International Journal of Computational Intelligence Systems

In Press, Uncorrected Proof

DOI: https://doi.org/10.2991/ijcis.d.200522.001; ISSN: 1875-6891; eISSN: 1875-6883
https://www.atlantis-press.com/journals/ijcis/

Research Article
An Intelligent Traffic Light System Using Object Detection
and Evolutionary Algorithm for Alleviating Traffic
Congestion in Hong Kong

Sin-Chun Ng*, , Chok-Pang Kwok

School of Science and Technology, The Open University of Hong Kong, 30 Good Shepherd Street, Ho Man Tin Kowloon, Hong Kong, China

ARTICLE INFO ABSTRACT

Article History High traffic flow is a typical characteristic of a mobilized city with a high population. Efficient traffic manage-
Received 29 Feb 2020 ment is a proper solution to reduce the stress and anxiety associated with driving or traveling. The road users can
Accepted 17 May 2020 have better timing for traveling as they will not experience journey delays due to traffic congestion. In Hong Kong,
until September 2017, over 2,100 kilometers long roads around Hong Kong island, Kowloon Peninsula, and New
Keywords Territories are serving over 762,000 vehicles, Hyd.gov.hk, Highways Department - Hong Kong Road Network, n.d.,
Intelligent traffic light system https://www.hyd.gov.hk/en/road_and_railway/existing/road_network/road.html. However, the current traffic signal control sys-
Traffic congestion tems used in Hong Kong are mainly pre-defined fixed-cycle traffic light systems. These systems do not consider too much about
Object detection the real traffic situations such as vehicle and pedestrian counts, delay and waiting time of the road users. They respond slowly to
Evolutionary algorithm regulate traffic flow especially when there is a high volume of traffic. A study of effective optimization technologies in controlling
traffic signals is conducted which aims to relief the congestion problem and increase road efficiency according to the specific
needs of Hong Kong. In this paper, a new traffic light system using machine learning with object detection and analyzing by the
evolutionary algorithm that aims to perform a real-time strategic signal switching arrangement to traffic lights at the intersection
was designed to reduce the waiting time of pedestrians and vehicles, and provide better traveling experience to road users.

© 2020 The Authors. Published by Atlantis Press SARL.

This is an open access article distributed under the CC BY-NC 4.0 license (http://creativecommons.org/licenses/by-nc/4.0/).

1. INTRODUCTION and large amounts of particulates are emitted. These pollutants are
trapped easily in urban because polluted air cannot be dissipated by
In city development, well infrastructure and management of the many tall buildings and high usage of roads. Hence, poor air quality
roads will allow the citizens to travel efficiently from one place will cause much illness, mainly in the respiratory system. Charpin
to another, especially during peak hours, to improve the traveling and Caillaud [3] proved that PM2.5, nitrogen oxides, and related
experiences and enjoy safety traveling. Designing an intelligent traf- pollutants caused the nasal allergic reaction to people, and ease to
fic light system (ITLS) that 100% fulfills the city’s requirement is aggravate other respiratory diseases like asthma, rhinosinusitis, and
critically hard for governors and stakeholders. middle ear disease. Also, mental illness such as traffic stress syn-
Traffic congestion is a serious issue, not just affecting the citizens, drome will be induced when people trapped in a narrow area with
but also reducing the interests of the business activities. Although long traffic queues.
no statistics are reflecting the economic loss by traffic congestion Traffic congestion is common to be seen on roads in Hong Kong.
in Hong Kong, but there are some proofs from the United States However, compared to other cities, the fixed-cycle traffic light
and European countries that traffic congestion inferred serious eco- (FCTL) system is currently in use for many years. Without consid-
nomic loss. In United States, a report released in 2014 [1] that ering the real-time conditions of the road to maintain a suitable
US$124 billion was spent in a year due to traffic congestion. In traffic arrangement, pedestrians and vehicles are usually in compet-
European Union, traffic congestion costs about 1% of its GDP [2]. itive status in using the cross side, so dangerous situations are easy
It shows that traffic congestion will lead to huge economic loss. to occur because both of them do not want to wait in a junction.
In addition to the economic issue, traffic congestion will dam- Hsieh and You [4] proposed a scheme by applying an evolutionary
age human health. Although many vehicles are using environmen- algorithm in managing visitor routines to exhibition rooms in the
tal fuels, for example, liquefied petroleum gas that emits fewer museum. The scheme successfully realized the flow management
pollutants, some public transport vehicles like buses, trucks and in the museum to have the best experience for all visitors. Similar
other heavy vehicles are using diesel as the fuel, so nitrogen oxides to the traffic control, vehicles can be considered as the visitors in
entering the exhibition room, and intersections are the exhibition
*
Corresponding author. Email: cpkwok@ouhk.edu.hk
rooms that can perform the best traffic management. In reducing
2 S.-C. Ng and C.-P. Kwok / International Journal of Computational Intelligence Systems, in press

