SYSTEMS FOR URBAN ROAD NETWORKS WITH CV 1731

frequently use queue length estimates as a parameter, and as a

quantity that is desirable to minimise [19]. Queue lengths are
determined from the distance of the furthest queuing vehicle
from the intersection. A vehicle is considered to be queuing if
its speed is less than 0.01 m/s (inferring that vehicles travelling
so slowly are at or approaching the end of the queue). The
queue clearance time for a lane is given by:
lqueue
tclear,queue = × tgreen,max (1)
lqueue,max
where tclear,queue is the queue clearance time, lqueue is the
queue length, tgreen,max is the maximum green time a stage
can have, and lqueue,max is the maximum length a queue may
have. Setting the queue clearance time in this way means that
as the queue length tends towards the maximum range of the Fig. 1. Flowchart for the MATS algorithm.
communication system, the initial green time tends towards
the maximum green time. The queue clearance calculation
is unique in that, neglects the start-up loss time that drivers time to clear the blocking traffic. Compared with back-pressure
need to react and accelerate. The reason start-up loss is routing approaches [30], the MATS algorithm maintains a
neglected is that the stage time is extended by the presence cyclic stage pattern and reduces the load on the downstream
of the connected vehicle at the tail of the queue if it has not intersection rather than modifying its stage to serve stages with
crossed the stop line. This allows the preliminary green time ‘high-pressure’.
to be automatically corrected if the queue clears slower than 3) Inductive Loop Data Integration: The green time exten-
expected. In comparison, the MOVA algorithm uses a queue sions are applied when vehicles are detected in real-time
length estimated from vehicle counts over its detectors, so the on the existing inductive loops in the road network. The
locations of each inductive loop restrict its estimation. MATS algorithm extends the stage by one extension interval
2) Blocking-Back Detection: Blocking-back occurs at if a vehicle is detected in the previous extension interval.
neighbouring intersections where queues of vehicles at a The actuation behaviour is defined based on the actuated
downstream intersection are long enough to obstruct subse- timing parameter recommendations of the Federal Highways
quent vehicles from joining the queue. Blocking-back can Administration Signal Timing Manual (STM) [31].
cause gridlock if traffic cannot proceed in any direction [27], 4) CV Data Integration: Real-time information from CVs
and is typically alleviated through signal coordination. For is used to derive a stage extension time dynamically. If a CV
example, the SCOOT algorithm measures the proportion of is detected close to the intersection in an active lane, the time
the cycle time where queues occupy its detectors. The queuing it takes for that vehicle to reach the intersections is estimated
information is passed to the optimiser, which then minimises from the driver’s current speed and position. This time is added
the likelihood of the upstream junction creating a blocking to the stage duration if it satisfies the acceptable travel time
queue [10]. Even though blocking-back is a well-understood requirements set by Highways England [32]. The acceptable
problem [28], the literature on traffic signal control for travel time factor is 1.67 times the free flow journey time. This
CVs appears to ignore the issue. In literature for traffic factor times the average time headway between vehicles gives
signal control strategies for CVs, Goodall et al. [19] and the time threshold for green extensions. The time for a CV to
He et al. [29] were the only studies to consider blocking-back. clear the intersection is:
Goodall et al. [19] detected blocking-back using CV data. d(xv , xi )
If vehicles were blocking a movement, then the movement tclear,CV = (2)
v vehicle
that clears the blocking vehicles was given higher priority.
In He et al. [29], vehicle platoon movements and ques lengths where tclear,CV is the time it takes a CV to clear the
were used to prevent the creation of queue spillback that would intersection. d(xv , xi ) is the Euclidean distance between the
cause blocking-back. 2-D Cartesian coordinates for the positions of the vehicle (xv )
Here the control is decentralised, so a method of and the intersection xi in meters. v vehicle is the speed of the
locally detecting blocking-back is developed. Blocking-back vehicle. This approach achieves control that is functionally
is detected by the MATS algorithm using CV position and similar to MOVA in that if continuous vehicle flow is present
speed data to determine if the vehicles are stationary during (oversaturation), the algorithm allows vehicles to proceed
a stage that should permit the vehicles to travel. If blocking- until the maximum green time is reached or the queue is
back is detected, the MATS algorithm ends the current stage dispersed, which maximises capacity. If the vehicle flow is
to allow vehicles in other lanes to traverse the junction undersaturated, the MATS algorithm allows vehicles to pass
on unobstructed routes. Although stage cancelling reduces as long as unacceptable gaps do not appear in the flow.
service to the vehicles in the cancelled stage, it gives vehicles 5) Algorithm Overview: Figure 1 shows the flowchart for
in other stages the opportunity to use the intersection to the MATS algorithm and highlights how the components of
increase throughput and gives the downstream intersection the algorithm integrate, and how MATS switches its mode of
1732 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 21, NO. 4, APRIL 2020

Fig. 2. SUMO model of the T-junction type intersection used in [3].

operation based on which data sources it has available to it.

