AUTOMATED TRAFFIC MANAGEMENT

2024 1st International Conference on Sustainable Computing and Integrated Communication in Changing Landscape of AI (ICSCAI) | 979-8-3503-5533-8/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICSCAI61790.2024.10866932

SYSTEM USING DEEP LEARNING BASED

OBJECT DETECTION
Dr. Poorna Shankar Dr. P. William Er. Shashank Shekhar Tiwari
Department of MCA, Indira College of Department of Information Assistant Professor, Department of
Engineering and Management, Pune Technology, Sanjivani College of Information Technology, REC
[email protected] Engineering, Kopargaon, SPPU, Pune Ambedkar Nagar, U.P, Dr. Abdul
[email protected] Kalam Technical University, Lucknow
[email protected]
Meenakshi Maindola Haider Alabdeli
Department of Computer Science & Department of computers Techniques
engineering, College of technical Kanchan Yadav
Engineering, Graphic Era Deemed to be Department of Electrical Engineering
engineering, The Islamic university,
University, Dehradun Najaf, Iraq GLA University, Mathura
[email protected] [email protected] [email protected]

Abstract— Despite having the second-largest road network management, which are dependent on human monitoring
in the world, India is responsible for ten percent of all deaths and control, are unable to deal with the complexity of
that occur on the world's roadways. In addition, there has modern traffic problems. Using the opportunities
been an increase in the amount of traffic congestion, air presented by cutting-edge technology, such as object
pollution, and noise pollution in smart cities as a result of the recognition that is based on deep learning, an increasing
proliferation of various kinds of vehicles, the technical number of individuals are showing an interest in
challenges associated with the technology used to operate
developing traffic management systems that are more
traffic signals, and the poor management of road traffic.
Even though they now rely on time-based methods, effective and efficient. This approach takes use of cutting-
traditional traffic management systems are unable to handle edge computer vision algorithms to automatically
the complicated flow of traffic that happens at intersections. identify, monitor, and analyse on the road any significant
This is the case even when these systems have been updated. components that may be present. These components may
Roadblocks may make travelling more stressful and cause include autos, people, and other things. When this is done,
further delays in addition to increasing the total number of it allows for decisions to be made in the moment that are
miles driven, the cost of transportation, and the amount of driven by data, which in turn supports traffic flow that is
pollution produced. We were able to accomplish this goal by both safer and more efficient [1].
developing a complete traffic light management system using
Through the use of object recognition and deep
the most recent version of YOLO, which was just recently
published. In terms of both speed and accuracy, the most learning, the purpose of this study is to evaluate the
current iteration of the YOLO algorithm, which goes by the feasibility of developing an Automated Traffic
name V7, outperforms all of the prior object recognition Management System (ATMS). By installing cutting-edge
models that have been developed. Because it is the fastest cameras, sensors, and deep learning algorithms at key
and most accurate real-time object identification model, it is intersections and highways, we can usher in a new era of
the best way to use in traffic management systems. As a traffic monitoring, analysis, and management. As a direct
result, it is the best approach to use. Therefore, doing so is consequence of this, we will have improved capabilities to
the most prudent thing to do in this situation. Yolo V7 is regulate and direct traffic. It is anticipated that the flow of
more than 120% faster than any of the previous
traffic will be maximised with the assistance of this
incarnations, and it also has the highest speed to accuracy
ratio of any of the versions. technology, which should help reduce the amount of
Keywords— Deep Learning, Traffic Management, Object congestion as well as the number of accidents while also
Detection, vehicle detection, YOLO, speed detection. contributing to a lower overall consumption of energy and
emissions [2-4].
I. INTRODUCTION In the parts that follow, we are going to discuss the
As the population of cities continues to expand at an primary attributes and benefits of the Automated Traffic
exponential rate, traffic management has emerged as one Management System that was recommended. We will
of the most pressing challenges confronting contemporary discuss the fundamental concepts that underpin deep
society. The growth of cities and the increase in the learning-based object recognition, as well as the required
number of cars on the road have both contributed to a hardware and software, as well as any potential
worsening of a variety of associated issues, including ramifications that this may have for public transit systems.
congestion, accidents, and a number of other related We will also investigate the system's social, economic,
issues. The quality of life of city dwellers is negatively and environmental advantages, with a particular focus on
impacted as a direct result of the aforementioned issues. It the environmental benefits' capacity to make urban areas
has been shown that traditional methods of traffic more livable. In this era of rapid technological