traffic congestion, an ITLS was developed by applying object detec- to optimize traffic control. It provides much valuable information
tion and evolutionary algorithm. Classification and determination to other researchers to develop an adaptive traffic control system
of the real-time situations captured on camera set at the junctions effectively. On Tai Tam Road (dam section), there is a specific traf-
were accomplished by object detection. Data were analyzed and fic condition, and fewer traffic situations occur due to lack of traffic
then performed a strategic arrangement to the traffic lights by the control systems to perform the signal timing optimization at junc-
evolutionary algorithm that aims to reduce the waiting time of vehi- tions in urban districts, especially in Central, Mong kok, and Tsim
cles and pedestrians at in intersection to increase efficiency of road Sha Tsui in Hong Kong. Based on the available open data, such as
usage. street-facing surveillance cameras and navigation applications like
the street view function in Google Maps, real-time detecting the
2. LITERATURE REVIEW number of vehicles and density of pedestrians in waiting or crossing
the road at a specific intersection can be realized. This information
2.1. Current Situation of Traffic Signal describes the real-time traffic conditions of road intersections more
clearly, which could be a useful resource to drive the development
Control in Hong Kong of intelligent traffic signal control for cities.
2.1.1. FCTL system
The system only uses a fixed-cycle time in switching the traffic lights 2.2. Review of Current Technologies
at each junction, which does not consider the real-time condition of
2.2.1. Object detection of machine learning
the corresponding intersection, such as the number of vehicles and
technology
pedestrians at the junction. And it does not take care of their needs,
especially the waiting time and pass-through time. Even the gov- From Ref. [8] that applied object detection to classify the type of
ernment carried out a research project about extending the crossing vehicle when the vehicle passes through the checkpoints in Tai
time for the elderly [5], the results of the on-site trial may not reflect Tam Road, object detection is applicable in determining the type
the actual traffic conditions since almost all of them are placed in of object and also locating its position by quoting a bounding box
old districts or under-saturated junctions. around it. Furthermore, it also presents the information of the
detected object in an image once classified [9]. Hence, object detec-
2.1.2. Area traffic control system tion is well in use for detecting vehicles and pedestrians in the real
scenario by setting up cameras at intersections.
Area Traffic Control (ATC) system has been applied to Hong Kong
since 1977, which is the first computerized adaptive traffic signal
control system [6]. The system deployed two different adaptive traf- 2.2.2. Simulation
fic control technologies, namely, Split Cycle Offset Optimization A simulation is an imitation of a model based on a system in reality.
Technique (SCOOT) and Sydney Coordinated Adaptive Traffic Sys- A simulation model can change different parameters, then perform
tem (SCATS) [7] in which traffic signal was controlled by detect- testing for a system that is costly or difficult to be constructed. A
ing the number of vehicles and calculating their speed when they simulation model can be used to investigate different states [10].
passed through the detector loops set on the vehicle lanes. The data
are then transferred to the control server to adjust the parameters A simulation environment can be built to show the scenario by
to the adaptive control algorithm that result to an optimal duration using the new traffic light application, comparing with the current
of green lights to release the vehicles, thus reducing traffic conges- FCTL system. Moreover, it can simulate the operations in different
tion. However, the system focuses on management of the vehicles traffic conditions by using the new application.
passing through the junction, but sacrifices the benefits of pedestri-
ans that shortened the crossing time for pedestrians which increase 2.3. Review of Existing Solutions
the risk for people trying to cross the road when the red light is
on pedestrians. Besides, Ref. [7] stated that SCOOT and SCATS Smith et al. [11] proposed a scalable urban traffic control system in
cannot provide rapid responses to traffic fluctuations at various 2013 that monitors the number of vehicles in real-time by installing
intersections. detectors at intersection scanning the sensors installed in vehicles.
The traffic data are gathered that transferred to traffic control to set
up the scheduled scheme then apply to intersections. Chavan et al.
2.1.3. Smart traffic light system
[12] proposed an intelligent traffic light controller using an embed-
Smart traffic light system set up for Tai Tam Road (dam section) ded system in 2009. The driver receives route suggestions as a refer-
was developed and implemented by Cheng and his research team ence to the destination by processing unit, which connecting with
at The Chinese University of Hong Kong in the year 2018 [8]. The analyzing the traffic flow data came from sensors set of intersec-
type of vehicle will be detected by eight sets of cameras installed at tions. Nie [13] developed an intelligent traffic lights system using
the street lamps on Tai Tam Road. It will then calculate the green object detection to find out the number of vehicles at an intersection
signal time, according to the data received by all cameras analyz- by cameras. Then fine-tune the traffic light switching by the results
ing the traffic queue of each side of the dam and the presence of of the traffic density. Although those systems performed well, they
heavy vehicles in the line resulting into an efficient light arrange- gave up other vital elements like the precision that an incorrect
ment to avoid the traffic congestion at the dam because the design vehicle queue was detected by a container truck. Also, they may be
of the dam cannot hold two heavy vehicles like a double-decker bus costly, especially in Ref. [11], that needs to install an extra compo-
and fire engine in parallel. This system is a pioneer in Hong Kong nent to vehicles which means the system cannot be operated if vehi-
applies video analytic technology by real-time traffic data collection cles pass through the road without installing sensors.
 S.-C. Ng and C.-P. Kwok / International Journal of Computational Intelligence Systems, in press 3

