sensors

Article
Cyber-Physical System for Smart Traffic Light Control
Siddhesh Deshpande and Sheng-Jen Hsieh *

Engineering Technology and Industrial Distribution Department, Texas A&M University,

College Station, TX 77843, USA; sid_2o@tamu.edu
* Correspondence: hsieh@tamu.edu

Abstract: In recent years, researchers have proposed smart traffic light control systems to improve
traffic flow at intersections, but there is less focus on reducing vehicle and pedestrian delays simulta-
neously. This research proposes a cyber-physical system for smart traffic light control utilizing traffic
detection cameras, machine learning algorithms, and a ladder logic program. The proposed method
employs a dynamic traffic interval technique that categorizes traffic into low, medium, high, and very
high volumes. It adjusts traffic light intervals based on real-time traffic data, including pedestrian
and vehicle information. Machine learning algorithms, including convolutional neural network
(CNN), artificial neural network (ANN), and support vector machine (SVM), are demonstrated to
predict traffic conditions and traffic light timings. To validate the proposed method, the Simulation
of Urban Mobility (SUMO) platform was used to simulate the real-world intersection working. The
simulation result indicates the dynamic traffic interval technique is more efficient and showcases a
12% to 27% reduction in the waiting time of vehicles and a 9% to 23% reduction in the waiting time
of pedestrians at an intersection when compared to the fixed time and semi-dynamic traffic light
control methods.

Keywords: cyber-physical system; machine learning; smart traffic lights

1. Introduction
The Manual on Uniform Traffic Control Devices (MUTCD) [1] states that traffic lights are
Citation: Deshpande, S.; Hsieh, S.-J. an important tool for improving road safety, reducing traffic congestion, and optimizing
Cyber-Physical System for Smart the use of roadways. Traffic lights are used to manage and control the flow of vehicular and
Traffic Light Control. Sensors 2023, 23, pedestrian traffic at intersections. By alternating the signal between green, yellow, and red,
5028. https://doi.org/10.3390/ traffic lights create a system of order and direction that reduces the likelihood of collisions
s23115028 and accidents. They also allow pedestrians to safely cross the road by stopping traffic and
Academic Editors: Joaquin Garcia- giving them the right of way.
Alfaro and Pantaleone Nespoli While traffic lights are an effective means of enhancing the safety of people at inter-
sections, they have the unintended consequence of increased travel times for drivers and
Received: 25 April 2023
pedestrians who must halt at red lights. Lv et al. [2] noticed that the likelihood of people
Revised: 16 May 2023
halting and waiting at an intersection increases as the number of road crossings grows. The
Accepted: 22 May 2023
traffic lights are designed to manage conflicting traffic flows, meaning they often prioritize
Published: 24 May 2023
one direction of traffic over others, according to the Signal Timing Manual (STM) [3]. This
strategy can delay vehicles or pedestrians waiting to cross the intersection. For example, if
more cars are on one road than another, the traffic light may give a longer green light to the
Copyright: © 2023 by the authors. road with more traffic, causing delays for the other route. The Highway Capacity Manual
Licensee MDPI, Basel, Switzerland. (HCM) [4] states that the travel times for vehicles and pedestrians increase if there is a lack
This article is an open access article of coordination between different intersections. In some cases, the timing of traffic lights
distributed under the terms and may not be synchronized, resulting in vehicle stop-and-go traffic and longer wait times for
conditions of the Creative Commons pedestrians.
Attribution (CC BY) license (https:// The shortcomings of the simple traffic light control system led researchers to inves-
creativecommons.org/licenses/by/ tigate smart traffic light control systems. The term “Smart Traffic Light Control System”
4.0/).

Sensors 2023, 23, 5028. https://doi.org/10.3390/s23115028 https://www.mdpi.com/journal/sensors

Sensors 2023, 23, 5028 2 of 25

refers to employing sensors and algorithms to control and optimize vehicular traffic. Real-
time traffic data analysis allows the system to fine-tune traffic lights and improve traffic
flow to minimize congestion and maximize safety. The control system may use machine
learning models and optimization techniques to reduce travel time or to find the relation
between various traffic signal parameters. An intelligent traffic light control system is gen-
erally more advanced and complex than traditional ones. Overall, an intelligent traffic light
system aims to manage and control traffic signals as per the traffic volume at intersections.
Intelligent traffic light control systems may also include anomaly detection, passage to
emergency vehicles, connection to public transport systems, and integration with other
smart city technologies.

2. Literature Review
The difficulties of urban traffic and intersections have prompted several proposed
remedies from researchers in recent years. Ghazal et al. [5] proposed a system that uses in-
frared sensors and microcontrollers to control traffic lights at simple two-way intersections
based on vehicle density. Diaz et al. [6] built an intelligent traffic light system with a Rasp-
berry Pi, PIR sensor, and LED traffic light for a simple, single-lane, two-way intersection.
These two approaches consider only single-lane traffic. Additionally, the traffic modeling is
not specified for the performance evaluation part.
Silva et al. [7] proposed intelligent traffic lights for low-traffic conditions. This work
presented a methodology that used a LanPro module installed on the vehicles to decide the
path and avoid the red traffic signals. However, as mentioned in the paper, this method is
only suitable for low traffic scenarios. Alharbi et al. [8] showcased a dynamic traffic light
management system, which increases the green light time if more vehicles are detected but
does not explicitly mention if the green time is reduced if fewer vehicles are detected.
Sen and Head [9] proposed a system that can skip green lights for the approaches/roads
without vehicle flow during a traffic light cycle. This technique is better when the traffic is
less on a particular route, and the traffic volumes do not change very often. Li et al. [10]
proposed a methodology to optimize the fixed traffic light timings on isolated intersections.
They optimized the traffic signal plan to reduce weighted vehicle and pedestrian delays.
However, their system is not smart, which will detect pedestrians and vehicles in their
respective lanes and adjust the traffic signal time as per the detected traffic.
Younis and Moayeri [11] worked on the simulation of dynamic traffic lights with one
method based on sensors and another based on communication between vehicles. This
work shows the results based on the through-vehicle movement but does not consider
turning vehicles at the intersection. Tchuitcheu et al. [12] considered adjusting traffic lights
based on the waiting queue of vehicles at the intersection in incoming and outgoing lanes.
Gandhi et al. [13] also considered a smart traffic light system, which calculates the green
traffic light timing based on vehicles detected by the camera sensor but does not consider
the phases which allow the signaling of green light to non-conflicting movements from
different incoming streets.
Pratama et al. [14] proposed a system to change the green light timings based on
the vehicle density in incoming and outgoing lanes. This method is effective when there
are a bunch of intersections close to each other, affecting the traffic density on the streets.
Chavan et al. [15] presented a sensor network with embedded technology to manage traffic
flows. In this work, they did not clearly mention the traffic light parameters for their
technique’s performance evaluation.
Hirankitti et al. [16] showcased an agent-based intelligent traffic light control model.
In this work, they used rules which will perform actions based on inputs such as current
traffic phase, traffic light timing, the queue length of vehicles, and incoming and outgoing
vehicles at an intersection and outline lane space availability, but did not mention the
sensor type or how frequently and when the data will be collected. Almawgani [17] used a
technique that applies different image processing algorithms for nighttime and daytime
Sensors 2023, 23, 5028 3 of 25

traffic detection. However, the work built a simple prototype to showcase the working of
an image processing system and did not include modeling traffic flows at an intersection.
Li et al. [18] used cameras to capture images or videos of traffic conditions and applied
machine learning algorithms to analyze and interpret the data. Wiering et al. [19] simulated
an intelligent model using reinforcement learning to optimize the traffic light setting.
They evaluated the performance of their technique on a group of intersections by making
decisions based on vehicles at the traffic intersection. This method is more focused on
optimizing traffic light parameters and, thus, there is no mention of the sensors to collect
the data.
Wei et al. [20] also used deep reinforcement learning to optimize traffic light timing.
Their system minimized the delays and improved traffic flow for through-movement
vehicles but did not consider the left-turning vehicles. Liu et al. [21] presented a multi-
agent Q-learning approach to control the traffic lights at the intersection. This method
mentions adapting to pedestrians, but the algorithm does not employ the technique when
pedestrians are detected at the intersection. Linag et al. [22] proposed a method that uses
deep reinforcement learning with multiple optimization elements. Their system improved
the vehicle waiting time by 20%. Their system inputs are vehicle position and vehicle speed
but do not include when and where the data points are captured and what should be the
frequency of data capturing in real-time. Göttlich et al. [23] utilized linear programming
optimization to determine the optimal traffic light timing for a given set of traffic conditions.
Park et al. [24] proposed genetic algorithms to optimize specific traffic signal parameters
only for high-traffic cases at intersections.
Genders and Razavi [25] proposed a technique using reinforcement learning and
a convolutional neural network to optimize the traffic light timings for an intersection.
The inputs to the neural network include the presence of the vehicles in a specified area,
vehicle speed, and the current phase of traffic lights. Oliveira et al. [26] used multiple
neural networks to optimize the traffic light timings. This method gives the same inputs to
different ANN, which predicts the same output. This work focuses more on neural network
performance and does not specify parameters such as traffic flows or traffic light cycles for
an intersection. Abbas et al. [27] presented a high-accuracy controller that can change the
next phase timing and net phase green light time based on the current phase data collected
from the roadside data collection (RSDC) units but did not mention the exact type of the
sensors and the frequency of data collection. Additionally, the system evaluation does not
mention the details of traffic light phases.
McKenney and White [28] applied an approach to control traffic signal lights based on
the number of vehicles in a section of a road near the intersections. The approach considers
various parameters related to data collection to decide the traffic light switching and green
light timing change. Zhou et al. [29] used a wireless sensor network to detect the traffic at
multiple intersections. This work used a technique that adjusted the traffic light phases and
timing based on detected vehicles’ information. Piris et al. [30] also proposed a technique
to optimally place the wireless sensors in a network of intersections. Additionally, this
work presented a traffic light management system that focused on controlling multiple
intersections and efficient communication of messages in a sensor network.
Navarro-Espinoza et al. [31] applied various machine learning models to predict the
traffic flow for intelligent traffic lights. Muntean [32] proposed a multi-agent system to
estimate the traffic volumes at junctions and car parking. Neither approach presented
any control strategy for the traffic lights. Artega et al. [33] proposed a fuzzy logic method
to control traffic lights based on the flow rate of vehicles. Nimac et al. [34] presented
pedestrian detection and a traffic light control scheme using radars, but this technique is
focused on the traffic light trigger mechanism and is not optimizing the pedestrian traffic
lights.
Table 1 summarizes the details of the methodologies discussed in the literature review.
From the table, it is clear that many researchers tend to give weightage to one factor: solving
the congestion of vehicles and optimizing the traffic light systems using various techniques
Sensors 2023, 23, 5028 4 of 25

and tools. These methodologies generally focus more on the vehicle parameters that impact
vehicle travel time. They try to reduce the delays at the intersection using sensors and
algorithms. Moreover, many approaches involve modeling the traffic flow through an
intersection and then finding the timing that minimizes some measure of congestion, such
as travel time or delay. They apply machine learning and optimization techniques to reduce
vehicle delay, but reducing pedestrian delay is not considered.