In order to reduce the computational load, the algorithm only
makes control decisions if the remaining green time is less
than a check threshold, which here is 5 s. The MATS algorithm
extends traditional vehicle actuated systems by incorporating
multiple data sources to compensate for the loss of CV or
inductive loop data. CV data is only used if the CV penetration
is high enough that using the data provides performance
superior to fixed-time control. Fig. 3. SUMO representation of the Selly Oak road network. Intersections
The pseudocode for the MATS algorithm is presented in with traffic signals are highlighted with the red-amber-green light block (12 in
total). The locations of the inductive loops are marked with yellow rectangles.
Algorithm 1. The semantics for the pseudocode are based
on a combination of the Python programming language [33]
and American Mathematical Society notation [34], where ‘//’ Selly Oak road network (Birmingham, UK) in SUMO [35].
infers a comment rather than a command, and ‘DO’ describes The extent of the network and the number of intersections are
in plain English an action to be taken by an external part of greater than the models typically used in the literature. The
the program. Four-Step Model (FSM) [36] approach was used for demand
modelling in this study and is described in the following
sections.
III. T ESTING F RAMEWORK
a) Trip generation: From the flow data publicly available
A. Case Study 1 in Birmingham and West Midlands traffic data portal, the time-
Figure 2 illustrates the T-junction type intersection used to dependent frequency of use for each origin and destination
assess the performance of the MOVA algorithm in [3]. The point in the network was estimated. In this study, flow data
OD matrix for the model is given in Table II. The OD matrix of weekdays (Monday-Friday, excluding holidays) in 2016 and
is reported to yield flows that operate the intersection close to 2017 were used. Each detector was matched to its correspond-
its saturation point. ing origin or destination lane in the road network model. The
flow information was then translated to the total trips per hour
B. Case Study 2: A Realistic Testing Framework going to/from the corresponding lane over 24 hours.
b) Trip distribution: Using the information from the
The microsimulation testing framework presented in this
previous step, the trips were distributed between connected
paper is based on the road network in the Selly Oak
Origin-Destination (OD) pairs, and an OD matrix was formed
area of Birmingham (UK): the roadway from Selly Oak
from collected flow data. As there were no prior travel survey
(latitude/longitude: 52.439177, −1.940248) to the Warwick-
information or turn counts available for the network, the initial
shire Country Cricket Club (latitude/longitude: 52.455288,
traffic was assigned proportionally using the methodology
−1.907067) (see Figure 1). This area covers 8.26 km 2 with
originally proposed by Robillard [37]. The OD matrix was
12 signalised intersections, and 64 inductive loop detectors.
then calibrated using the Furness method [38], to ensure the
This area is chosen since it covers the route with the highest
OD matrix was consistent with the loop detector flow data
number of working loop detectors within Birmingham. Also,
after assignment. All vehicle flows are uniformly distributed
traffic signal data from this area were available for public
across their insertion intervals.
use [25]. Additionally, high volumes of traffic were observed
c) Mode choice assignment: With the trips determined,
in this area due to the presence of large retail and residential
the next step was to allocate each trip a mode of travel (e.g.
centres, and key sites of trip generation such as a 1000+ bed
car, motorcycle, Light Goods Vehicle (LGV), Heavy Goods
hospital and the University of Birmingham. The specification
Vehicle (HGV)), bus). The UK Department for Transport
of the testing framework and details of how it overcomes
provides information about the distribution of different vehicle
the limitations of previous work identified in Section I are
types on a regional basis in the VEH0104 dataset [39]. The
presented in the following sections.
distribution of vehicles registered in the West Midlands area
1) Road Network Modelling: A road network model for this
of the UK, where Birmingham is located, is:
testing framework was created using OpenStreetMap (OSM)
data [26] and vehicle flow data from the Birmingham and West Cars 82.7% Motorcycles 3.0% LGVs 12.3%
Midlands traffic data portal [25]. Figure 3 shows the modelled HGVs 1.6% Buses 0.4%
RAFTER et al.: AUGMENTING TRAFFIC SIGNAL CONTROL SYSTEMS FOR URBAN ROAD NETWORKS WITH CV 1733

Algorithm 1 MATS Algorithm Pseudocode

1 begin MATS
2 DO: Gather CV data from the communications channel, collect flow data from inductive loops
3 remainingTime ← stageDuration − elapsedTime
4 if remainingTime ≤ checkThreshold then
// Get loop extension time if loop data available
5 if loopDataForControlledLanes then
6 if ANY(lastDetectTime ≤ extensionThreshold) then
7 loopExtendTime ← loopStageExtension
8 else
9 loopExtendTime ← 0
10 else
11 loopExtendTime ← NONE
// Get CV extension time if CV data available
12 if CVpenetration > CVPthreshold then
13 if nearestVehicleSpeed ≥ 0.01 and nearestVehicleIsInRange then
14 cvExtendTime ← nearestVehicleDistance / nearestVehicleSpeed
15 if cvExtendTime > 2×loopStageExtension then
16 cvExtendTime ← 0
17 else
18 cvExtendTime ← 0
19 else
20 cvExtendTime ← NONE
// Select extension from the available data, default to fixed-time plan
21 if loopExtendTime = NONE and cvExtendTime = NONE then
22 stageExtendTime ← max(loopExtendTime, cvExtendTime)
23 else if loopExtendTime = NONE and cvExtendTime = NONE then
24 stageExtendTime ← cvExtendTime
25 else if loopExtendTime = NONE and cvExtendTime = NONE then
26 stageExtendTime ← loopExtendTime
27 else
28 stageExtendTime ← max(0, fixedTimeDuration − elapsedTime)
// Update stage time to fall within the upper and lower green time bounds
29 stageDuration ← elapsedTime + max(stageExtendTime, remainingTime)
30 stageDuration ← max(stageDuration, minGreenTime)
31 stageDuration ← min(stageDuration, maxGreenTime)
// If this is a new stage set a preliminary green time based on the queue length
32 else if newStage and numberOfCVs > 0 then
33 if lastVehicleDistance = NULL then
34 queueClearanceTime ← lastVehicleDistance × (maxGreenTime/maxQueueLength)
35 stageDuration ← max(queueClearanceTime, minGreenTime)
36 stageDuration ← min(queueExtendTime, maxGreenTime)
37 else
38 stageDuration ← minGreenTime
// If no vehicles are moving due to blocking back then end stage
39 else if elapsedTime > minGreenTime and remainingTime > checkThreshold and numberOfCVs > 0 and not queueIsMoving then
40 DO: Set stage to end
41 else
42 DO: Continue
// Continue stage if time remaining, else transition to next stage
43 if elapsedTime < stageDuration then
44 elapsedTime ← elapsedTime + timeStep
45 else
46 DO: Transition to next stage
47 elapsedTime ← 0
48 stageDuration ← 0
1734 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 21, NO. 4, APRIL 2020