According to the old-styled applications that cannot satisfy the The traffic models were used to simulate the situation so that the
requirements in modern cities, Emami et al. [14] suggested that current FCTL system and the proposed ITLS were implemented
machine learning technology is a way to solve the problem. Apply- at the intersection. Finally, the simulation displayed traffic models
ing to traffic control is increasing the efficiency in processing a visually.
large amount of data while data of traffic conditions are sets of
images. Furthermore, in using deep learning, more precise infor-
mation such as the type and speed of the vehicle can be realized. 3.3. Traffic Condition Videos
Fuzzy expert systems are introduced by Hawi et al. [15] in 2015
and they stated that the fuzzy logic controller proposed by Khiang Videos for different traffic conditions were recorded as sources for
et al. by implementing fuzzy logic started by Zadeh [16]. Fahmy the object detection application. Cameras were set at positions in
[17] presented with another system called FLATSC included wait- which traffic flow of the intersection can be seen clearly. The footage
ing time as a criterion in managing four intersection roundabout. was used for the object detection applications.
Besides, Hawi et al. also showed wireless sensor networks proposed
by Yousef et al. [18] in 2010 that the traffic data are shared between 3.4. Object Detection Application
intersections to optimize the traffic flow. While fuzzy logic and
wireless technologies become famous in solving this type of issue, The system applies object detection technology using Tensorflow
the fundamental problem has not been resolved. That is, sensors [20] to analyze the video footage. As shown in Figure 2, the system
are required to maintain the detection; hence much more costs can recognize the object type appeared in the footage. It can be used
will be required based on the well-developed traffic network such to calculate the number of vehicles and pedestrians near an inter-
as Hong Kong. Furthermore, to conduct the data sharing between section in a specific period of time.
intersections, and massive power to support the data exchange is
also needed.
While performing site testing may not be permitted in Hong Kong,
simulation can perform different conditions virtually that will not
affect the real traffic, especially in high traffic flow districts. SUMO
[19] simulator is applied to generate the actual scenario in Mong
kok. It creates the road map according to the online maps. Then,
the traffic flow is made and imported to the simulator. Although
SUMO can generate traffic flow very well, it cannot fulfill the needs
of this paper. The traffic map cannot be modified once created so
that there is a barrier to making a change to the road according to
rapid constructions made or new facilities being implemented.
In summary, sensors are frequently used in different solutions in
detecting traffic conditions. However, limited data gathering and
maintenance of these devices become complicated [13]. In improv-
ing traffic control, object detection technology becomes more flex-
ible and more practicable.