Table 1. Comparison of various intelligent traffic light systems.

Controller Simulation Machine

Sensor(s) Type Results Categorization
Previous Work Type Platform Learning
Mentioned Presented of Traffic
Mentioned Mentioned Used
Yes: high
PIC 16F877A
Ghazal et al. [5] IR sensors No Yes No density and low
microcontroller
density
Diaz et al. [6] PIR sensors Raspberry Pi No No No No
Cameras,
Silva et al. [7] No No No Yes No
radars
WSN, RFID
Alharbi et al. [8] No MATLAB No Yes No
tags
Yes: light traffic
Younis and Piezoelectric Raspberry Pi 3
C++ No Yes and heavy
Moayeri [11] (proposed) (proposed)
traffic
Tchuitcheu et al. [12] WSN No SUMO No Yes No
Python,
Gandhi et al. [13] Cameras No Yes Yes No
Pygames
PIC 16F876A
Pratama et al. [14] IR sensors N/A No Yes No
microcontroller
Microcontroller
AT 89C51
Chavan et al. [15] IR sensors assembly No Yes No
microcontroller
language
NetLogo
Hirankitti et al. [16] No No No Yes No
simulator
Almawgani [17] Camera Arduino No No Yes No
UAV/
Li et al. [18] No SUMO Yes Yes No
cameras
Green Light
Wiering et al. [19] No No District Yes Yes No
simulator
Wei et al. [20] Camera No C++, MATLAB Yes Yes No
Liu et al. [21] Cameras No SUMO Yes Yes No
Vehicular/
Linag et al. [22] Sensor No SUMO Yes Yes No
networks
Genders and
No No SUMO Yes Yes No
Razavi [25]
Yes: small,
Oliveira et al. [26] No No SUMO Yes Yes medium, and
large vehicles
Sensors 2023, 23, 5028 5 of 25

Table 1. Cont.

Controller Simulation Machine

Sensor(s) Type Results Categorization
Previous Work Type Platform Learning
Mentioned Presented of Traffic
Mentioned Mentioned Used
SIRDA
intersection
Abbas et al. [27] No No No Yes No
simulator,
MATLAB
McKenney and
No No SUMO No Yes No
White [28]
Zhou et al. [29] WSN No No No Yes No
Piris et al. [30] WSN No SUMO No Yes No
Navarro-Espinoza et al.
No No No Yes Yes No
[31]
Muntean [32] WSN No No Yes Yes No
Yes: very low,
On-road low, medium,
Artega et al. [33] No SUMO No Yes
sensors high, and very
high
Nimac et al. [34] Radar No No No No No

Thus, this paper’s objective is to propose an intelligent traffic light system and dynamic
traffic interval technique that considers both vehicle and pedestrian traffic volumes to
minimize waiting time at intersections. The next sections of this paper will describe the
proposed system’s design, implementation, and evaluation.

3. Methodology
The traffic lights control the vehicle and pedestrian flow at intersections where traffic
travels in different directions. The traffic flows moving in various directions are called
movements and are categorized as left-turn, through, right-turn, and pedestrian move-
ments, as shown in Figure 1. To differentiate the traffic flows, each movement is identified
by a separate number (HCM) [4]. The traffic signal phases are then used to group certain
movements, which allows the traffic to move in an orderly manner.
Figure 2 depicts a ring and barrier diagram for an 8-phase intersection. In this diagram,
rings organize phases so that they are synchronized without interference and consist of a
series of competing stages. At barriers, phases in both rings conclude concurrently. This
enables dual or two-ring operations, allowing compatible phases to run concurrently with
those of the opposite ring. The phases are an important part of the system as they control the
timing and sequence of the green, yellow, and red lights for the incoming traffic movements
at the intersection. Typically, they distinguish between major and minor street segments.
Additionally, the movements are categorized as permitted and protected movements.
 The traffic lights control the vehicle and pedestrian flow at intersections where traf-
fic travels in different directions. The traffic flows moving in various directions are
called movements and are categorized as left-turn, through, right-turn, and pedestrian
movements, as shown in Figure 1. To differentiate the traffic flows, each movement is
Sensors 2023, 23, 5028 identified by a separate number (HCM) [4]. The traffic signal phases are then used 6 ofto
25
group certain movements, which allows the traffic to move in an orderly manner.

Sensors 2023, 23, x FOR PEER REVIEW 6 of 26

Figure
Figure1.1.Traffic
Trafficmovement
movementat
atan
anintersection.
intersection.

Figure 2 depicts a ring and barrier diagram for an 8-phase intersection. In this dia-
gram, rings organize phases so that they are synchronized without interference and con-
sist of a series of competing stages. At barriers, phases in both rings conclude concur-
rently. This enables dual or two-ring operations, allowing compatible phases to run con-
currently with those of the opposite ring. The phases are an important part of the system
as they control the timing and sequence of the green, yellow, and red lights for the in-
coming traffic movements at the intersection. Typically, they distinguish between major
and minor street segments. Additionally, the movements are categorized as permitted
and protected movements.

Figure 2. Ring and barrier diagram.

3.1. Dynamic Traffic Intervals

3.1. Dynamic Traffic Intervals
As the number of vehicles or pedestrians waiting at the intersection increases or
As the number of vehicles or pedestrians waiting at the intersection increases or de-
decreases, the time required for them to pass through the intersection changes. Thus,
creases, the time required for them to pass through the intersection changes. Thus, the
the traffic light timings need to be appropriately planned so they do not cause delays
traffic light timings need to be appropriately planned so they do not cause delays at the
at the intersection. To effectively manage the traffic flow at the intersection, this work
intersection. To effectively manage the traffic flow at the intersection, this work proposes
proposes a dynamic traffic interval technique where the vehicle and pedestrian volumes
a dynamic traffic interval technique where the vehicle and pedestrian volumes waiting
waiting at a red light at the intersection are divided into low, medium, high, and very
at a red light at the intersection are divided into low, medium, high, and very high cate-
high categories. This is done to efficiently use the traffic light timings for all the phases
gories. This is done to efficiently use the traffic light timings for all the phases and
and movements. Dividing the traffic into intervals allows setting the time of the green
movements. Dividing the traffic into intervals allows setting the time of the green light
light only for the vehicles and pedestrians that request to cross the intersection. This work
only for the vehicles and pedestrians that request to cross the intersection. This work se-
lects the four categories of traffic based on the simulation parameters and results. Each
category represents a group of vehicles waiting at an intersection at a red light. The
number of vehicles can be different for each category for different intersections and de-
pends on the demand for the service requested by the user. To establish the categories of
Sensors 2023, 23, 5028 7 of 25

selects the four categories of traffic based on the simulation parameters and results. Each
category represents a group of vehicles waiting at an intersection at a red light. The number
of vehicles can be different for each category for different intersections and depends on
the demand for the service requested by the user. To establish the categories of traffic, it
is important to perform a critical movement analysis and define the detection zone at an
intersection, as explained further in this section.
Figure 3 shows dynamic traffic light timing distribution for an 8-phase intersection.
Each phase has green light timing for four categories of traffic. The timing of the green light
can be adjusted as per the traffic category detected. Based on the green light timing of an
ongoing phase, the red-light timing for other phases and the total cycle length will change.
The cycle length is the time it takes for a traffic signal to complete a full cycle of all its signal
phases. The yellow clearance time alerts drivers that the right-of-way assignment at the
intersection is going to change. The red clearance time allows vehicles that entered the
intersection during the yellow change period to reach a safe position before the next phase
begins (STM) [3]. HCM [4] has suggested the red and yellow clearance interval timing
for different vehicles approaching speeds. The timing for both lights is referred to as the
Sensors 2023, 23, x FOR PEER REVIEW
change period. This change period is fixed and is based on the average approaching vehicle 7 of 26
speed at the intersection.

Figure 3. Traffic light timing distribution.

Figure 3. Traffic light timing distribution.

Critical movement
Critical movement analysis
analysisisisperformed
performedtotocalculate
calculatethe
thetraffic
traffic light
light timings.
timings. Criti-
Critical
cal movement analysis is a methodology used to identify the movements
movement analysis is a methodology used to identify the movements of vehicles that have of vehicles that
have the highest potential for conflict at an intersection. Critical movement
the highest potential for conflict at an intersection. Critical movement analysis aims to analysis aims
to evaluate
evaluate phasing
phasing requirements
requirements and and
signalsignal
timingtiming parameters.
parameters. This analysis
This analysis uses con-
uses conflicting
flicting often
phases, phases, often a phase
a left-turn left-turn
andphase and anthrough-movement
an opposing opposing through-movement phase,the
phase, to identify to
identifyphase
crucial the crucial phase
pairings. pairings. To
To determine thedetermine the critical
critical phase phase
pair, the total pair,
volumethe of
total
carsvolume
for all
of cars
sets for all setsphases
of conflicting of conflicting phasesFigure
is compared. is compared.
4 shows Figure 4 shows
the vehicle the vehicle
volumes volumes
associated with
associated with each phase. The following equations determine the critical
each phase. The following equations determine the critical volume for an intersection. volume for an
intersection.
CPP1 = max(v1 + v2 , v5 + v6 ) (1)
𝐶𝑃𝑃 max 𝑣 𝑣 ,𝑣 𝑣 (1)

𝐶𝑃𝑃
CPP max
2 = max (v3𝑣+ v4𝑣, v, 𝑣7 + v𝑣8 ) (2)
(2)

𝐶𝑉 𝐶𝑃𝑃 𝐶𝑃𝑃 (3)

where
𝐶𝑃𝑃 : Critical phase pair for street 1.
Sensors 2023, 23, 5028 8 of 25

Sensors 2023, 23, x FOR PEER REVIEW CV = CPP1 + CPP2 8 of(3)

26

where
CPP1 : Critical phase pair for street 1.
𝐶:
CPPOptimum, minimum delay cycle length (seconds).
2 : Critical phase pair for street 2.
𝐿: Lost time per
v1−8 : Critical lanecycle (seconds).
volume for phases 1 to 8 (vehicles/h).
𝑌:
CV: Critical volume for anvolumes
Sum of the critical lane divided by saturation flow rate.
intersection.