Authorized licensed use limited to: Yenepoya Institute of Technology. Downloaded on March 26,2025 at 08:30:47 UTC from IEEE Xplore. Restrictions apply.
advancement, the development of an Automated Traffic do this in order to prevent any slowing down of the
Management System (ATMS) that can recognise objects detection speed. Second, training ends up taking a lot of
through the application of a methodology that is based on time and space because the R-CNN design concept
deep learning is one of the most important steps that can requires four steps, each of which must be trained
be taken towards the creation of cities that are smarter, independently. These steps include the candidate box
safer, and more environmentally responsible[5]. The selection and adjustment strategy, the training of the
primary goal of this proposal is to provide the framework convolution neural network, the training of the SVM
for a comprehensive research of this paradigm-shifting classifier, and the training of the final regression
traffic management method. In doing so, the quality of operation. All of these steps need to be trained separately.
life for people who live in urban locations across the Throughout the course of the research, participants had
globe will be improved. This innovative new method to access to both the Fast R-CNN as well as the Faster
traffic regulation is going to be thoroughly analysed, and RCNN. They were able to overcome the limitations of R-
the purpose of this proposal is to provide the groundwork CNN, which included a significant demand on
for doing so. The creation of RCNN (Regions with CNN computational resources and challenging training, by
characteristics), which was an important step in the deleting the candidate box approach and increasing the
process of using CNN to identify targets, was an classifier's training. Because of this, they were able to
important milestone. With the powerful feature extraction circumvent the limitations of R-CNN. They provide a
and classification capabilities of CNN, the Region high degree of detection accuracy and have become the
Proposal methodology may be able to improve target technique of choice in deep learning for detecting targets
identification. The process may be broken down into the because of their effectiveness in this area [7-8].
following four parts for the purpose of making it easier to
grasp. The border regression, feature extraction, II. REVIEW OF LITERATURE
classification, and selection of an acceptable region are Despite having the second-largest road network in the
the four procedures that make up this technique. The world, India is responsible for ten percent of all deaths
method known as "Region Proposal," which simulates that occur on the world's roadways. In addition, there has
sliding windows of varying heights and widths, is often been an increase in the amount of traffic congestion, air
used in the process of extracting areas. Sliding windows pollution, and noise pollution in smart cities as a result of
provide an opportunity to snap photographs of potential the proliferation of various kinds of vehicles, the technical
targets. After the target photographs for the proposal have challenges associated with the technology used to operate
been discovered, they are modified such that it seems as traffic signals, and the poor management of road traffic.
though they were taken from CNN. After selecting a Even though they now rely on time-based methods,
method for classification, the next step is to use bounding- traditional traffic management systems are unable to
box regression so that the precise location of the target handle the complicated flow of traffic that happens at
zone may be determined [6]. intersections. This is the case even when these systems
have been updated. Roadblocks may make travelling more
stressful and cause further delays in addition to increasing
the total number of miles driven, the cost of
transportation, and the amount of pollution produced. We
were able to accomplish this goal by developing a
complete traffic light management system using the most
recent version of YOLO, which was just recently
published. In terms of both speed and accuracy, the most
FIGURE: 1AN EXAMPLE OF THE R-CNN DETECTION current iteration of the YOLO algorithm, which goes by
TECHNIQUE the name V7, outperforms all of the prior object
In light of the fact that the actual aim will result recognition models that have been developed. Because it
in the formation of a great deal of sub-regions, the is the fastest and most accurate real-time object
objective is to zero in on the future target for completing identification model, it is the best way to use in traffic
the classification as precisely as is humanly feasible. Bits management systems. As a result, it is the best approach
were concatenated so as to avoid repeated detections of to use. Therefore, doing so is the most prudent thing to do
the same pattern. The R-CNN technique, which is simple in this situation. Yolo V7 is more than 120% faster than
to implement, substitutes the artificial feature extraction any of the previous incarnations, and it also has the
process with a trained convolutional neural network, highest speed to accuracy ratio of any of the versions [9].
which results in an instant improvement of more than Accidents on the roads and other forms of traffic risk
10% in detection accuracy. The disadvantage of R-CNN are occurring increasingly often, which is resulting in a
may also be recognised in a rather apparent manner. There substantial loss of life and property. This problem has an
is a significant amount of calculation required right off the effect on many facets of modern life, including the
bat. When doing a selective search on an input picture, it amount of time spent, the frequency of automobile
is not uncommon for the number of legal candidate boxes accidents, the quantity of greenhouse gas emissions, and
to exceed 1000 throughout the course of the process. In the development of the economy, to name just a few of
order to function properly, a convolutional neural network them. In order to find a solution to this problem, the
has to carry out over a thousand forward operations. This objective of this project is to develop a traffic
process requires considerable patience. It is necessary to management system (TMS) that is based on machine