Figure 1 Product diagram of the system.

3. METHODOLOGY
3.1. Overview of Methodology
This paper applies a build-test methodology for the solution and
evaluation design. For the solution design, the system is developed
for an intelligent traffic light to control the traffic light signal auto-
matically based on the machine learning with object recognition
on the number of vehicles and pedestrians passing through a road
intersection in Hong Kong.
For the evaluation design, the simulation demonstrated the com-
parison between the current FCTL system and the proposed ITLS.

3.2. Design of the Solution

There were five components of the design shown in Figure 1. Videos
for different traffic conditions were taken at an intersection. The
object detection application analyzed the video footage. The data
of video footage analyzed were used for building traffic models. Figure 2 Object detection.
4 S.-C. Ng and C.-P. Kwok / International Journal of Computational Intelligence Systems, in press

3.5. FCTL Traffic Model such as another vehicle or a pedestrian was detected at the
front side of it.
This traffic model simulated the traffic flow of the current FCTL
at an intersection. After the simulation, the traffic model gener- • There were different types of vehicles at the intersection.
ated the average waiting time of the pedestrians and vehicles at an
intersection. The data were stored for algorithm development and
testing purposes. 3.8. Detailed Design of Algorithms
The traffic light control algorithms were shown in Figures 4 and 5.
Figure 4 shows the algorithm design of the ITLS. Figure 5 shows
3.6. Intelligent Traffic Light Traffic Model the current FCTL system FCTL used in Hong Kong. The condi-
The intersection was simulated with ITLS. The simulation was tions determine the flow of intelligent traffic light switching. The
based on the number of pedestrians and vehicles estimated from the “threshold” in the figure were the parameter thresholds of ITLS.
object detection application. The traffic light signal will be switched These parameter thresholds determine how the light signals were
dynamically due to different traffic conditions. This traffic model switched based on different situations.
aims to improve the efficiency of the traffic lights system currently The parameters thresholds were
used in Hong Kong.
• carMaxNumAtJunction—The maximum number of vehicles at
the junction.
3.7. Simulation
• carLightGreenMinTime—The minimum time of a vehicle
The pygame application simulates the above traffic models visually. green signal should last.
It used the data on traffic models to build a simulation that visual-
ized the concept of the traffic light system as shown in Figure 3. • carMaxWaitingtimeAtJunction—The maximum waiting time
for a vehicle at the junction.
There were several features to improve the realistic situation of the
traffic model: • pedMaxNumAtJunction—The maximum number of
pedestrians at the junction.
• The number of pedestrians and vehicles was set according to • pedLightGreenMinTime—The minimum duration time of the
the output of the object detection application. pedestrian green signal.
• The traffic light signal switching method was based on the • pedMaxWaitingtimeAtJunction—The maximum waiting time
analyzed data from the video footage. for a pedestrian at the junction.
• The road map was drawn according to the actual intersection • pedLightFlashLongerTime—The time limit to extend the
in Prince Edward in Kowloon, Hong Kong. pedestrian flashing green signal when there are still pedestrians
• The speed of pedestrians and vehicles were simulated to crossing the street.
real-life situation. There was also a probability that a person will
walk slower than average to simulate the movement of elderly. The default traffic light signal for vehicles and pedestrians was
green and red, respectively. When the value of input parameters was
• Vehicle movement. The vehicle slowed down if it is too close to higher than the threshold values, the traffic light signals will be trig-
a vehicle in front of it. The vehicle stopped either the traffic gered to change. There was three seconds for the vehicle traffic light
light is red or enforced to stop while a barrier was detected, to change from green to yellow, and from yellow to red to let the
pedestrians cross the road by switching the red pedestrian light to
green.
The cross-time for pedestrians was also based on the parameters.
Before the pedestrian traffic light becomes blinking green, the sys-
tem would keep checking if vehicle queue is less than the threshold
or the number of waiting pedestrians are more than the threshold,
then the time of green pedestrian traffic light would be extended. If
not, the pedestrian traffic light would change to blinking green.
While the green pedestrian traffic light is extended, the system
would keep checking if the vehicle queue is more than the thresh-
old, the pedestrian traffic light would then change to blinking green
immediately and let the vehicles move quickly.
Evolutionary algorithm is used to find the best set of parameter
thresholds for the system to improve the performance of the ITLS.
Before using the evolutionary algorithm, a regression model was
Figure 3 Pygame simulator presenting the traffic flow at an built to predict the simulator outcome based on the parameter set.
intersection. Several simulation runs have generated the average waiting time of
 S.-C. Ng and C.-P. Kwok / International Journal of Computational Intelligence Systems, in press 5