Figure
Figure 4.
4. Vehicle
Vehicle volume
volume for
for different
different phases.

This critical
Then, volume
the timing of isthe
then divided
green lightinto
for four
eachvolumes
phase iswhere the traffic
estimated usingisequations
categorized5
as low,
and medium,
6 [36] for fourhigh,
trafficand very high
conditions. traffic.
This Thus,
time is using the
determined critical
based volumes
on the forcycle
expected each
category,
length andthe four movement
critical different cycle lengths are calculated by the following equation based
analysis.
on Webster’s Least delay cycle [35]. The cycle length will be fixed for each of the four
categories as it is based on the critical 𝐴volume 𝐶 for∑𝐶𝑃
an intersection. This equation considers (5)
the saturation flow rate of 1900 passenger cars per hour per lane as stated by HCM [4].
𝑉
𝐺 𝐴 (6)
𝑉 +5
1.5L
C= (4)
where 1.0 − Y
𝐴 : Available time to apportion between all phases’ green interval (seconds).
where
𝐶:
C: Calculated
Optimum, cycle length
minimum (seconds).
delay cycle length (s).
𝐶𝑃 : Change period (yellow
L: Lost time per cycle (s). change interval plus red clearance interval) for each phase
Y: Sum(seconds).
of the critical lane volumes divided by saturation flow rate.
𝐺 : Phase green interval for each phase (seconds).
Then, the timing of the green light for each phase is estimated using Equations (5)
𝑉 : Critical lane volume for phase i (vehicles/h).
and (6) [36] for four traffic conditions. This time is determined based on the expected cycle
𝑉length
: Sumandof critical
critical lane volumes
movement for all phases (vehicles/h).
analysis.
The green light timing for each traffic category
is based on the sensor detection
zone. The detection zone is the areaAwhere t =C− the∑ CPi collects the data. For this research,
sensor (5)
the cameras are considered as sensors installed at the intersection. Cameras provide da-
ta-rich information, and recent advancements VA in machine learning models prove that vi-
G
sion systems are effective in detecting traffic. i = × Awork
This t assumes that cameras will have (6)
VT
a clear view of the lanes and will detect the traffic at night when the lights are installed
where
at the intersection.
At : Available time tozone
The occupancy apportion between
is divided into all
fourphases’ greenasinterval
segments, shown(s).
in Figures 5 and 6 for
C: Calculated
vehicles cycle lengthrespectively,
and pedestrians, (s). according to the critical volumes for each category
CPtraffic.
of i : Change
Eachperiod (yellow
segment change the
represents interval
number plusofred clearance interval) that
vehicles/pedestrians for each
will phase (s).
be wait-
G
ing : Phase green interval for each phase (s).
i at the red light with respect to the critical volume. The traffic will occupy a certain
area of the zone and the number of vehicles and pedestrians will be calculated based on
these four segments. For the pedestrians, the cameras will take the image of curbside
where pedestrians wait on a red signal. The traffic category and its respective occupancy
Sensors 2023, 23, 5028 9 of 25

VA : Critical lane volume for phase i (vehicles/h).

VT : Sum of critical lane volumes for all phases (vehicles/h).
The green light timing for each traffic category is based on the sensor detection zone.
The detection zone is the area where the sensor collects the data. For this research, the
cameras are considered as sensors installed at the intersection. Cameras provide data-
rich information, and recent advancements in machine learning models prove that vision
systems are effective in detecting traffic. This work assumes that cameras will have a clear
view of the lanes and will detect the traffic at night when the lights are installed at the
intersection.
The occupancy zone is divided into four segments, as shown in Figures 5 and 6 for
vehicles and pedestrians, respectively, according to the critical volumes for each category
of traffic. Each segment represents the number of vehicles/pedestrians that will be waiting
at the red light with respect to the critical volume. The traffic will occupy a certain
area of the zone and the number of vehicles and pedestrians will be calculated based
on these four segments. For the pedestrians, the cameras will take the image of curbside
where pedestrians wait on a red signal. The traffic category and its respective occupancy
percentages are given in Table 2. The 100% occupancy is the area under the detection where
the camera sensor can clearly see vehicles and pedestrians.

Table 2. Traffic categories and corresponding occupancy.

Sensors 2023, 23, x FOR PEER REVIEW Traffic Category Occupancy 9 of 26

Low traffic 0–25%
Medium traffic 26–50%
High traffic 51–75%
percentages areVery
given
highintraffic
Table 2. The 100% occupancy is the area under the detection
76–100%
where the camera sensor can clearly see vehicles and pedestrians.

Figure 5. Vehicle occupancy zone.

Figure 5. Vehicle occupancy zone.
Sensors 2023, 23, 5028 10 of 25

Figure 5. Vehicle occupancy zone.

Figure6.6.Pedestrian
Figure Pedestrianoccupancy
occupancyzone.
zone.

Sensors 2023, 23, x FOR PEER REVIEW

Table
3.2. 2. Traffic categories and corresponding occupancy.
Framework 10 of 26
Figure 7 depicts an overview
Traffic Categoryof the framework. The proposed framework
Occupancy is organized
into three systems: physical, cyber,
Low traffic and control. The physical system consists of a junction
0–25%
where two or more Medium
streets cross, vehicles and pedestrians who wish to cross the junction,
junction, sensors that collecttraffic
data, traffic lights that allow or prevent26–50%
vehicles and pedes-
sensors that collect data, traffic lights that allow or prevent vehicles and pedestrians from
trians from crossing, and Highatraffic
programmable logic controller that controls
51–75%the traffic lights.
crossing, and a programmable logic controller that controls the traffic lights. The cyber
The cyber system employs
Very high machine
traffic learning (ML) algorithms to 76–100%
predict traffic condi-
system employs machine learning (ML) algorithms to predict traffic conditions and light
tions and light timings, reducing delays for pedestrians and automobiles. The control
timings, reducing delays for pedestrians and automobiles. The control system, which
3.2. Framework
system, which includes a ladder logic program, influences the physical system by alter-
includes a ladder logic program, influences the physical system by altering the on and off
ing theFigure
on and7 off timesan
depicts of the traffic lights.
times of the traffic lights. overview of the framework. The proposed framework is orga-
nized into three systems: physical, cyber, and control. The physical system consists of a
junction where two or more streets cross, vehicles and pedestrians who wish to cross the

Figure 7.
Figure Overview of
7. Overview of Cyber-Physical
Cyber-Physical System.
System.

The operation
The operationofofthe
thesmart
smarttraffic
traffic light
light control
control system
system is given
is given in Figure
in Figure 8. pro-
8. The The
proposed system’s role is to detect traffic conditions and regulate intersection traffic
posed system’s role is to detect traffic conditions and regulate intersection traffic lights. lights.
The sensors
The sensors are
are cameras
cameras that
that supply
supply the
the information
information required
required by
by the
the cyber
cyber system.
system. TheThe
CNN model
CNN model determines
determines thethe traffic
traffic category, while the
category, while the ANN/SVM predicts green
ANN/SVM predicts green light
light
timings based on traffic photos, vehicle count, and pedestrian count. After
timings based on traffic photos, vehicle count, and pedestrian count. After that, the an- that, the
anticipated values are saved in the DDE client file, which communicates
ticipated values are saved in the DDE client file, which communicates with the DDE with the DDE
server and ladder logic. To eliminate delays, the smart traffic light control system can adjust
server and ladder logic. To eliminate delays, the smart traffic light control system can ad-
just the timing of the green and red lights based on predictions. MATLAB was used to
create machine learning models. Visual Basic for Applications (VBA) was used to convey
data to ladder logic, and the control system runs on the RSLinx Classic and RSLogix
platforms.
Sensors 2023, 23, 5028 11 of 25

the
Sensors 2023, 23, x FOR PEER REVIEW timing of the green and red lights based on predictions. MATLAB was used to11 create
of 26
machine learning models. Visual Basic for Applications (VBA) was used to convey data to
ladder logic, and the control system runs on the RSLinx Classic and RSLogix platforms.

Figure
Figure 8.
8. Operation
Operationof
ofsmart
smart traffic
traffic light
light control
control system.
system.

3.2.1. Physical
3.2.1. PhysicalSystem
System
This work
This workfocuses
focuseson onanan intersection
intersectionto toreduce
reducevehicle
vehicleand andpedestrian
pedestriantraveltraveldelays.
delays.
The traffic intersection is a central part of the physical system
The traffic intersection is a central part of the physical system and framework. Traffic and framework. Traffic
in-
tersections can be regulated by traffic lights, stop signs, yield signs, or other traffictraffic
intersections can be regulated by traffic lights, stop signs, yield signs, or other con-
control
trol devices
devices to ensure
to ensure safesafeandand efficient
efficient trafficflow.
traffic flow.They
Theyare arecritical
criticalcomponents
components of of
transportation systems in urban areas and are designed to manage traffic movements and
transportation systems in urban areas and are designed to manage traffic movements
minimize the risk of collisions and accidents.
and minimize the risk of collisions and accidents.
The collector roads are common in cities, and they connect local streets to arterial roads
The collector roads are common in cities, and they connect local streets to arterial
and are designed to provide access to local destinations, such as homes and businesses. This
roads and are designed to provide access to local destinations, such as homes and busi-
research considers a four-way junction on collector roads, where many users can approach
nesses. This research considers a four-way junction on collector roads, where many users
the intersection at once. The junction has two streets crossing each other. Each street can
can approach the intersection at once. The junction has two streets crossing each other.
have one or more incoming and outgoing vehicle lanes and two lanes for pedestrians
Each street can have one or more incoming and outgoing vehicle lanes and two lanes for
on either side. The pedestrians can cross the junction using marked crosswalks at the
pedestrians on either side. The pedestrians can cross the junction using marked cross-
intersection.
walks at the intersection.
The traffic lights are usually installed at the intersection on poles or mast arms over the
The traffic lights are usually installed at the intersection on poles or mast arms over
roadway. The lights face each lane, indicating when vehicles and pedestrians are allowed to
the
moveroadway.
and when Thetheylights face
must each
stop. Thelane, indicating
location when vehicles
and placement andlights
of traffic pedestrians are al-
at intersections
lowed to move and when they must stop. The location and placement
are carefully planned to convey the information and right-of-way to the users clearly. The of traffic lights at
intersections are carefully planned to convey the information
MUTCD [1] recommends the traffic light types and installation procedures based on the and right-of-way to the us-
ers clearly. The
intersection MUTCD [1]This
characteristics. recommends
work considers the traffic light yellow,
the green, types and andinstallation
red lights for proce-
each
dures based on the intersection characteristics. This work considers
lane. Additionally, a flashing yellow light is considered for the permitted movements of the green, yellow,
and red lights for each lane. Additionally, a flashing yellow light is considered for the
the vehicle.
permitted movements
A traffic detectionofcamera
the vehicle.
is the most common type of sensor utilized at intersections
to detect traffic. The cameras areisusually
A traffic detection camera the most common
installed type
on the of sensor
traffic utilized
poles with at intersec-
a certain height
tions
to get a specific view of the lanes. The camera specification is chosen based on theademand
to detect traffic. The cameras are usually installed on the traffic poles with certain
height to get
for traffic a specific
signal serviceview of the
and local lanes. Theguidelines
jurisdiction camera specification
TCM [4]. Inisthis chosen
work,based on the
the cameras
demand for traffictosignal
are programmed captureservice
images and oflocal jurisdiction
the lanes when the guidelines TCM
traffic light [4]. Infrom
changes this work,
red to
the cameras are programmed to capture images of the lanes
green. Then, the captured photos of the intersection are sent to computer vision algorithms when the traffic light
changes from red to green.
to estimate traffic patterns and volume. Then, the captured photos of the intersection are sent to
computer vision algorithms to estimate traffic patterns and volume.
Traffic light controllers are electronic devices that control the operation of traffic
lights at intersections. These are the cabinets that can be installed in a central control
Sensors 2023, 23, 5028 12 of 25