Authorized licensed use limited to: Yenepoya Institute of Technology. Downloaded on March 26,2025 at 08:30:47 UTC from IEEE Xplore. Restrictions apply.
learning. The vast majority of collisions are brought about III. USING DEEP LEARNING FOR OBJECT DETECTION
because drivers were going faster than the posted speed IN AN AUTONOMOUS TRAFFIC COUNTING SYSTEM
limit in their vehicles. When people are in a hurry to get Both the building of new infrastructure and the
to work on time or to any other appointment on time, they design of new urban areas need efficient traffic
often drive at speeds that are in excess of the legal limit, management as a prerequisite. As city populations grow
which may lead to accidents. Because of this, the project and road networks become more congested, there has
also contains a component that makes use of OpenCV and never been a time when there was a greater need for
image processing and has the ability to calculate the pace efficient traffic monitoring and management systems that
at which moving objects are travelling [10]. are driven by data. Traditional methods of traffic
Even if only a small portion of the traffic at an urban monitoring, which often include human counting or
crossroads actually moves when the light turns green, simple sensor technologies, are unable to provide
long wait times are still a regular problem at these comprehensive data in real time, which renders them
crossings. Within the scope of this research project, we unsuitable for the control of traffic flow. There is a
provide a machine-vision control mode approach that considerable amount of interest in establishing an
incorporates traffic signal management [11]. The Automatic Traffic Measurement System (ATMS)
approach that has been suggested as a result of this employing cutting-edge technology, notably identification
investigation comprises the use of a monitoring camera in of objects based on deep learning. This is being done so
order to collect real-time video footage of the road, the that these problems may be resolved. This ground-
counting of the number of automobiles at different points, breaking technology makes use of cutting-edge computer
and the utilisation of dynamic regulation of the traffic vision algorithms to automatically detect, classify, and
light. Because of our real-time technology, traditional track a variety of relevant roadside items, such as
traffic bottlenecks are less likely to occur, and cars will be automobiles, people, and other relevant things. This
able to cross the intersection without encountering any includes the ability to monitor these objects in real time.
difficulties [12]. Deep learning algorithms and high-resolution sensors may
One of the most fundamentally significant aspects of be used by this system to collect real-time, granular data
the design, construction, and management of highway on the state of the traffic, which will allow for more
networks is the estimation of macroscopically scaled efficient and well-informed decision-making. Figure 2
traffic flow characteristics. This study presents MOM-DL, shows the available roadway configuration with its four
an innovative new approach for the autonomous lanes. In order to establish lanes, the PAGE programme
collection of traffic data. MOM-DL was developed by the will need to be utilised. The video captured by the traffic
authors of this paper. This strategy makes use of the cameras in each lane will be shown. The model will keep
moving observer technique (MOT), as well as the track of how many vehicles are occupying each lane of
YOLOv3 algorithm and deep learning. The MOT is also the roadway at any one time.
often referred to as the moving observer method (MOT)
[13]. The proposed method is able to recognise
automobiles automatically inside a traffic stream and
estimate the traffic variables flow q, space mean speed vs,
and vehicle density k for highways when the traffic
conditions are stationary and uniform. The MOM-DL
method was first used by engineers on a stretch of an
Italian autoroute. Throughout the whole of these tests, a
survey vehicle equipped with a camera has been used.
Using deep learning and YOLOv3, a count was taken of
the number of autos that were parked on the street, which
was the focus of the investigation. The Wardrop
associations were used in order to extract the factors that
were relevant to the flow of traffic. The MOM method
and the MOM-DL approach are highly comparable,
FIGURE 2: REAL-TIME TRAFFIC CONTROL SYSTEM
according to the first results, which reveal that they are in
great agreement with one another. Despite the fact that The model, as we can see, constantly monitors
MOM-DL made a few errors during the vehicle all of the lanes and utilises the information that it collects
identification stage due to a number of causes, the MOM to determine which kinds of cars are located in each one.
method and the MOM-DL approach are quite similar. The The YOLO algorithm is helpful in accomplishing this
values of the macroscopically relevant traffic variables goal. In the course of this investigation, the YOLO
that were derived using the Drakes' traffic flow model and algorithm is put to use in order to recognise different
the recommended approach (MOM-DL) are pretty kinds of vehicles and place them into the appropriate
comparable to those that were obtained from the standard category. These categories include automobiles,
method (MOM). This is due to the fact that the largest motorcycles, buses, and trucks, among others.
possible percentage change is less than 3% [14].