TABLE II TABLE III

OD M ATRIX FOR THE T-J UNCTION M ODEL . ROWS D ENOTE T HE K RAUSS C AR -F OLLOWING M ODEL PARAMETERS FOR
O RIGINS , C OLUMNS D ENOTE D ESTINATIONS . F LOWS THE C ONSIDERED V EHICLE T YPES [43]
A RE IN V EHICLES PER H OUR

The vehicle type distribution data from the

VEH0104 dataset are used for mode choice assignment,
improving the testing framework’s realism over studies that
only consider passenger cars.
d) Route choice allocation: With the trips and travel
modes determined, the routes used for each trip can be
calculated. The Selly Oak network is a local area with few
alternative paths between origins and destinations. It is reason- considered vehicle modes (cars, motorcycles, LGVs, HGVs,
able to assume that drivers in the network follow the shortest and buses). Both connected and unconnected vehicles have
path to their destination. Dijkstra’s algorithm [40] for finding the same car-following parameters as it is assumed that CVs
the shortest path between two nodes was used for this localised ability to share data with a traffic signal controller does not
road network. affect driver behaviour.
4) Control Strategies: In order to assess the performance
C. Simulation Parameters of the proposed MATS algorithm, the results were compared
Microsimulation was used to test whether the MATS algo- with those from TRANSYT plans calibrated on an average
rithm offers improved intersection management compared to flow case in the Selly Oak area. TRANSYT signal timing
TRANSYT and MOVA. The simulations were performed plans were produced using the TRANSYT 15 software. Sep-
using the open-source SUMO (version 0.30.0) microsimula- arate timing plans were calibrated for off-peak (00:00-06:00,
tion environment [35], and were controlled using Python [33]. 20:00-00:00), peak (06:00-11:00, 16:00-20:00) and inter-peak
The MATS algorithm was tested with and without loop data, flows (11:00-16:00).
and under non-ideal communication channel conditions. The 5) Intersection Configuration: Intergreen times for each
parameter configurations for the various simulation scenarios intersection were set per the UK Government guideline inter-
are explained below. green times [44]. The MATS algorithm (Algorithm 1) was
1) CV Penetration: In order to understand how the number configured with an extension interval of 2 s per the work
of CVs present in the network can affect intersection manage- of Bonneson and McCoy [45]. The minimum and maximum
ment, simulations were run across a range of CV penetrations. green times were 2 and 10 times the intergreen time for
The CV presence in the network was incremented from 0% the intersection, respectively. The junction control region was
to 100% in steps of 10%. defined as a circle with radius 250 m [46] centred on the
2) Traffic Demand: The amount of traffic in the network is intersection. The check interval for the MATS algorithm
a contributory factor in determining how effective a traffic was 5 s, not 1 s, to allow sufficient time for decisions in
signal control strategy is. Case Study 1 uses one hour of the event of long communication latencies and high levels of
static demand that is close to the intersections saturation packet loss.
point. Case Study 2 tests 24-hours of low, average, and high 6) Stochastic Effects: Many of the processes within the
flow levels so that the change in performance of the signal simulation rely on randomness to generate values, especially
control strategies at varying demand levels can be assessed. the traffic generation process. All random number generators
By analysing the data collected from the loop detectors in the used in the codes for this research were drawn from seeded
Selly Oak study, the base case shows average flow levels. The uniform distributions so that the results are repeatable. As the
high and low demand cases were defined as being ±20% of system is stochastic, each simulation must be repeated to
the average flow experienced by the detectors, respectively. create a sample space, and the results averaged in order to
3) Car-Following Model: Here, the Krauss model [41] determine the typical performance resulting from the under-
was used as the car-following model as it produces stable lying system dynamics. The sample space for this research
collision-free traffic flow and is well validated. The Krauss contains 50 repetitions per experiment.
model has been shown to outperform other traffic-flow models 7) Communication Channel, Errors, and Delays: Where
in mixed traffic scenarios [42]. It is also stable at the 0.1 s CVs transmit information, data were sent at a rate of
simulation time-step, which was needed to represent the wire- 10 Hz based on the ETSI CAM [47] specification. Mes-
less communication system dynamics. Table III describes the sages were transmitted via an IEEE 802.11p [48] Dedi-
parameters used in the Krauss car-following model for the five cated Short-Range Communication (DSRC) channel. Research
RAFTER et al.: AUGMENTING TRAFFIC SIGNAL CONTROL SYSTEMS FOR URBAN ROAD NETWORKS WITH CV 1735