Sensors 2023, 23, x FOR PEER REVIEW 12 of 26

Traffic light controllers are electronic devices that control the operation of traffic lights
at intersections. These are the cabinets that can be installed in a central control room,
room, allowing
allowing operators operators
to monitor to monitor andsignal
and adjust adjusttiming
signal across
timing multiple
across multiple intersec-
intersections in
tions
real in real
time. time. Alternatively,
Alternatively, controllerscontrollers can be distributed
can be distributed across individual
across individual intersec-
intersections and
tions and communicate
communicate with each other withtoeach other tosignal
coordinate coordinate
timingsignal timing and
and optimize optimize
traffic traffic
flow across a
flow across
wider a wider
area. The area. The
controllers are controllers
programmed areto
programmed to control
control the on/off theofon/off
times all thetimes of
traffic
all the
lights at traffic lights at the
the intersection. intersection.
Sensors installedSensors
at theinstalled at the
intersection intersectioncommunicate
continuously continuously
communicate real-time information to the controllers. The controller
real-time information to the controllers. The controller then sends the data to the then sends thecyber
data
to the cyber system for
system for further processing. further processing.

3.2.2.
3.2.2.Cyber
CyberSystem
System
The
The cybersystem
cyber systemconsists
consistsofofmachine
machinelearning
learningmodels
modelsto toprocess
processthe thesensor
sensordata.
data. In
In
this work, the four traffic categories are predicted using different kinds
this work, the four traffic categories are predicted using different kinds of pre-trained of pre-trained CNN
models. In CNNInmodels,
CNN models. the images
CNN models, are represented
the images as numeric
are represented matrices
as numeric where awhere
matrices singlea
image
singleisimage
divided into pixels,
is divided intoand eachand
pixels, pixel is assigned
each a numerica value.
pixel is assigned numeric These matrices
value. These
are processed
matrices by CNN models
are processed by CNN using filtersusing
models or kernels, which
filters or havewhich
kernels, been trained
have been to identify
trained
features. Thisfeatures.
to identify process is known
This as convolution.
process is known asThe convolutional
convolution. The layer applies numerous
convolutional layer ap-
convolution filters to the image. Then, the pooling, or subsampling layer
plies numerous convolution filters to the image. Then, the pooling, or subsampling layer acts to decrease
the spatial
acts dimensions
to decrease of thedimensions
the spatial data, which of helps to reduce
the data, which computational
helps to reduce needs and control
computational
overfitting. Finally, the fully connected layer, functioning like a traditional
needs and control overfitting. Finally, the fully connected layer, functioning like a tradi- multi-layer
perceptron, usually perceptron,
tional multi-layer employs a SoftMax
usually activation
employs afunction
SoftMaxtoactivation
classify the input image
function into
to classify
different
the input categories
image into based on the
different training dataset.
categories based on the training dataset.
The
TheCNN
CNNmodels
modelsare areprovided
provided with simulation
with simulation images
imagesto detect the the
to detect traffic condition,
traffic condi-
astion,
shown in Figure 9. Here, the images of the vehicles and pedestrians
as shown in Figure 9. Here, the images of the vehicles and pedestrians are given are given to theto
algorithms separately. The output of the models is the probability for
the algorithms separately. The output of the models is the probability for the low, medi- the low, medium,
high,
um, and
high,very
andhigh
verycategories.
high categories.

Figure9.9.CNN
Figure CNNarchitecture.
architecture.

Thiswork
This workuses
usesaafeed-forward
feed-forwardartificial
artificialneural
neuralnetwork
network(ANN)
(ANN) and and support
support vector
vector
machine(SVM)
machine (SVM)model
modeltotopredict
predictthethe green
green light
light time
time forfor a through
a through movement
movement of vehi-
of vehicles
atcles at an intersection
an intersection by using bythe
using the data
sensor sensor data asIt input.
as input. It is assumed
is assumed that the
that the sensor sensor
provides
provides
the numberthe of number
vehicles of
andvehicles and pedestrians
pedestrians as an input as an input
to the to the algorithms.
algorithms. The out-
The output from the
put from the models is the green light timings for the ongoing phase.
models is the green light timings for the ongoing phase. For the ANN model, one hidden For the ANN mod-
el, one
layer hidden
with layer with
five neurons five
was neuronsaswas
selected, selected,
shown as shown
in Figure 10. Thein Figure 10. The
following follow-
equations
ing equations
describe the ANN describe
modelthe ANN model parameters.
parameters.
𝑢 2∑
𝑤 𝑥 (7)
uk = ∑ wkj x j (7)
𝑦 j=φ1 𝑢 𝑏 (8)
where
y k = ϕ ( u k + bk ) (8)
𝑢 : linear combiner output;
𝑤 : weights of neuron k;
𝑥 : inputs to the neural network;
 Sensors 2023, 23, 5028 13 of 25

where
Sensors 2023, 23, x FOR PEER REVIEW 13 of 26
uk : linear combiner output;
wkj : weights of neuron k;
x j : inputs to the neural network;
𝑦yk::output of kth
output of kthneuron;
neuron;
Φ:
ϕ:activation function;
activation function;
𝑏 : bias.
b k : bias.

Figure
Figure 10.
10. Feed-forward ANNarchitecture.
Feed-forward ANN architecture.

The SVM
The SVM model modelmaps
mapsthe
theinput data
input intointo
data highhigh
dimensional spacespace
dimensional via mapping the
via mapping the
function f ( x ). The following is the equation for the SVM model.
function 𝑓 𝑥 . The following is the equation for the SVM model.
f ( x ) = hw, x i + b (9)
𝑓 𝑥 𝑤, 𝑥 𝑏 (9)
where
where
𝑓f (𝑥x )::mapping
mapping function;
function;
x: input vector;
𝑥: input vector;
w: weight vector perpendicular to the higher plane;
𝑤: weight vector perpendicular to the higher plane;
b: bias term.
𝑏: bias term.
The predicted data are then stored in a DDE client. Dynamic data exchange (DDE) is
The predicted
a protocol data
that allows are then stored
communication in a applications.
between DDE client. Here,
Dynamic dataclient
the DDE exchange
is the (DDE)
isExcel
a protocol
file where the data are stored. In this work, to initiate the conversation with the DDEclient is
that allows communication between applications. Here, the DDE
server
the andfile
Excel thewhere
ladder the
logic, a VBA
data arecode wasIn
stored. used.
thisThe client
work, tothen connects
initiate with the control
the conversation with the
system
DDE and sends
server theladder
and the input data to the
logic, DDEcode
a VBA server to adjust
was used. the
Thetraffic
clientlight
thentimings.
connects with the
control system and sends the input data to the DDE server to adjust the traffic light tim-
3.2.3. Control System
ings.
The control system consists of a DDE server that receives the data from the cyber
system. The process begins with the Excel sheet requesting data to be entered in the ladder
3.2.3.
logic.Control System
The server receives the request and retrieves the specific instruction from its memory
whereThe
thecontrol
data needsystem consists The
to be updated. of acontroller
DDE server that the
then uses receives
data inthe
the data
ladderfrom
logicthe
to cyber
change the
system. Thetraffic light
process timing.
begins Whenever
with thesheet
the Excel data in the Excel file
requesting change,
data to be an update
entered inisthe lad-
der logic. The server receives the request and retrieves the specific instruction from its
memory where the data need to be updated. The controller then uses the data in the
ladder logic to change the traffic light timing. Whenever the data in the Excel file change,
an update is sent to the ladder logic, ensuring that the controller sending the signal to
the traffic light is always up to date.
 Sensors 2023, 23, 5028 14 of 25

sent to the ladder logic, ensuring that the controller sending the signal to the traffic light is
always up to date.
Sensors 2023, 23, x FOR PEER REVIEW 14 of 26
A ladder logic program is utilized to implement the dynamic traffic interval strategy.
It is a programming language used in programmable logic controllers (PLCs) to create
control systems for industrial processes and machines. The language is based on ladder
The ladder
diagrams, logic represents
which represent the control
the physical componentssystem as control
of the a seriessystem,
of rungs
suchonasatraffic
ladder. Each
rung consists of one or more input
lights, and the logic that connects them. conditions that must be met for the output to be acti-
vated.TheThe input
ladder conditions
logic representsare
the connected
control system by as
logical operators
a series such
of rungs on as AND,
a ladder. EachOR, and
rung consists
NOT, whichofallow
one orcomplex
more input conditions
logical that must
operations tobebemet for the output
performed. Thetooutput
be activated.
of one rung
The input conditions are connected by logical operators such as AND, OR,
can be connected to the input of another rung, allowing for more complex control sys- and NOT, which
allow to
tems complex logical operations to be performed. The output of one rung can be connected
be created.
to the input of another rung, allowing for more complex control systems to be created.
In this work, the ladder logic is separated into main routines and sub-routines. The
In this work, the ladder logic is separated into main routines and sub-routines. The
main routineisisthe
main routine theprimary
primary program
program thatthat controls
controls the overall
the overall operation
operation of the system.
of the system. It It
consists
consists ofofaaseries
seriesofofrungs
rungs that
that areare executed
executed sequentially,
sequentially, withrung
with each eachcontrolling
rung controlling
a a
specific sub-routine, as shown in Figure 11. The sub-routines represent
specific sub-routine, as shown in Figure 11. The sub-routines represent the timing operation the timing opera-
tion
of theoflow,
themedium,
low, medium,
high, and high,
very and
high very high as
categories, categories, as shown
shown in Figure in Figure 12. Sub-
12. Sub-routines
can be called
routines canfrom withinfrom
be called the main
withinroutine, allowing
the main them to
routine, be used as
allowing building
them to beblocks
usedtoas build-
create more complex control systems.
ing blocks to create more complex control systems.