Authorized licensed use limited to: Yenepoya Institute of Technology. Downloaded on March 26,2025 at 08:30:47 UTC from IEEE Xplore. Restrictions apply.
 FIGURE 3: KEEPING AN EYE OUT FOR RESCUE SQUADS

The picture that you see above depicts what is known matter. The collection has photographs that were taken by
as an ambulance, a kind of vehicle that is used in times of a drone as it flew rather low above the busy crossroads.
emergency. As a consequence of this, the length of the The dataset, which included a total of 335 photos, was
signal will immediately become shorter once it recognises partitioned into a train set and a valid set. In contrast, the
the presence of an ambulance. Utilising the Alex Net idea valid set only features a total of sixty photos, while the
allowed for the successful completion of this mission. In training set has a whopping two hundred and seventy-five
order to train the Alex Net model, a significant number of different ones. The fact that there are a lot of different
pictures of a variety of different types of emergency road segments inside the scene is the first problem that
vehicles were employed. One of them consists of the arises with dense estimations that are based on vehicle
ambulance as well. counts. There are further automobiles that vary in size
from very little to very large. Because of this, detection
methods that exclusively rely on counting might
sometimes provide conclusions that are not accurate. The
original YOLO model, which was trained on the COCO
dataset, does not perform well on this dataset since there
is a lot of diversity in the view angles. This dataset was
used to train the model.

FIGURE 4: THE SCIENCE OF DETERMINING HOW FAST CARS

GO
It is feasible to determine the speed of moving
autos on a road by using OpenCV in conjunction with a
deep learning model that has previously been trained.
Always ensure that the camera that was used to capture FIGURE 5: THE TRAFFIC LIGHTS SHOULD BE MANAGED
the film is pointing in a direction that is parallel to the USING A FINITE STATE MACHINE.
road. The speed of the vehicle may be calculated by first
When heading from north to south on State 1, the
finding the distance between each set of waypoints and
traffic light is red, however when moving from east to
then multiplying that distance by the length of time it
west on State 1, it is green. In state 2, the traffic signal
takes for the vehicle to go that distance. The timestamps
shows a red light for drivers heading east to west and a
that are generated by a vehicle are collected by the code at
green light for those heading north to south. If you are
each of the four waypoints (A, B, C, and D). The
heading east to west on State 3, the traffic light will be
anticipated speeds are converted from miles per second to
green for you, but it will be red if you are travelling north
kilometres per hour using an average computation, and
to south. When travelling east to west across state 4, the
then they are converted back to miles per hour. The code
traffic light is red, however when travelling north to south
displays the amount of time that has elapsed, an estimate
it is green. The traffic light heading east to west in state 5
of how many frames are gathered every second, the speed
is coloured yellow. The traffic signal heading east-west on
of the vehicle that has most recently gone by, and the
State 6 is now red, while the traffic signal heading north-
amount of time that has elapsed.
south is currently yellow.
IV. PERFORMANCE EVALUATION
Our model is evaluated using a dataset that
includes information about aerial vehicles as its subject