on IEEE 802.11p networks shows that signal strength is

high enough within a 250 m radius to allow messages to
be received correctly [46], [49], and that packet latency of
approximately 50 ms are achievable at vehicles speeds of up
to 90 km/h [49]. In this research, CAMs were received by the
intersection controller with ideal information content, but with
a delay of 100 ms.
In order to assess the lower-bound performance of the
MATS algorithm, it was tested under non-ideal conditions.
In the non-ideal case, the MATS algorithm was tested with a
CAM generation rate of 1 Hz instead of the usual 10 Hz. The
packet loss in the system was set to 50%. Finally, Gaussian
noise of the form X ∼ N (μ, σ 2 ), with mean μ = 0 and
variance σ 2 = 2.79, was added to GPS measurements (i.e. the Fig. 4. Comparison of the mean delay of the MATS algorithm with MOVA
position ±5 m in both the x and y coordinates, typical for on Case Study 1.
differential GPS systems [50]).

IV. R ESULTS AND D ISCUSSION data, three varieties of the MATS algorithm are defined and
compared:
A. Performance Indicators
• MATS-FT: The MATS algorithm combining data from
Mean travel time delay and mean stops were selected as the fixed-time plans and CVs
performance indicators for this research. Delay and stops are • MATS-HA: The MATS algorithm with hybrid actuation,
primary components on which TRANSYT optimises signal combining data from fixed-time plans, inductive loops,
timings [51] and allow comparison. Free-flow travel times and CVs
were the basis for delay calculations. In this study, free-flow • MATS-ERR: MATS-FT but under imperfect communi-
travel time is the time a vehicle takes to make its journey at the cation channel conditions.
speed limit, unimpeded by external factors such as other traffic The fixed time plan was derived from the TRANSYT plan.
or signalised intersections. Travel-time delay characterises the The times given by the TRANSYT plan were truncated in
excess time a vehicle takes to complete its journey compared the MATS algorithm if they exceed the junction’s maximum
to the free-flow travel time along the same route. As simulation green time value. During initial testing, it was found that
was used to study the traffic dynamics, the time delay Tdelay the use of loop data was detrimental to the performance of
for a vehicle is defined as: the MATS-HA variant at 0% CV penetration. The negative
Tdelay = Texit − Tenter − Tfreeflow (3) behaviour was due to placing the algorithm in a network with
imperfect coverage. In the case study, the loop placement was
where Tenter and Texit are the times a vehicle enters and exits for a SCOOT system. It was found during testing that the
the simulation, respectively. Tfreeflow is the time it takes the loop coverage is too degraded to be beneficial at 0% CV
vehicle to make its journey on an unobstructed route. Delay penetration, so loop detectors are only used in the presence of
time indicates the amount of time actually saved compared CV data for this case study, and the base level of performance
to the complete journey time and highlights the performance is fixed-time.
limitations of each method. In Sections IV-B and IV-C, the results of the tests on the
In this study, a vehicle is defined as being stopped if its two case studies are presented. First, the results comparing the
speed is less than 0.01 m/s. The total number of stops a vehicle MATS algorithm to MOVA are discussed. Secondly, the results
makes on its journey were recorded for analysis. To normalise comparing the MATS algorithm to TRANSYT on the realistic
the results in Case Study 2, the mean delay and mean stops are testing framework are discussed.
represented per kilometre. The mean data points are banded
by the 5th and 95th percentiles of their corresponding dataset
as indicators of variability. B. Case Study 1: Comparison With an Actuated Controller
The results also compare the mean delays and the mean The MATS algorithm is compared to the state-of-the-art
stops in terms of the percentage reduction between them. The vehicle actuated signal controller of MOVA using the single
percentage reduction is calculated by: intersection case study developed in [3]. In that study, it was
  demonstrated that at a single intersection with ideal loop detec-
x
100 1 − (4) tor data and traffic conditions just below saturation, the average
x ref delay is 20.3 s using MOVA. Figure 4 shows the difference
which yields the percentage reduction between a result value in delay between MOVA and the MATS algorithms for CV
from the MATS algorithm x, and the corresponding reference penetrations from 0%–100%. Table IV shows the percentage
value from the MOVA or TRANSYT results, x ref . difference in delay between the MATS algorithm and MOVA.
To establish how the performance of the MATS algorithm Under the same traffic conditions and ideal communication
differs depending on the quality and availability of input conditions, the MATS-FT algorithm showed lower mean delay
1736 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 21, NO. 4, APRIL 2020

TABLE IV TABLE V
B ENCHMARK R ESULTS OF THE MATS A LGORITHM A GAINST MOVA ON B ENCHMARK R ESULTS OF THE MATS A LGORITHM A GAINST TRANSYT
C ASE S TUDY 1. T HE P ERCENTAGE R EDUCTION IN AVERAGE D ELAY AT FOR THE T HREE D EMAND C ASES . T HE P ERCENTAGE R EDUCTION IN
10%, 30% 50%, 70% AND 100% CV P ENETRATION A RE S HOWN AVERAGE D ELAY AND AVERAGE N UMBER OF S TOPS AT 10%, 50%,
AND 100% CV P ENETRATION A RE S HOWN