Figure 11.Ladder
Figure 11. Ladderlogic
logic program—Main
program—Main routine.
routine.
Sensors 2023,
Sensors 2023,23,
23,x5028
FOR PEER REVIEW 15 of 25 15 of 2

Figure12.
Figure 12.Ladder
Ladder logic
logic program—Sub
program—Sub routine.
routine.

4.
4. Simulation
Simulationand
andModelling
Modelling
AAsimulation
simulation model
model waswas
builtbuilt
in the
inSimulation of Urban
the Simulation Mobility
of Urban (SUMO)(SUMO)
Mobility to mimic to mim
the real-world intersection and the dynamic traffic interval strategy. It is a widely used
ic the real-world intersection and the dynamic traffic interval strategy. It is a widely used
open-source microscopic traffic simulation software used for various research purposes [37].
open-source microscopic traffic simulation software used for various research purpose
SUMO is primarily used for modeling and analyzing different traffic light control strategies.
[37]. SUMOuse
Researchers is primarily used fordifferent
SUMO to simulate modeling and analyzing
strategies, different traffic
such as fixed-time light contro
or actuated
strategies. Researchers use SUMO to simulate different strategies,
control, and then compare their performance in various metrics, such as average wait such as fixed-time o
actuated
time or fuelcontrol, and then
consumption. Thiscompare their performance
can help understand the relativein performance
various metrics, such as aver
of different
strategies
age wait and
timeinform
or fuel theconsumption.
design and operation
This canof help
trafficunderstand
light systems.the relative performance o
This research
different strategiesusesand
a common
informfour-way intersection
the design in the SUMO
and operation modellight
of traffic to demonstrate
systems.
the dynamic traffic light logic shown in Figure 13. This intersection
This research uses a common four-way intersection in the SUMO has eight phases, where
model to demon
the left turn movements and the through movements on the same road are divided into
strate the dynamic traffic light logic shown in Figure 13. This intersection has eigh
separate phases. The traffic light phases follow the National Electrical Manufacturers
phases, where the left turn movements and the through movements on the same road
Association (NEMA) convention. Each incoming road is divided into four lanes, and the
are divided
outgoing roadinto separate
is divided intophases. The traffic
three lanes. There islight phasesfor
a sidewalk follow
everythe
roadNational
that onlyElectrica
Manufacturers Association (NEMA) convention. Each incoming
pedestrians can use. The pedestrians use the crosswalks at the intersection during road is divided
the into
four lanes, and thephases.
through-movement outgoing road is divided into three lanes. There is a sidewalk for eve
ry road that only pedestrians can use. The pedestrians use the crosswalks at the intersec
4.1. Parameters
tion during the through-movement phases.
The simulation model defined the traffic flow for each lane to represent real-world
traffic scenarios. The highest critical volume used in the simulations was 1000 passenger
cars per hour per lane. Additionally, the flows were defined to allow vehicles or pedestrians
coming from one direction to travel in all directions. The dimensions of the intersections
were default values generated by SUMO. The length of the edges was 150 m. The passenger
vehicles were considered to have a length of 5 m, and the minimum gap between vehicles
was presumed as 2 m when the vehicles were fully stopped at the intersection. The speed
limit was 25 mph for vehicles with a 10% deviation in speed. The pedestrians were
Sensors 2023, 23, 5028 16 of 25

considered adults with an average speed of 1.06 m/s with a speed deviation of 10%.
Additionally, they would occupy a 0.3 m2 area while waiting at the curbside [38]. The
detection zone length for the through-movement lanes was 61 m. The pedestrian detection
zone was defined before the area where the pedestrian stops, as SUMO has no default option
to place the lane area detectors where the pedestrians stop. For turning-movement lanes
(left and right), the volume of vehicles was considered as 75% of the through movement
lanes. Using the parameters for vehicles and pedestrians, the number of vehicles and
pedestrians waiting at the intersection was calculated for low, medium, high, and very high
categories, as shown in Table 3. The parameters were kept the same for all traffic 16
Sensors 2023, 23, x FOR PEER REVIEW flows
of 26
coming from all directions.

Figure13.
Figure 13.A
A4-way,
4-way, 8-phase
8-phase intersection.
intersection.

4.1. Parameters
Table 3. Parameters for traffic categories.
The simulation model defined the traffic flow for each lane to represent real-world
Traffic Conditions
traffic scenarios. The highest critical volume used in the simulations was 1000 passenger
Parameters
cars per hour per lane. Additionally, the flows Low were definedMedium vehicles orVery
to allowHigh pedes-
High
trians coming from one direction to travel in all directions. The dimensions of the inter-
sectionsCritical
werevolume
default(vehicles/h/lane)
values generated by SUMO. 416 The length718 of the 920 edges was1071150 m.
TheDetection
passenger zonevehicles Vehicle (length) (m) to have16
were considered a length of325 m, and the 48 minimum 61 gap
Pedestrian (area) (m2 ) 1.86 3.72 5.58 7.44
between vehicles was presumed as 2 m when the vehicles were fully stopped at the in-
tersection. The speed limitLeft
No. of vehicles wasvehicles 0–1 with a 2–3
25 mph for vehicles 10% deviation4–5 in speed.
6–7 The
Through vehicle 0–2 3–4
pedestrians were considered adults with an average speed of 1.06 m/s with a speed 5–6 7–8 de-
viation of 10%. No. Additionally,
of pedestrians they would occupy 0–2 a 0.3 m3–4 2 area while 5–6 waiting7–8 at the
curbside [38]. The detectionPedestrian
zone length for the through-movement
4 5 lanes
6 was 61 m.7 The
Green Light timing
pedestrian detection zoneLeft was defined before the
movement 9 area where 12 the pedestrian
15 stops,
19 as
(s)
SUMO has no default option Through tomovement 16–19detectors
place the lane area 16–19
where the20pedestrians 26 stop.
For turning-movement lanes (left and right), the volume of vehicles was considered as
75%Traffic
4.2. of theDistributions
through movement lanes. Using the parameters for vehicles and pedestrians,
the number
To generateof vehicles
a certainand pedestrians
category waiting
of traffic flow at the simulation,
in the intersection this
waswork
calculated
uses the for
low, medium, high, and very high categories, as shown in Table 3. The
Poisson and binomial distribution for a vehicle flow approaching an intersection. Past lineparameters were
kept
of workthe[39–42]
same for all traffic
shows flows
that the coming
Poisson from alldistribution
probability directions. can be used to approximate
the behavior of low traffic flow, and the binomial probability distribution can be used
to estimate the high traffic flow. Table 4 contains Poisson probabilities for the low and
medium traffic flow, whereas the binomial probabilities were used to model the high
Sensors 2023, 23, 5028 17 of 25

and very high traffic flow. Similarly, the Poisson probabilities were used to model the
pedestrians approaching an intersection. These probabilities would generate four different
traffic conditions for vehicles and pedestrians.

Table 4. Probabilities for four traffic conditions.

Traffic Lane Probabilities (Vehicles/s) Pedestrian Probability

Conditions Through Vehicle Left Vehicle (Pedestrians/s)
Low 0.033 0.024 0.012
Medium 0.057 0.042 0.024
High 0.065 0.049 0.035
Very High 0.079 0.059 0.047

4.3. Dynamic Traffic Light Algorithm

This work uses the Simulation of Urban Mobility (SUMO) Traffic Control Interface
(TraCI) module that controls the SUMO model using dynamic traffic logic. It allows the
retrieval of item values and the live modification of their behavior by granting access to a
running road traffic simulation [37]. Algorithm 1 shows the pseudo-code for adjusting the
green traffic light timing based on the detected number of pedestrians and traffic detected
for the through traffic phase.

Algorithm 1 Dynamic Traffic Light adjustment

1 Input: n ← number of through phases
2 i ← current through phase
3 vi ← detected vehicles
4 pi ← detected pedestrians
5 ti ← traffic light timing
6 Output: W1 ← average vehicle delay
7 W2 ← average pedestrian delay
8 Procedure: AdjustTrafficLightTiming
9 pi = GetHaltedPedestrians (Tr → Condition)
10 vi = GetHaltedVehicles (Tr → Condition)
11 for i= (1 to n) do
12 if Tr = low then
13 ti = low
14 elseif Tr = medium then
15 ti = medium
16 elseif Tr = high then
17 ti = high
18 else Tr = very high then
19 ti = very high
20 end if
21 end for
22 Output: W1 , W2

Figure 14 shows the architecture of the dynamic traffic light interval control strategy. At
the beginning of each phase, the sensors input the dynamic traffic logic, and the green light
timing will be altered according to the dynamic logic. This research utilizes a combination
of pedestrian and vehicle green time in the simulation to reduce the waiting time for
automobiles and pedestrians. The TraCI module interacts with the simulation online,
whereas the communication between the TraCI and the cyber system happens in fixed
iterations.
 Figure 14 shows the architecture of the dynamic traffic light interval control strate-
gy. At the beginning of each phase, the sensors input the dynamic traffic logic, and the
green light timing will be altered according to the dynamic logic. This research utilizes a
combination of pedestrian and vehicle green time in the simulation to reduce the waiting
Sensors 2023, 23, 5028 time for automobiles and pedestrians. The TraCI module interacts with the simulation 18 of 25
online, whereas the communication between the TraCI and the cyber system happens in
fixed iterations.

Figure 14. Dynamic traffic light algorithm.