Figure 4 Flows of signal switching in intelligent traffic light system (ITLS).

The best parameter set by evolutionary algorithm will be deter-

mined by following steps:

1. Randomly generate a set of parameter thresholds (as stated in

Section 3.8) for vehicles and pedestrians as initial values.
2. Evaluate the fitness by the average waiting time for vehicles and
pedestrians using the parameter coefficient. Each coefficient is
the weighting factor for the parameter while the value is set in
the range according the actual conditions at the conjunction.
3. Sum up all partial values of the waiting period of the different
pedestrians and vehicles to become the waiting time for vehi-
cles and pedestrians.
4. Compare the new value with the current best value of the wait-
ing time (fitness), then pick the parameter set with the goal of
minimizing the fitness.
5. Repeat Steps 2 to 4 for various conditions at the junction to
perform the suitable fitness.

4. RESULT OF IMPLEMENTATION
Figure 5 Flows of signal switching in fixed-cycle traffic light
(FCTL).
The purpose of ITLS was to find a solution to alleviate traffic con-
gestion in Hong Kong. The implementation result of ITLS is pre-
pedestrians and vehicles. The simulation data were used for training sented in this section. It is used to evaluate the performance of ITLS
a regression model to predict the simulation output when the ITLS in improving the efficiency of road usage.
was implemented with different sets of parameter thresholds.
To evaluate the model, several traffic condition videos were
After the regression model was built, the evolutionary algorithm recorded, including non-peak and peak traffic periods. Each video
was used to find the optimal set of parameter thresholds in order recorded the traffic condition at the intersection between Prince
to achieve the best fitness with minimum average waiting time of Edward Road West and Embankment Road in Hong Kong for one
pedestrians and vehicles. hour to perform the simulation.
6 S.-C. Ng and C.-P. Kwok / International Journal of Computational Intelligence Systems, in press

The results were the comparison between the simulations that are From the result generated by the evolutionary algorithm as shown
using the ITLS for traffic control and the FCTL system. If the ITLS in Figure 6, there was a decreasing trend of the average waiting
decreased the waiting time of cars and pedestrians at junctions, time for vehicles and pedestrians at the junction after each genera-
the efficiency of road space usage increased. The simulations using tion of the evolutionary algorithm. There was a decrease of 44% to
ITLS to manage the regular traffic hours were run many times. The over 60% of the total waiting time with the optimal set of parame-
results were stated in Sections 4.1 and 4.2. ter thresholds that fulfills the requirements, for both peak and non-
peak hours.