than MOVA above 20% CV penetration, with mean delay

reductions of 20%-28% above 30% CV penetration. Under
non-ideal communication conditions, the MATS algorithm
reduces mean delay better than MOVA above 40% CV pen-
etration, with reductions in mean delay between 19%-29%
above 40% CV penetration. When both inductive loop and
CV data are used in the MATS-HA algorithm, the MATS-HA
algorithm reduces mean delay between 12%-15% compared
with MOVA for CV penetrations ≥10%. The MATS-FT and
MATS-ERR use fixed-time plans up to 10% and 20% CV
penetration, respectively, indicating that for this case study,
there is a threshold below which CV data is not beneficial. The
MATS-HA algorithm does the best at 0% CV penetration as it
can use loop detectors to actuate signal timings, but is worse
than MOVA as its actuation strategy is not as sophisticated.
The results show that the MATS algorithm is better than the
state-of-the-art vehicle actuation strategy MOVA, and that loop
detector data is useful at low CV penetrations but can limit which stop many times, stop less frequently when there are
performance at high CV penetrations. CVs present in the network. The variability remains wide even
at high CV penetrations due to the varied route lengths in the
network resulting from its large size.
C. Case Study 2: Test on a Realistic Road Network In addition, there are visible peaks in the MATS-FT and
Tables V show the results from the benchmarking of MATS MATS-HA variants at 10% CV penetration in the high demand
against TRANSYT on the low, average, and high traffic case. The peaks occur as a result of frequent switching
demand cases. The numbers in Tables V(A), (B), and (C) show between fixed-time and adaptive modes due to the sparsity
the percentage reduction in the average delay and the average of CV data, resulting in a slightly higher frequency of stops.
number of stops compared to TRANSYT at 10%, 50%, and Conversely, MATS-ERR does not exhibit the same peaks due
100% CV penetration. to a relaxation in the frequency of switches between modes
Figure 5 shows the plots of mean delay per kilometre due to communication delays. However, it can be seen that the
and the mean number of stops per kilometre for three traffic behaviour of the MATS algorithm variants stabilises for CV
demand scenarios. As shown in Figure 5, the MATS algorithm penetrations above 10%.
resorts to using the fixed-time plan in the absence of CV data From Table V(B), it can be seen that the MATS algorithm
(0% CV penetration). As shown in Figure 5, the grey lines for provides reductions of up to 35% over TRANSYT under
both delay and number of stops are straight for TRANSYT. average traffic demand. Additionally, the effects of imperfect
These lines are expected to be straight as TRANSYT does not communication conditions do not significantly impact the
use CV data, so its performance is invariant with increases performance of the MATS algorithm in terms of the number of
in CV penetration. The results are discussed in further detail stops made. From Tables V(A) and (C), it can be seen that the
under two categories (stops and delays) in the following MATS algorithm can achieve greater reductions in the number
sections: of stops as the traffic demand increases.
1) Stops: It can be seen in Figures 5(b), (d), and (f) 2) Delay: Looking at results in Figures 5(a), (c), and (e)
that there is little difference in the average number of stops and Tables V(A), (B), and (C), it can be seen that the
made across the MATS algorithm variants compared with MATS algorithm offers significant reductions in delay at all
TRANSYT, even with the addition of inductive loops. Across levels of traffic demand. In addition to reducing the mean
all of the stop results, it can be seen that as the CV penetration delay, the MATS algorithm significantly reduces the delay
increases the variability in the number of stops made by vehi- variability experienced by drivers with CV penetrations above
cles reduces. The reduction in variability shows that vehicles 10%. In the case of imperfect communication conditions,
RAFTER et al.: AUGMENTING TRAFFIC SIGNAL CONTROL SYSTEMS FOR URBAN ROAD NETWORKS WITH CV 1737

Fig. 5. Plots of mean delay per kilometre and mean stops per kilometre for each of the three flow scenarios (low, average, high). Each plot compares the
performance of the MATS algorithm with and without loop information (MATS-FT), and MATS with errors (MATS-ERR), to TRANSYT. The bands on the
data represent the 5th and 95th percentiles of the data as indicators of variability.

the MATS algorithm’s performance is degraded compared to all demand cases. The discrepancy between MATS-ERR
MATS-FT but still offers reductions in delay and variability and the other MATS algorithm variants can be attributed
compared with TRANSYT, emphasising its robustness. Across to MATS-ERR overestimating or underestimating the queue
the results, there is some variability in the delay even at high clearance time and stage extensions due to the noise, error,
CV penetrations due to the varied route lengths in the network and delay in the communication channel. However, the effects
resulting from its large size. of the non-ideal communication channel conditions were
In Figures 5(a), (c), and (e), the MATS-ERR variant overcome by 30% CV penetration. Despite the non-ideal
under-performs the MATS-FT and MATS-HA variants across communication channel, MATS-ERR reduces delay and delay
1738 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 21, NO. 4, APRIL 2020