Figure 14. Dynamic traffic light algorithm.
4.4.Design
4.4. DesignofofExperiments
Experiments
To check
To check the theperformance
performance of the the dynamic
dynamiccontrol
controllogic,
logic,it itwas
wascompared
compared with
withthethe
pre-timed and semi-dynamic traffic systems. The simulation parameters
pre-timed and semi-dynamic traffic systems. The simulation parameters were the same for were the same
allfor
thealltraffic
the traffic
light light
control control strategies
strategies exceptexcept the traffic
the traffic lightlight timings.
timings. In pre-timed
In the the pre-timedtraffic
traffic
light light system,
system, the green thelight
green light timings
timings wereirrespective
were fixed fixed irrespective
of the of the traffic
traffic condi-The
condition.
tion. light
green The green
timings lightfortimings for each
each phase phase
were set were set to
to allow allow
the the highest
highest volumevolume
to passtothrough
pass
through
the the intersection
intersection without green without green
light light extension.
extension. So, the So, the pre-timed
pre-timed traffic traffic light sys-
light system used
tem used the timing of the very heavy traffic condition category.
the timing of the very heavy traffic condition category. For the semi-dynamic traffic light For the semi-dynamic
traffic light
system, system,
the traffic thetimings
light traffic light
weretimings
dividedwereintodivided into two rush
two categories: categories: rushnon-rush
hours and hours
and non-rush hours. So, here, the timings of medium and very
hours. So, here, the timings of medium and very heavy traffic conditions categories were heavy traffic conditions
categories were used for the green light during non-rush hours and rush hours, respec-
used for the green light during non-rush hours and rush hours, respectively. The traffic
tively. The traffic demand was built in a way to mimic real-world traffic conditions for 24
demand was built in a way to mimic real-world traffic conditions for 24 h. Table 5 gives a
h. Table 5 gives a detailed breakdown of the vehicle and pedestrian volumes at different
detailed breakdown of the vehicle and pedestrian volumes at different times of the day.
times of the day.
The distribution of the traffic flow was based on the peak hours for urban areas
Table 5. Vehicle and Pedestrian volume for 24 h simulation.
from 6 am to 10 am and 4 pm to 8 pm as stated by FHWA [43], where the vehicle and
pedestrian categories used around this time are medium, Traffichigh, and very high. Addition-
Conditions
ally, toTime of thedemand
establish Day at an intersection, the travel time for Houston was considered
Vehicles Pedestrians
12:00 a.m.–5:00 a.m. Low Low
5:00 a.m.–6:00 a.m. Low Medium
6:00 a.m.–7:30 a.m. Medium Medium
7:30 a.m.–9:00 a.m. Very high High
9:00 a.m.–10:30 a.m. Very high Very high
10:30 a.m.–12:00 p.m. High Medium
12:00 p.m.–1:30 p.m. Medium High
1:30 p.m.–3:00 p.m. Low Very high
3:00 p.m.–4:30 p.m. Low High
4:30 p.m.–6:00 p.m. Medium Very high
6:00 p.m.–6:45 p.m. High High
6:45 p.m.–7:30 p.m. High Very high
7:30 p.m.–8:15 p.m. Very high Medium
8:15 p.m.–9:00 p.m. Very high Low
9:00 p.m.–10:30 p.m. High Low
10:30 p.m.–12:00 a.m. Medium Low

The distribution of the traffic flow was based on the peak hours for urban areas
from 6 am to 10 am and 4 pm to 8 pm as stated by FHWA [43], where the vehicle and
pedestrian categories used around this time are medium, high, and very high. Additionally,
to establish demand at an intersection, the travel time for Houston was considered over a
period of day for the year 2022 [44]. The demand was built to test the control strategy for
different combinations of vehicle and pedestrian volume. This demonstrates the variability
in the volume of vehicles and pedestrians over the course of a day. In the simulations,
the overall waiting time was checked and compared for pre-timed, semi-dynamic, and
dynamic traffic systems.
Sensors 2023, 23, 5028 19 of 25

5. Data Analysis and Findings

Simulations were conducted using the parameters given in Table 6 to check the effect
of pedestrian volume on vehicle waiting time and vehicle volume on pedestrian waiting
time. In total, 160 simulations were conducted, where ten samples were taken for each
combination of pedestrian and vehicle volume. The pedestrian volume was changed, and
the average waiting of the vehicles and pedestrians was noted over the one-hour simulation
period with an average of 30 traffic cycle lengths, keeping the vehicle’s volume constant. All
the parameters were kept same in one combination set of vehicle and pedestrian volume,
while changing the seed value for each simulation. This would generate slightly different
spawning patterns of vehicles and pedestrians in the simulation.

Table 6. Parameter settings for hypothesis test simulations.

Volume Category Traffic Light Timing

Through Left Turn
Vehicle Pedestrian Pedestrians
Vehicles Vehicles
Low 12 4 9
Medium 12 5 9
Low
High 12 6 9
Very high 12 7 9
Low 12 4 12
Medium 12 5 12
Medium
High 12 6 12
Very high 12 7 12
Low 16 4 15
Medium 16 5 15
High
High 16 6 15
Very high 16 7 15
Low 22 4 19
Medium 22 5 19
Very high
High 22 6 19
Very high 22 7 19

A one-way ANOVA (analysis of variance) test was performed to compare the differ-
ence of mean among the various combinations of traffic category groups as listed. It is a
statistical method used to analyze the differences between the means of three or more inde-
pendent groups based on a single categorical independent variable. A one-way ANOVA
aims to determine whether there is a significant difference among the means of the groups,
suggesting that the factor has a significant effect on the continuous dependent variable.
Table 7 lists the vehicle waiting time for pedestrian volume change, and Table 8 lists the
pedestrian waiting time for vehicle volume change.
The results show that the p-value for each group combination is less than the 0.05 sig-
nificance level. Thus, the null hypothesis is rejected, and there is a difference in the mean of
the groups. Similarly, for the pedestrian waiting time, the p-value for each group combina-
tion is less than the significance level, indicating that the change in user volumes affects the
other users at the intersections. Additionally, an examination of the average waiting times
reveals that the traffic waiting time increases as the volumes change from low to high. The
results are given in Tables 9 and 10.
Sensors 2023, 23, 5028 20 of 25

Table 7. Effect of pedestrian volume change on vehicle waiting time.

Vehicle Pedestrian
Vehicle Waiting Time (s) Average Std. Dev.
Volume Volume
Low 12.3 11.4 12 12.4 12.3 12.2 12.1 12.3 12.2 11.7 12.084 0.31711
Medium 13.1 12.4 12.8 12.6 12.7 12 12.1 11.7 12.1 12.7 12.422 0.45161
Low
High 13.1 13.8 13.6 12.9 13 12.5 12.8 12.9 13 13 13.065 0.37041
Very high 13.3 13.7 13 13.5 13.2 13.4 13 13.8 13.6 13 13.334 0.30537
Low 16.4 16.3 15.9 15.7 16.7 15.9 15.2 15.9 15.7 15.1 15.863 0.50953
Medium 16.4 16.2 15.7 17 16.1 15.6 16.1 17 17 16.6 16.368 0.52434
Medium
High 16.1 17.4 17 16.5 17.5 17 16.4 16 16.3 17.7 16.787 0.61056
Very high 17.5 17.2 16.5 16.4 17.7 16.9 17.2 16.5 16.5 18 17.037 0.55331
Low 19.4 18.5 19 20.1 19.4 19.3 18.6 18.8 19.2 18.6 19.077 0.48585
Medium 19.9 21.2 20.6 19.2 20 20.7 20.4 20.1 20 20.1 20.207 0.55542
High
High 20.8 21.1 22 20.1 19.9 19.7 19.9 20.9 20.6 21.5 20.63 0.75673
Very high 20.9 19.6 19.9 21.3 21.9 21.7 22.3 21.8 20.1 21.1 21.062 0.91183
Low 28.6 29.8 26.8 28.2 25.6 27.5 26 32.5 28.1 27.9 28.095 1.9663
Medium 32.6 27.5 30.8 28.2 26.3 32 27.1 30.1 32.4 29.8 29.668 2.31398
Very high
High 32.1 28.1 33.3 33.2 30.8 30.3 32.5 31.5 29.2 31.3 31.227 1.68779
Very high 32.1 31.6 30.7 31.1 30.2 32.9 33 33.3 28.8 31.6 31.516 1.40426

Table 8. Effect of vehicle volume change on pedestrian waiting time.

Pedestrian Vehicle
Pedestrian Waiting Time (s) Average Std. Dev.
Volume Volume
Low 29.1 26.3 26.2 27.7 29.4 27.9 28.3 27.7 27.9 27.9 27.825 0.97328
Medium 31 28.5 28.7 30.6 29.8 29.1 28.5 27.5 29.3 30.5 29.343 1.04039
Low
High 36 33.4 34.3 36.7 35.4 34.4 34.4 33.9 33.8 35.1 34.737 0.97946
Very high 42.2 44.3 41.8 41.9 40.1 41 41 39.7 42.3 41.3 41.547 1.21186
Low 29.6 30.1 29.9 30 30.6 30.4 30 27.5 29.3 28.4 29.566 0.92792
Medium 31.1 30.7 30.7 32.2 31.7 31.2 32.2 32 30.7 31.4 31.381 0.58911
Medium
High 35.9 35.1 36 35.2 36.8 37.7 37.1 36.2 35.5 36 36.154 0.80135
Very high 42.8 43.3 42.4 41.8 42.9 43.8 40.7 42.3 42.8 44.6 42.742 1.0153
Low 32.2 31.6 28.3 29.9 30 31.5 30.9 30.4 30 30.6 30.539 1.0476
Medium 32.7 33.5 32.5 32.7 32.2 32.3 32.8 32.3 32.1 32.7 32.565 0.37824
High
High 37 37.8 37.4 38.4 36.8 38.2 36 35.9 37.5 38.6 37.349 0.90547
Very high 43 42.6 44.8 44.8 44 44.7 42.3 44 45.8 44.2 44.019 1.0458
Low 31.7 30.6 31.3 31.4 31.4 31 31.2 30.4 31.7 32.2 31.28 0.49786
Medium 33.6 33.8 35.3 33.5 33.8 33.4 33.5 33.8 33.6 43.3 34.748 2.88001
Very high
High 39.2 37.4 38.5 40 38.4 41 38.6 39.2 39.5 39.5 39.137 0.94161
Very high 47 45.4 44.8 43.6 45.5 46.3 46.2 45.4 45.3 46.6 45.619 0.92143

Table 9. Results of ANOVA test for vehicle waiting time.

One-Way ANOVA Test F p-Value F Critical Hypothesis Result

Low vehicle vs. variable pedestrian volume 24.65366825 7.58 × 10−9 2.866265551 Reject null hypothesis
Medium vehicle vs. variable pedestrian volume 8.715096705 0.000177 2.866265551 Reject null hypothesis
High vehicle vs. variable pedestrian volume 14.92596812 1.79 × 10−6 2.866265551 Reject null hypothesis
Very high vehicle vs. variable pedestrian volume 7.101865068 0.000721 2.866265551 Reject null hypothesis
Sensors 2023, 23, 5028 21 of 25

Table 10. Results of ANOVA test for pedestrian waiting time.