4.1. Results of Non-peak Traffic Hours 4.3. Optimal Set of Parameter Thresholds
In Hong Kong, non-peak traffic hours are considered as the periods
With the use of evolutionary algorithm, an optimal set of param-
between 10:00 o 17:00 and 20:00 to 07:00 (the next day) in working
eter thresholds was found. The optimal set worked on both peak
days.
and non-peak traffic hours. The results showed that even drivers
Table 1 shows the simulation results of 4 test cases with different or pedestrians were waiting in peak hours, the waiting period was
time slots. By applying the ITLS at non-peak traffic hours, the aver- close to the time in non-peak traffic hours. By the results, the sys-
age waiting time of all road users reduced from around 50 seconds tem enhanced the efficiency when crossing the road or traffic light.
to less than 30 seconds in all cases. It reflected that the efficiency in
By using the optimal set of parameters, the average waiting time of
terms of the throughputs of the road can be increased 44% to over
drivers was about 10 seconds, and the average waiting time of pedes-
60% by using ITLS. Test cases 2 and 3 resulted in a lower increment
trians was approximately 13 seconds in non-peak traffic hours.
of efficiency because case 2 falls in the lunch hour that there are
probably more pedestrians and vehicles on the road. And the time The average waiting time of drivers was about 10.81 seconds, and
is close to commuter time and cut-off time of trucks for cargoes the average waiting time of pedestrians was approximately 14 sec-
delivery. onds in peak traffic hours.
According to the results simulated, the optimal set of parameter
thresholds were determined as listed in Table 3.
4.2. Results of Peak Traffic Hours
Peak hours are the periods between 07:00 to 10:00 and 17:00 to
20:00 in working days. 5. EVALUATION
Table 2 shows the simulation results of 4 peak traffic periods in 5.1. Evaluation of Object Detection Model
working days. Since there are many vehicles and pedestrians waiting
on the road intersection during peak hours. ITLS greatly reduced The dataset were divided into a training set and a testing set for
the total waiting time of road users. The overall efficiency (in terms the evaluation. Intersection over Union (IoU) was applied to deter-
of reduction in total waiting time) has been increased 54% to 57% mine the accuracy. We referred to the actual bounding box, and the
for all four cases. It reflected that the throughput at the intersection
has been increased a lot.

Table 1 Simulation results on non-peak traffic hours.

Overall Average Waiting Time (in Seconds)

Case Time Slot FCTL ITLS Efficiency Gained by
ITLS (%)
1 10:30–11:30 48 21 56.25
2 13:00–14:00 47 25 48.89
3 16:00–17:00 45 25 44.44
4 20:30–21:30 53 20 62.26
FCTL, fixed-cycle traffic light; ITLS, intelligent traffic light system.

Table 2 Simulation results on peak traffic hours.

Overall Average Waiting Time (in Seconds)

Case Time Slot FCTL ITLS Efficiency Gained by
ITLS (%)
1 07:30–08:30 52 23 55.77
2 09:00–10:00 49 21 57.14
3 17:30–18:30 48 22 54.17
4 19:00–20:00 50 22 56.00
FCTL, fixed-cycle traffic light; ITLS, intelligent traffic light system. Figure 6 Evolutionary algorithms progress.
 S.-C. Ng and C.-P. Kwok / International Journal of Computational Intelligence Systems, in press 7