variability. Interestingly, the mean in the MATS-ERR plot lines V. C ONCLUSION

lies close to the upper-bound of the error bars, suggesting
some skew towards vehicles on longer journeys being less In this paper, the MATS algorithm that augments existing
well served than those on shorter journeys. As with the results traffic signal control systems with CV data was introduced.
for the number of stops, the skew results from lower mode A microsimulation testing framework that comprehensively
switching frequency due to the sparsity of CV data are greater covers the parameters that need to be explored to simulate
under high traffic demand. These results suggest that at low traffic signal controllers that use CV data was also presented.
CV penetrations and higher traffic demands, less frequent data In initial tests, the MATS algorithm reduced delays better than
transmissions reduce the amount of mode switching, which is the MOVA algorithm. Under the testing framework, the MATS
beneficial in decreasing the number of stops but detrimental algorithm was shown to outperform the industry-standard
to reducing delay. TRANSYT traffic signal timing algorithm under increasing
It can be seen from the comparison between the MATS algo- traffic demands through the combination of data from CVs,
rithm and TRANSYT, that under average demand, the MATS existing fixed-time plans, and roadside infrastructure. Further-
algorithm reduces delay up to 96%. Table V(B) indicates that more, the framework highlights that assessing connected traffic
by 10% CV penetration up to 93% delay reduction can be signal control strategies at multiple CV penetrations and in the
achieved, which was most of the maximum achievable delay presence of communication errors is critically important.
reduction of 96%. Additionally, Figures 5(a), (c), and (e) Simulating the MATS algorithm on the introduced testing
show that by 30% CV penetration MATS-ERR reduces delay framework’s realistic simulation case study in Selly Oak,
almost as well as MATS-FT. These results suggest that with Birmingham, UK demonstrated that the MATS algorithm was
a relatively low proportion of CVs in the road network, highly robust across a spectrum of scenarios. The MATS
significant delay reductions can be achieved using the MATS algorithm outperforms TRANSYT at minimising delay across
algorithm even under non-ideal communication channel con- varying traffic demands and offers significant performance
ditions. These findings were consistent when comparing the improvements even in the presence of imperfect communica-
MATS algorithm with TRANSYT under low and high traffic tion conditions and measurement noise. The MATS algorithm
demands seen in Tables V(A) and (C). offers reductions in the average delay of up to 96% and the
Comparing MATS-FT to MATS-HA in Figures 5(a), (c), number of stops up to 33% over TRANSYT, for networks
and (e), the results suggest that in the presence of CV with 0-100% connected vehicle presence. The results confirm
data, the data from inductive loops does not significantly that significant reductions in delay can be achieved for CV
improve the performance of the algorithm. In terms of deploy- penetrations as low as 10%, highlighting that not all the
ing the MATS algorithm, given an existing system with vehicles in the road network need to be connected to achieve
degraded roadside infrastructure, it would be more beneficial delay reductions. In comparison with MOVA, the MATS
to re-calibrate the fixed-time plan than to restore the loop algorithm reduced mean delay better than MOVA for CV
hardware. Having a reliable fixed-time instance provides the penetrations above 20%, and achieved reductions in the mean
most robust fall-back behaviour for the MATS algorithm, and delay of 20%-28% for CV penetrations above 30%.
that having even a small amount of CV data provides most The findings showed that for networks where the loop
of the benefits of using the proposed adaptive traffic control placement is imperfect, such as in older, degraded urban
system. Therefore, MATS-FT is the best implementation for networks, the performance of the MATS algorithm is hindered
use in existing but degraded road networks. at low CV penetrations when attempting to use data from these
3) Hypothesis Testing: As the simulations were stochastic, loops. Further work should address the issues of data from
hypothesis tests were performed on the delay and stop data in partial or degraded roadside infrastructure in order to improve
order to assess its statistical independence across the 50 exper- how the MATS algorithm generalises to other road networks.
iment runs, and incremental increases in CV penetration. Here, In conclusion, this study has found that the MATS algorithm
the null hypothesis H0 was that the mean stops and delay data addresses the identified issue of enhancing existing traffic
at CV penetrations greater than 0% were drawn from the same signal control systems in urban environments with increasing
normal distribution as the mean delay for 0% CV penetration. numbers of CVs, and operates robustly in realistic scenar-
The alternative hypotheses H1 tested were that the mean ios provided under a comprehensive testing framework. The
delay and stop data for all simulated CV penetrations greater presented testing framework improves on those in previous
than 0%. A two-sample independent T-test was performed literature in terms of scale, traffic demand levels, vehicle
between H0 and each H1 , and the p-value was determined. types considered, resolution of CV penetrations tested, and
The hypothesis testing determined that all scenarios reject communication channel error sources considered. In future
the null-hypothesis with p < 0.001. The results indicate that work, the testing framework can be improved by accounting
the addition of connected vehicles into the transport network for both pedestrian and vehicle movements, and through esti-
changes the MATS algorithm such that it meaningfully impacts mating the prevailing CV penetration. Further testing scenarios
the delays and number of stops experienced by road users in could include lane closures due to disabled vehicles, and
all cases where CVs were present. The rejection of the null response to emergency service vehicles. Future trials should
hypothesis also confirms that there was a significant reduction also investigate how the MATS algorithm applies to other
in delay in all cases for CV penetrations as low as 10%, which urban road networks with different demands, and in other
addresses the gap in previous research. countries.
RAFTER et al.: AUGMENTING TRAFFIC SIGNAL CONTROL SYSTEMS FOR URBAN ROAD NETWORKS WITH CV 1739

ACKNOWLEDGMENT [27] K. Wood, C. Bielefeldt, F. Biora, and G. Kruse, “COSMOS-congestion