One-Way ANOVA Test F p-Value F Critical Hypothesis Result

Low pedestrian vs. variable vehicle volume 311.4556 8.51 × 10−6 2.866266 Reject null hypothesis
Medium pedestrian vs. variable vehicle volume 432.7093 2.77 × 10−28 2.866266 Reject null hypothesis
High pedestrian vs. variable vehicle volume 409.7162 7.21 × 10−28 2.866266 Reject null hypothesis
Very high pedestrian vs. variable vehicle volume 133.925 1.39 × 10−19 2.866266 Reject null hypothesis

The machine learning performance was also tested. The ANN and SVM models were
given 1200 data samples to predict the green light timing for detected traffic conditions.
Out of all the samples in the dataset, 70% of the data were used for training, and 30% were
used in testing the neural networks. Table 11 shows the mean square error (MSE), R2 value,
average training, and testing accuracy of both models. The results indicate that both models
performed well, with the SVM model performing slightly better than the ANN model.

Table 11. Performance of ANN and SVM models.

Neural Training Validation

MSE R2
Networks Accuracy Accuracy
ANN 97.897% 97.678% 0.44834 0.93813
SVM 98.557% 98.191% 0.47739 0.93412

SVM’s performance was better in this case because the relationship between the input
and output data was fairly linear so that the model could find an optimal decision boundary,
called the maximum–margin hyperplane, which tends to generalize well to unseen data.
Additionally, the models did not need to predict any complex patterns in the traffic light
timing. Moreover, SVM models tend to be more interpretable than ANNs, as they involve
finding a hyperplane that separates the data. This can be important when the goal is not
only to make predictions but also to understand the underlying patterns in the data.
On the other hand, the ANN model accuracy was very close to the SVM model
accuracy. However, ANN models are more vulnerable to overfitting, which may reduce the
accuracy of the ANN model in predicting the output.
Ten different pre-trained CNNs were used to predict the traffic condition from images.
The pre-trained models are typically trained on a large dataset and have learned valuable
features that can be reused in a new task. The training and testing accuracies of all the
networks are listed in Table 12. The results show that almost all the models performed
well in detecting the vehicles. Darknet53 achieved the highest accuracy of 97.5% validation
accuracy compared to others. Darknet53 is a CNN model that has a total of 53 layers.
This model performs well in detecting the objects from the given image dataset. Simi-
larly, Darknet19, Alextnet, and Inceptionv3 are known for their high accuracy in object
recognition.
The performance of the different CNN models depends on their architecture. Different
architectures have unique strengths and weaknesses. Resnet18 is designed to handle
deeper networks by introducing skip connections, while Mobilenetv2 uses depth-wise
separable convolutions for efficient computation. More complex models such as Resnet101
can capture more intricate patterns but may require more computation and be prone
to overfitting. Simpler models such as SqueezeNet and Mobilenetv2 are designed for
efficiency, trading off some accuracy for reduced computational requirements. Thus, their
validation accuracies are 47.5% and 80%, respectively, which are the least compared to
other models.
Sensors 2023, 23, 5028 22 of 25

Table 12. Performance of CNN models.

Vehicle Detection Pedestrian Detection

Pre-Trained
Training Validation Training Validation
Networks
Accuracy Accuracy Accuracy Accuracy
squeezenet 46.875% 47.5% 36.7188% 42.5%
googlenet 69.5313% 82.5% 54.6875% 57.5%
resnet18 92.1875% 90% 73.4375% 55%
mobilenetv2 88.2813% 80% 78.1250% 62.5%
resnet50 92.1875% 82.5% 85.9375% 57. 5%
resnet101 100% 85% 100% 60%
inceptionv3 98.4375% 90% 95.3125% 53.5%
alexnet 87.5% 92.5% 75% 67.5%
darknet19 100% 90% 100% 60%
darknet53 100% 97.5% 100% 55%

However, the validation accuracy of all models plunged for pedestrian detection
cases. This happened because of the incorrect data from the simulation. Some images of
the pedestrian lanes taken during the simulation represented more pedestrians than the
detected category. Pedestrians can move in two directions on a single lane, whereas the
vehicle forms a proper queue and moves in a single direction only on a single lane. So,
while capturing the images of the pedestrians looking towards the signal, the camera also
captured the pedestrians moving in the opposite direction of the traffic signal.
The method proposed in this work was compared with the time-gap-based and
delay-based traffic light actuation methods [45,46]. These two methods are established
methods and have been proven to reduce traffic delays at intersections in real settings and
in simulations. Table 13 compares the average waiting time of vehicles for 12 h simulations
for low, medium, high, and very high volumes of vehicles. In these simulations, pedestrian
flows were not considered. From the table, some instructive points can be noted. The
time-gap-based and delay-based traffic light control strategy performed better for the low
and medium volume of vehicles, whereas the waiting time was similar for a high traffic
volume. The dynamic traffic interval technique presented in this work outperformed the
others under the condition of very high traffic volume. This is because, in the dynamic
traffic intervals technique, the times are fixed for a detected volume of traffic for a given
phase in a cycle. Thus, even if there are more cars coming at an intersection after a certain
gap or time delay, the signal will remain green for the fixed period.

Table 13. Waiting time of vehicles for three different traffic light control strategies.

Dynamic Traffic
Traffic Volume Gap-Based (s) Delay-Based (s)
Intervals (s)
Low 12.13 9.72 9.09
Medium 13.81 13.22 12.27
High 15.06 15.77 15.06
Very high 20.19 28.46 35.48

The dynamic traffic light interval technique was also tested alongside fixed time and
semi-dynamic traffic light plans to evaluate the performance in terms of the delay. In these
simulations, the pedestrian and vehicle flows were considered. Twenty-five simulations
were conducted to test out the proposed technique thoroughly. The experimental design
was structured to adjust the categories according to the time of day. The fixed-time traffic
light plan used very-high-category-traffic timing, and it remained fixed irrespective of the
change in input probabilities. Similarly, the semi-dynamic traffic plan used the timings of
medium- and very-high-category traffic. The time in semi-dynamic traffic lights would
change based on the rush and non-rush hours of the day.
Sensors 2023, 23, 5028 23 of 25

Based on the simulation results, there was a 20.11% to 26.67% reduction in the overall
vehicle waiting time for an intersection compared to the fixed-time traffic light system
and an 11.99% to 18.09% reduction compared to the semi-dynamic traffic light system.
Additionally, a 20.38% to 23.16% drop was observed in the overall pedestrian waiting
time at an intersection compared to the fixed-time traffic light system and an 8.55% to
10.95% reduction compared to the semi-dynamic traffic light system. Table 14 shows the
performance comparison for the three different traffic light control strategies.

Table 14. Performance comparison of different traffic light control strategies.

Vehicle Waiting Time (s) Pedestrian Waiting Time (s)

Technique
Average Std. Dev. Minimum Maximum Average Std. Dev Minimum Maximum
Fixed time 24.58 0.5616 23.93 26.25 42.36 0.1533 42.22 42.53
Semi-
22.03 0.5171 21.25 23.5 36.16 0.1654 36.63 37
dynamic
Dynamic 18.77 1.4653 18.28 19.25 33.16 0.2576 32.76 33.33

6. Conclusions and Future Work

This work proposed a system that simultaneously reduced the waiting time of vehicles
and pedestrians by categorizing the vehicle and pedestrian volume into categories. The
model achieved this result by adjusting the green light timings of the ongoing phase.
Reducing the waiting time of vehicles at the intersection means indirectly lowering vehicle
emissions, reducing fuel consumption, and effectively utilizing the traffic signal cycle length.
Moreover, lowering the delays for pedestrians decreases the chances of the dangerous
behavior of people trying to cross the roads.
This work also demonstrated a framework to implement the dynamic traffic interval
strategy. The cyber system, using CNN, proved its usefulness in detecting traffic condi-
tions at the intersection. ANN/SVM helped to set the timing for the combination of the
pedestrian and vehicle traffic condition. Moreover, ladder logic programming controlled
the traffic light timings based on the traffic condition detected. In this model, all vehicles
were assumed as passenger cars having an equal length, and each car approaching the
intersection maintained the same distance from the other. Additionally, the simulation
considered the same critical volume for all through movements of vehicles. As this tech-
nique is dependent on the service requested by the users, it will not be applicable at all
intersections. This is particularly true in areas where the demand is very low, where few
pedestrians are requesting a service, or when traffic variation is minimal over the course of
the day.
In the future, more parameters can be provided as inputs to machine learning mod-
els to predict traffic conditions accurately. Moreover, green light timing was calculated
based on the number of vehicles and pedestrians. In the future, more parameters will
be considered, such as pedestrians and vehicle speed, and other vehicles, such as trucks,
buses, and motorbikes, to adjust the green light time. One crucial assumption in this
research was that the cameras provide all the data to the cyber system. However, this may
not be possible in all scenarios; thus, data from different sensors such as ultrasonic and
infrared will be considered in the future. Additionally, this work focused on the working of
signal intersections. So, future work will include a control strategy considering multiple
coordinating intersections.

Author Contributions: Methodology, S.D.; Formal analysis, S.D.; Writing—original draft, S.D.;
Supervision, S.-J.H. All authors have read and agreed to the published version of the manuscript.
Funding: The work is partially supported by Texas A&M-FAPESP award No. 2019-3-01. Any
opinions, findings, and conclusions or recommendations expressed in this material are those of the
authors and do not necessarily reflect the views of Texas A&M University.
Institutional Review Board Statement: Not applicable.
Sensors 2023, 23, 5028 24 of 25

Informed Consent Statement: Not applicable.