predicted bounding box and calculated the overlapped area of the the number of vehicles and pedestrians in the video footage. A sim-
two bounding boxes. Table 4 shows the related outcomes. ulation was built for the comparison between the ITLS and the cur-
rent FCTL system. It included features to improve the realism of the
The average precision (AP) of the vehicle was 56% and for the
simulation. Besides, the evolutionary algorithm was developed to
pedestrian was 15%. The reason for AP of pedestrians resulting in
optimize the parameter thresholds and control the best traffic light
low accuracy is that both the ground-truth and prediction bounding
signal switching method for the ITLS. The average waiting time of
box are small; therefore, they are challenged to have a high overlap-
the pedestrians and the vehicles was used in finding the optimal
ping rate. Also, concerning more on the counting of objects rather
set of thresholds by using evolutionary algorithm. With the use of
than the IoU in implementation. The result meets the expectation
ITLS, it was shown that the efficiency of the road usage has been
for the ITLS with further improvements.
improved, and hence the congestion problem can be alleviated.

5.2. Evaluation of the Evolutionary

Algorithm ACKNOWLEDGMENTS
The result of the evolutionary algorithm reduces the average waiting The authors would like to express our great appreciation to Mr. Yui-Chung
time of vehicles and pedestrians. Fung, Mr. Chun-Yung So and Mr. Yuen-Ho Lam for collecting data and
performing experiments of this research works.
Figure 6 shows the results of each generation of the evolutionary
algorithm. The average waiting time of pedestrians and vehicles
were reduced by each generation. The average waiting time by com-
bining the pedestrian and vehicle was about 65.92 seconds in the REFERENCES
first generation. It became lower, with approximately 22.84 seconds
in the tenth generation. The total average waiting time was reduced [1] F. Guerrini, Traffic congestion costs Americans $124 billion a year,
by 65% by using evolutionary algorithm in finding the optimal set report says, Forbes, 2014.
of parameter thresholds. [2] The Economist, The cost of traffic jams, 2014. https://
www.economist.com/blogs/economist-explains/2014/11/
The evolutionary algorithm initialized the parent parameter set and economist-explains-1
then calculated the fitness of average waiting time of pedestrians [3] D. Charpin, D.M. Caillaud, Air pollution and the nose in chronic
and vehicles. After that, the process determined the parameter set respiratory disorders, The Nose and Sinuses in Respiratory Dis-
and calculated the eligibility of the parameter thresholds. When orders (ERS Monograph), in: C. Bachert, A. Bourdin, P. Chanez
compared the current fitness with the best fitness value, the result (Eds.), European Respiratory Society, Sheffield, UK, 2017,
optimized the choice of parameter set in each generation. The ITLS pp. 162–176.
applied the optimal set of parameter thresholds to model the traf- [4] Y. Hsieh, P. You, An efficient encoding scheme for a new multiple-
fic conditions in peak hour and non-peak hours in controlling the type museum visitor routing problem withmust-see and select-see
traffic flow. exhibition rooms, Int. J. Comput. Intell. Syst. 10 (2017), 677–689.
[5] Td.gov.hk, Transport department - smart devices for elderly using
signalised pedestrian crossing, n.d. https://www.td.gov.hk/en/
6. CONCLUSION smart_device/index.html
[6] C.K. Lee, K. Chan, S.T. Ho, An overview on traffic signal
In this paper, the ITLS based on machine learning with object detec- system and its maintenance in Hong Kong, in Proceeding
tion and the evolutionary algorithm was presented. An object detec- of 2014 International Conference on Electrical Engineering,
tion model was built to analyze traffic video footage. It outputted Jeju, Korea, 2014, pp. 15–19. https://www.emsd.gov.hk/
filemanager/conferencepaper/en/upload/55/cnfrnc-paper-
20140615-19-1.pdf
Table 3 Optimal set of parameter thresholds.
[7] A. Stevanovic, C. Kergaye, P.T. Martin, SCOOT and SCATS: a
Parameter Value closer look into their operations, in 88th TRB Annual Meeting,
Maximum number of vehicles at junction 12 Transportation Research Board, Washington, DC, USA, 2009.
Minimum time a car green signal should last 21 [8] C.H. Cheng, Smart traffic lights at Tai Tam Road
Maximum waiting time for a car at junction 12 (Dam Section), CU Engineering, 2019. http://cloud.
Maximum number of pedestrians at junction 11 itsc.cuhk.edu.hk/enewsasp/app/article-details.aspx/
Minimum time a pedestrian green signal should last 40 5BB75A736B42D708826F19533FFE0AF1/
Maximum waiting time for a pedestrian at junction 40
Time limit to extend the pedestrian flashing green signal [9] A.R. Pathak, M. Pandey, S. Rautaray, Application of deep learn-
when there are still pedestrians crossing the street 5 ing for object detection, Procedia Comput. Sci. 132 (2018),
1706–1717.
[10] A. Maria, Introduction to modelling and simulation, in Proceed-
Table 4 Average Precision(AP) of vehicle and pedestrian.
ings of the 29th conference on Winter simulation, IEEE Computer
AP Accuracy (%) Society, Atlanta, GA, USA, 1997, pp. 7–13.
0.5 IoU for vehicle 56.78 [11] S.F. Smith, G.J. Barlow, X.F. Xie, Z.B. Rubinstein, Surtrac: scalable
0.5 IoU for pedestrian 15.48 urban traffic control, in 92nd Annual Meeting of the Transporta-
AP, average precision; IoU, Intersection over Union. tion Research Board, Washington, DC, USA, 2013.
8 S.-C. Ng and C.-P. Kwok / International Journal of Computational Intelligence Systems, in press