management strategies and methods in urban sites,” Transp. Res. Lab.,
The authors would like to thank TRL for providing access Crowthorne, U.K., Tech. Rep., 1998.
to and guidance on TRANSYT. [28] M. Smith, “Traffic signal control and route choice: A new assignment
and control model which designs signal timings,” Transp. Res. C, Emerg.
technol., vol. 58, pp. 451–473, Sep. 2015.
R EFERENCES [29] Q. He, K. L. Head, and J. Ding, “PAMSCOD: Platoon-based arterial
[1] “The future economic and environmental costs of gridlock in 2030,” multi-modal signal control with online data,” Transp. Res. C, Emerg.
CEBR, London, U.K., Tech. Rep. 1, 2014. Technol., vol. 20, no. 1, pp. 164–184, 2012.
[2] “Global traffic scorecard,” INRIX, Kirkland, WA, USA, Tech. Rep. 1, [30] T. Wongpiromsarn, T. Uthaicharoenpong, Y. Wang, E. Frazzoli, and
2017. D. Wang, “Distributed traffic signal control for maximum network
[3] B. Waterson and S. Box, “Quantifying the impact of probe vehicle throughput,” in Proc. 15th Int. IEEE Conf. Intell. Transp. Syst.,
localisation data errors on signalised junction control,” IET Intell. Sep. 2012, pp. 588–595.
Transp. Syst., vol. 6, no. 2, p. 197, 2012. [31] P. Koonce, L. Rodegerdts, K. Lee, and S. Quayle, Traffic Signal Timing
[4] D. I. Robertson, “TRANSYT: A traffic network study tool,” Transp. Manual. Washington, DC, USA: Federal Highway Administration, 2008.
Road Res. Lab., Tech. Rep. LR253, 1969. [32] H. England, “Operational metrics manual,” Highways England, Guild-
[5] G. R. Vincent and J. R. Peirce, “MOVA: Traffic responsive, self- ford, U.K., Tech. Rep. 1, 2019.
optimising signal control for isolated intersections,” Transp. Road Res. [33] G. Rossum, “Python reference manual,” Python Softw. Found., Amster-
Lab., Tech. Rep. RR170, 1988. dam, The Netherlands, Tech. Rep. 1, 1995.
[6] M. G. Bell, “Future directions in traffic signal control,” Transp. Res. A, [34] S. Pakin, “The comprehensive LATEX symbol list,” CTAN,
Policy Pract., vol. 26, no. 4, pp. 303–313, Jul. 1992. Tech. Rep. 1, 2015.
[7] P. Diakaki, D. Kotsialos, and Wang, “Review of road traffic control [35] D. Krajzewicz, M. Bonert, and P. Wagner, “The open source traffic
strategies,” Proc. IEEE, vol. 91, no. 12, pp. 2041–2042, Dec. 2003. simulation package SUMO,” RoboCup,to be published.
[8] S. J. Agbolosu-Amison, A. W. Sadek, and W. Eldessouki, “Inclement [36] J. De Dios Ortúzar and L. G. Willumsen, Modelling Transport. Hoboken,
weather and traffic flow at signalized intersections: Case study NJ, USA: Wiley, 2011.
from Northern New England,” Transp. Res. Rec., vol. 1867, no. 1, [37] P. Robillard, “Estimating the OD matrix from observed link volumes,”
pp. 163–171, Jan. 2004. Transp. Res., vol. 9, nos. 2–3, pp. 123–128, 1975.
[9] H. C. Council, “Intelligent transport systems strategy package report,” [38] K. P. Furness, “Time function iteration,” Traffic Eng. Control, vol. 7,
Hertfordshire County Council, Hertford, U.K., Tech. Rep. 1, 2011. no. 7, pp. 458–460, 1965.
[10] P. Hunt, D. Robertson, R. Bretherton, and R. Winton, SCOOT: A Traffic [39] Licensed Vehicles By Body Type, By Region and Per Head of Popula-
Responsive Method Coordinating Signals. Crowthorne, U.K.: TRRL, tion: Great Britain and United Kingdom, document VEH0104, U.K.,
1981. Government Dept. Transp., 2017.
[11] P. R. Lowrie, “Scats, sydney co-ordinated adaptive traffic system: [40] E. W. Dijkstra, “A note on two problems in connexion with graphs,”
A traffic responsive method of controlling urban traffic,” Roads Traffic Numer. Math., vol. 1, no. 1, pp. 269–271, Dec. 1959.
[41] S. Krauß, “Microscopic modeling of traffic flow: Investigation of
Authority New South Wales, Darlinghurst, Austrailia, Tech. Rep. 1,
collision free vehicle dynamics,” Ph.D. dissertation, Mathematisch–
1990.
[12] R. Engineering, “InSync: The traffic Bot,” Rhyhtm Eng., Lenexa, KS, Naturwissenschaftlichen Fakultät der Universität zu Köln, Cologne,
USA, Tech. Rep. 1, 2019. Germany, 1998.
[13] R. Bretherton, “SCOOT urban traffic control system–philosophy and [42] T. Mathew and K. Ravishankar, “Car-following behavior in traffic
evaluation,” in Control, Computers, Communications in Transportation. having mixed vehicle-types,” Transp. Lett., vol. 3, no. 2, pp. 109–122,
Amsterdam, The Netherlands: Elsevier, 1990. Apr. 2011.
[14] A. Stevanovic, “Adaptive traffic control systems: Domestic and foreign [43] “SUMO vehicle type parameter defaults,” Cologne, Germany, DLR,
state of practice a synthesis of highway practice,” NCHRP, Washington, Tech. Rep. 1, 2018.
[44] Traffic Advisory Leaflet 1/06—General Principles of Traffic Control by
DC, USA, Tech. Rep. 403, 2010.
[15] L. Li, D. Wen, and D. Yao, “A survey of traffic control with vehicular Light Signals, London, U.K., Government Dept. Transp., 2006.
[45] J. A. Bonneson and P. T. McCoy, Manual Traffic Detector Design.
communications,” IEEE Trans. Intell. Transp. Syst., vol. 15, no. 1,
Austin, TX, USA: Texas Transportation Institute, Texas A&M Univ.
pp. 425–432, Feb. 2014.