Data Availability Statement: The data presented in this study are openly available in FigShare at
https://doi.org/10.6084/m9.figshare.22689676, reference number 22689676 (accessed on 12 May 2023).
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Manual on Uniform Traffic Control Devices for Streets and Highways; FHWA: Washington, DC, USA, 2009. Available online:
https://mutcd.fhwa.dot.gov/pdfs/2009/mutcd2009edition.pdf (accessed on 10 May 2023).
2. Lv, W.; Zhou, X.; Fang, Z.; Huo, F.; Li, X. Simulation study of vehicle travel time on route with signals considering comprehensive
influencing factors. Phys. A Stat. Mech. its Appl. 2019, 530, 121389. [CrossRef]
3. Urbanik, T.; Tanaka, A.; Lozner, B.; Lindstrom, E.; Lee, K.; Quayle, S.; Beaird, S.; Tsoi, S.; Ryus, P.; Gettman, R.; et al. Signal Timing
Manual; Transportation Research Board: Washington, DC, USA, 2015; Volume 1.
4. Highway Capacity Manual: A Guide for Multimodal Mobility Analysis, 6th ed.; Transportation Research Board: Washington, DC, USA,
2016.
5. Ghazal, B.; ElKhatib, G.; Chahine, K.; Kherfan, M. Smart traffic light control system. In Proceedings of the 2016 Third International
Conference on Electrical, Electronics, Computer Engineering and Their Applications (EECEA), Beirut, Lebanon, 21–23 April 2016;
pp. 140–145.
6. Díaz, N.; Guerra, J.; Nicola, J. Smart traffic light control system. In Proceedings of the 2018 IEEE Third Ecuador Technical Chapters
Meeting (ETCM), Cuenca, Ecuador, 15–19 October 2018; pp. 1–4.
7. Silva, C.M.; Aquino, A.L.L.; Meira, W. Smart traffic light for low traffic conditions: A solution for improving the drivers safety.
Mob. Netw. Appl. 2015, 20, 285–293. [CrossRef]
8. Alharbi, A.; Halikias, G.; Sen, A.A.A.; Yamin, M. A framework for dynamic smart traffic light management system. Int. J. Inf.
Technol. 2021, 13, 1769–1776. [CrossRef]
9. Sen, S.; Head, K.L. Controlled optimization of phases at an intersection. Transp. Sci. 1997, 31, 5–17. [CrossRef]
10. Li, M.; Alhajyaseen, W.; Nakamura, H. A traffic signal optimization strategy considering both vehicular and pedestrian flows. In
Proceedings of the Compendium of Papers CD-ROM, the 89th Annual Meeting of the Transportation Research Board, Washington,
DC, USA, 8–10 January 2010; pp. 10–14.
11. Younis, O.; Moayeri, N. Employing cyber-physical systems: Dynamic traffic light control at road intersections. IEEE Internet
Things J. 2017, 4, 2286–2296. [CrossRef]
12. Tchuitcheu, W.C.; Bobda, C.; Pantho, M.J.H. Internet of smart-cameras for traffic lights optimization in smart cities. Internet Things
2020, 11, 100207. [CrossRef]
13. Gandhi, M.M.; Solanki, D.S.; Daptardar, R.S.; Baloorkar, N.S. Smart control of traffic light using artificial intelligence. In
Proceedings of the 2020 5th IEEE International Conference on Recent Advances and Innovations in Engineering (ICRAIE), Jaipur,
India, 1–3 December 2020; pp. 1–6.
14. Pratama, B.; Christanto, J.; Hadyantama, M.T.; Muis, A. Adaptive traffic lights through traffic density calculation on road pattern.
In Proceedings of the 2018 International Conference on Applied Science and Technology (iCAST), Manado, Indonesia, 26–27
October 2018; pp. 82–86.
15. Chavan, S.S.; Deshpande, R.S.; Rana, J.G. Design of intelligent traffic light controller using embedded system. In Proceedings of
the 2009 Second International Conference on Emerging Trends in Engineering & Technology, Nagpur, India, 16–18 December
2009; pp. 1086–1091.
16. Hirankitti, V.; Krohkaew, J. An agent approach for intelligent traffic-light control. In Proceedings of the First Asia International
Conference on Modelling & Simulation (AMS’07), Phuket, Thailand, 27–30 March 2007; pp. 496–501.
17. Almawgani, A.H.M. Design of real time smart traffic light control system. Int. J. Ind. Electron. Electr. Eng. 2018, 6, 43–47.
18. Li, Y.; Chen, Y.; Yuan, S.; Liu, J.; Zhao, X.; Yang, Y.; Liu, Y. Vehicle detection from road image sequences for intelligent traffic
scheduling. Comput. Electr. Eng. 2021, 95, 107406. [CrossRef]
19. Wiering, M.A.; van Veenen, J.; Vreeken, J.; Koopman, A. Intelligent Traffic Light Control; Utrech University: Utrech, The Netherlands,
2004.
20. Wei, H.; Zheng, G.; Yao, H.; Li, Z. Intellilight: A reinforcement learning approach for intelligent traffic light control. In Proceedings
of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, London, UK, 19–23 August 2018;
pp. 2496–2505.
21. Liu, Y.; Liu, L.; Chen, W.-P. Intelligent traffic light control using distributed multi-agent Q learning. In Proceedings of the 2017
IEEE 20th International Conference on Intelligent Transportation Systems (ITSC), Yokohama, Japan, 16–19 October 2017; pp. 1–8.
22. Liang, X.; Du, X.; Wang, G.; Han, Z. A deep reinforcement learning network for traffic light cycle control. IEEE Trans. Veh. Technol.
2019, 68, 1243–1253. [CrossRef]
23. Göttlich, S.; Potschka, A.; Ziegler, U. Partial outer convexification for traffic light optimization in road networks. SIAM J. Sci.
Comput. 2017, 39, B53–B75. [CrossRef]
24. Park, B.; Messer, C.J.; Urbanik, T. Traffic signal optimization program for oversaturated conditions: Genetic algorithm approach.
Transp. Res. Rec. 1999, 1683, 133–142. [CrossRef]
Sensors 2023, 23, 5028 25 of 25

25. Genders, W.; Razavi, S. Using a deep reinforcement learning agent for traffic signal control. arXiv 2016, arXiv:1611.01142.
26. De Oliveira, M.B.W.; De Almeida Neto, A. Optimization of traffic lights timing based on multiple neural networks. In Proceedings
of the 2013 IEEE 25th International Conference on Tools with Artificial Intelligence, Herndon, VA, USA, 4–6 November 2013;
pp. 825–832.
27. Abbas, M.K.; Karsiti, M.N.; Napiah, M.; Samir, B.B.; Al-Jemeli, M. High accuracy traffic light controller for increasing the given
green time utilization. Comput. Electr. Eng. 2015, 41, 40–51. [CrossRef]
28. McKenney, D.; White, T. Distributed and adaptive traffic signal control within a realistic traffic simulation. Eng. Appl. Artif. Intell.
2013, 26, 574–583. [CrossRef]
29. Zhou, B.; Cao, J.; Wu, H. Adaptive traffic light control of multiple intersections in wsn-based ITS. In Proceedings of the 2011 IEEE
73rd Vehicular Technology Conference (VTC Spring), Budapest, Hungary, 15–18 May 2011; pp. 1–5.
30. Cruz-Piris, L.; Rivera, D.; Fernandez, S.; Marsa-Maestre, I. Optimized sensor network and multi-agent decision support for smart
traffic light management. Sensors 2018, 18, 435. [CrossRef]
31. Navarro-Espinoza, A.; López-Bonilla, O.R.; García-Guerrero, E.E.; Tlelo-Cuautle, E.; López-Mancilla, D.; Hernández-Mejía, C.;
Inzunza-González, E. Traffic Flow Prediction for Smart Traffic Lights Using Machine Learning Algorithms. Technologies 2022, 10,
5. [CrossRef]
32. Muntean, M.V. Multi-Agent System for Intelligent Urban Traffic Management Using Wireless Sensor Networks Data. Sensors
2022, 22, 208. [CrossRef]
33. Madrigal Arteaga, V.M.; Pérez Cruz, J.R.; Hurtado-Beltrán, A.; Trumpold, J. Efficient Intersection Management Based on an
Adaptive Fuzzy-Logic Traffic Signal. Appl. Sci. 2022, 12, 6024. [CrossRef]
34. Nimac, P.; Krpič, A.; Batagelj, B.; Gams, A. Pedestrian Traffic Light Control with Crosswalk FMCW Radar and Group Tracking
Algorithm. Sensors 2022, 22, 1754. [CrossRef]
35. Webster, F.V. Traffic Signal Settings. Road Research Technical; Paper no. 39; Transport Research Laboratory: Berkshire, UK, 1958.
36. Traffic Design Manual Traffic Signal Design—Operations and Coordination; TDOT: Nashville, TN, USA, 2020.
37. Lopez, P.A.; Behrisch, M.; Bieker-Walz, L.; Erdmann, J.; Flötteröd, Y.P.; Hilbrich, R.; Lücken, L.; Rummel, J.; Wagner, P.; Wießner, E.
Microscopic Traffic Simulation using SUMO. In Proceedings of the 2019 IEEE Intelligent Transportation Systems Conference
(ITSC), Maui, HI, USA, 4–7 November 2018; IEEE: Piscataway, NJ, USA, 2018; pp. 2575–2582.
38. Rouphail, N.; Hummer, J.; Allen, P.; Milazzo, I.I. Capacity Analysis of Pedestrian and Bicycle Facilities; No. FHWA-HRT-98-106; US
Department of Transportation. Federal Highway Administration, Research, Development, and Technology, Turner-Fairbank
Highway Research Center: McLean, VA, USA, 1998.
39. Darroch, J.N. On the traffic-light queue. Ann. Math. Stat. 1964, 35, 380–388. [CrossRef]
40. Little, J.D.C. Approximate expected delays for several maneuvers by a driver in Poisson traffic. Oper. Res. 1961, 9, 39–52.
[CrossRef]
41. Newell, G.F. Approximation methods for queues with application to the fixed-cycle traffic light. SIAM Rev. 1965, 7, 223–240.
[CrossRef]
42. Otis, S.C.; Machemehl, R.B. Analysis of Pedestrian Signalization in Suburban Areas; No. SWUTC/99/472840-00065-l; University of
Texas at Austin, Center for Transportation Research: Austin, TX, USA, 1999.
43. Transportation Performance Management—Peak Hour Excessive Delay Measure; FHWA: Washington, DC, USA, 2017.
44. Houston Traffic Report. 2022. Available online: https://www.tomtom.com/traffic-index/houston-traffic/ (accessed on 10 May
2023).
45. Oertel, R.; Wagner, P. Delay-time actuated traffic signal control for an isolated intersection. In Proceedings of the 90th Annual
Meeting Transportation Research Board (TRB), Washington, DC, USA, 23–27 January 2011.
46. Erdmann, J.; Oertel, R.; Wagner, P. VITAL: A simulation-based assessment of new traffic light controls. In Proceedings of the 2015
IEEE 18th International Conference on Intelligent Transportation Systems, Gran Canaria, Spain, 15–18 September 2015; pp. 25–29.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.