[12] S.S. Chavan, R.S. Deshpande, J.G. Rana, Design of intelligent [17] M.M.M. Fahmy, An adaptive traffic signaling for roundabout
traffic light controller using embedded system, in 2009 Second with four approach intersections based on fuzzy logic, J. Comput.
International Conference on Emerging Trends in Engineering & Inf. Technol. 15 (2007), 33–45. http://cit.srce.unizg.hr/index.php/
Technology, IEEE, Nagpur, India, 2009, pp. 1086–1091. CIT/article/view/1625/1329
[13] Y., Nie, Intelligent traffic lights based on MATLAB, AIP Conf. [18] K.M. Yousef, J.M. Al-Karaki, A.M. Shatnawi, Intelligent traffic
Proc. 1995 (2018), 040139. light flow control system using wireless sensor networks, J. Inf.
[14] P. Emami, A. Rangarajan, S. Ranka, L. Elefteriadou, Machine Sci. Eng. 26 (2010), 753–768. http://www.iis.sinica.edu.tw/page/
learning for intelligent transportation systems, in UF Informatics jise/2010/201005_02.pdf
Institute Student Data Analysis Seminar, 2018. [19] B. Michael, et al., SUMO? Simulation of urban mobility: an
[15] R. Hawi, G. Okeyo, M. Kimwele, Techniques for smart traffic con- overview, in Proceedings of SIMUL 2011, The Third Interna-
trol: an in-depth review, Int. J. Comput. Appl. Technol. Res. 4 tional Conference on Advances in System Simulation, Think-
(2015), 566–573. Mind, 2011.
[16] L.A. Zadeh, Fuzzy sets, Inf. Control. 8 (1965), 338–353. http:// [20] M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean,
www.cs.berkeley.edu/~zadeh/papers/Fuzzy\__xunadd_text_ M. Kudlur, Tensorflow: a system for large-scale machine learning,
character:nN{\textbackslash}{\}%20SetsInformation\__xunadd_ in OSDI, Savannah, GA, USA, 2016, vol. 16, pp. 265–283.
text_character:nN{\textbackslash}{\}%20and\__xunadd_text_
character:nN{\textbackslash}{\}%20Control-1965.pdf