[16] Y. Feng, “Intelligent traffic control in a connected vehicle environment,” System, 2005.
[46] Z. H. Mir and F. Filali, “LTE and IEEE 802.11 p for vehicular
Ph.D. dissertation, Dept. Syst. Ind. Eng., Univ. Arizona, Tucson, AZ,
networking: A performance evaluation,” EURASIP J. Wireless Commun.
USA, 2015.
[17] D. Fajardo, T.-C. Au, S. T. Waller, P. Stone, and D. Yang, “Automated Netw., vol. 2014, p. 89, Dec. 2014.
[47] Intelligent Transport Systems (ITS); Vehicular Communications; Basic
intersection control: Performance of future innovation versus current
Set of Applications; Part 2: Specification of Cooperative Awareness
traffic signal control,” Transp. Res. Rec., vol. 2259, no. 1, pp. 223–232,
Basic Service, V1.3.2, document ETSI EN 302 637-2, ETSI, 2014.
2011. [48] IEEE Standard for Information Technology–Local and Metropolitan
[18] T. Litman, Autonomous Vehicle Implementation Predictions. Victoria,
Area Networks–Specific Requirements—Part 11: Wireless LAN Medium
BC, Canada: Victoria Transport Policy Institute, 2019.
[19] N. J. Goodall, B. L. Smith, and B. Park, “Traffic signal control with Access Control (MAC) and Physical Layer (PHY) Specifications Amend-
connected vehicles,” Transp. Res. Rec., vol. 2381, no. 1, pp. 65–72, ment 6: Wireless Access in Vehicular Environments,’ IEEE Stan-
Jan. 2013. dard 802.11p-2010, 2010.
[20] B. Beak, K. L. Head, and Y. Feng, “Adaptive coordination based on [49] I. C. Msadaa, P. Cataldi, and F. Filali, “A comparative study between
connected vehicle technology,” Transp. Res. Rec., vol. 2619, no. 1, 802.11 p and mobile WiMAX-based V2I communication networks,”
pp. 1–12, Jan. 2017. in Proc. 4th Int. Conf. Next Gener. Mobile Appl., Services Technol.,
[21] S. Ilgin Guler, M. Menendez, and L. Meier, “Using connected vehicle Jul. 2010, pp. 186–191.
technology to improve the efficiency of intersections,” Transp. Res. C, [50] S. Box and B. J. Waterson, “Signal control using vehicle localization
Emerg. Technol., vol. 46, pp. 121–131, Sep. 2014. probe data,” in Proc. 42nd UTSG Conf., 2010, pp. 1–12.
[22] K. Yang, S. I. Guler, and M. Menendez, “Isolated intersection con- [51] J. C. Binning, M. Crabtree, and G. Burtenshaw, “TRANSYT 15 user
trol for various levels of vehicle technology: Conventional, connected, guide,” TRL Softw., Crowthorne, U.K., Tech. Rep. Issue F, 2013.
and automated vehicles,” Transp. Res. C, Emerg. Technol., vol. 72,
Craig B. Rafter (Student Member, IEEE) received
pp. 109–129, Nov. 2016.
the M.E. degree in electronics and computer engi-
[23] C. B. Rafter, B. Anvari, and S. Box, “Traffic responsive intersection
neering from University College Dublin in 2015.
control algorithm using GPS data,” in Proc. IEEE 20th Int. Conf. Intell.
He is currently pursuing the Ph.D. degree with
Transp. Syst. (ITSC), Oct. 2017, pp. 1–6.
[24] C. B. Rafter, B. Anvari, and S. Box, “A hybrid traffic responsive inter- the University of Southampton, sponsored by the
section control algorithm using global positioning system and inductive Transportation Research Laboratory, with a focus
loop data,” in Proc. Transp. Res. Board 97th Annu. Meeting, 2017, on effects of traffic signal control using connected
pp. 1–19. vehicle data on the transport network. He has com-
[25] B. C. Council, “Birmingham and West Midlands real-time traffic data,” pleted the Ph.D. training program in next generation
Birmingham City Council, Birmingham, U.K., Tech. Rep. 1, 2016. computational modeling from the Engineering and
[26] Open Street Maps, OpenStreetMap Foundation, Coldfield, U.K., 2018. Physical Sciences Research Council Centre.
1740 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 21, NO. 4, APRIL 2020

Bani Anvari (Member, IEEE) received the M.Sc. Tom Cherrett is currently a Professor of logis-
degree in advanced architectural studies from UCL tics and transport management at the University of
in 2009, and the Ph.D. degree in mixed traffic Southampton. He lectures on the topics of transport
modeling from Imperial College London in 2014. planning, freight and passenger systems, and con-
She is currently a Lecturer in intelligent mobil- struction management. His research interests include
ity with the Centre for Transport Studies, UCL. core goods distribution and retail logistics optimiza-
Her research interests include traffic and infrastruc- tion within and between our urban areas, smartphone
ture modeling, autonomous vehicles, and vehicle-to- technology use in logistics, and using remote mon-
vehicle/vehicle-to-infrastructure communications in itoring technology with optimization techniques to
order to create intelligent transportation systems. effectively manage waste and recyclable collection
in urban areas.

Simon Box received the Ph.D. degree in automotive

sensor simulation from the University of Cambridge
in 2007. He currently holds a visiting researcher
position at the University of Southampton, and he
is also an Architect for simulation at Aurora Inno-
vation. His research interests include autonomous
vehicle systems simulation, traffic simulation, and
sensor simulation.