Authorized licensed use limited to: Yenepoya Institute of Technology. Downloaded on March 26,2025 at 08:30:47 UTC from IEEE Xplore. Restrictions apply.
FIGURE 6: OUR ARCHITECTURE'S TINY YOLOV3 TRAINING OUTCOMES. FOR EACH CASE, 1200 EPOCHS WERE USED IN THE TRAINING
OF THE NETWORK. THE UPDATED LAYOUT PROMOTED HIGHER EFFICIENCY AND A QUICKER LEARNING CURVE.

In this section, we compare the proposed architecture decided to go with the model that had the highest amount
to a model called tiny-YOLOv3 that was trained with the of intersection over union (IoU) when compared to the
same dataset and number of iterations. The strategy that actual data. In YOLO, the amount of loss is computed by
we recommend leads to a reduction in the amount of time using the sum-squared error of the difference between the
spent training while at the same time improving predictions and the actual data. The classification loss, the
performance. Keep in mind that the quality of the road localization loss (errors between the predicted boundary
density estimate has a significant association with the box and the actual boundary box), and the confidence loss
detection accuracy as well as the recall of the information. make up the loss function. These three components work
You really must keep this in mind moving forward. As a together to form the loss. - A decrease in classification If
direct consequence of this, we restrict the scope of our an item is found, the squared error of the class conditional
investigation of detection to a basic comparison. In order probability is equal to the classification loss at each cell.
to conduct a more in-depth analysis of the counting This is the case regardless of which class is being
process, we make use of the video recordings that were considered. This takes occur after a location for an item
created by the public security cameras known as has been determined.
EarthCam and Cerevo live camera. Since the majority of
cameras do not cover the whole path of the junction, the
output from numerous cameras may be utilised as the
input to the traffic management system. One illustration If there is an object in cell I, then the value of
of this may be seen in the camera that is installed on Iijobj will be 1, but if there isn't, it will be 0. This is
Route 66. Before we can make an accurate estimate of the because p(c) represents the conditional class probability
flow of traffic, we need to begin by dividing the route into for class c in cell i. - analyses the deviations that exist
pieces of about the same size. between the predicted border box sizes and positions and
the actual border box sizes and positions. This component
V. RESEARCH METHODOLOGY of the loss of localisation. It turned out that the only box
The study strategy to construct the Automated that was counted was the one in which the item had been
Traffic Management System (ATMS) that employs object found. You may determine the size of the bounding box in
identification for deep learning requires a methodical YOLO v3 by using the following equation:
approach and several essential phases. The system would
be built this way. First, project objectives, parameters, and
traffic management challenges will be determined. These
include traffic congestion, accident prevention, and flow
improvement. A detailed literature review will reveal the
latest advances in deep learning-based object detection for
traffic control. A comprehensive traffic dataset will
follow. This collection will feature photographs and where the model's projected outputs, tx, ty, tw, and
videos from many urban areas to accurately reflect th, stand in for the relative position coordinates of the
various events. We will build a hardware and software centre and the width and height of the enclosing box,
infrastructure for real-time data collection, model respectively. The width and height of the predicted front
deployment, and deep learning model selection and fine- bounding box are denoted by the characters pw and ph,
tuning for object recognition. Building it will support this respectively, while the net is denoted by the symbols cx
infrastructure. and cy. Last but not least, the real width, height, and
 Loss Calculation prediction-based centre coordinates of the bounding box
YOLO will generate several bounding boxes for each are given by the variables bx, by, bw, and bh respectively.
grid cell that will be utilised in the research. In order to These numbers are located in the very last column of the
compute the percentage of the total gain that is very last row in the table.
attributable to the loss, one of them will need to be in
control of the scenario. As a consequence of this, we

Authorized licensed use limited to: Yenepoya Institute of Technology. Downloaded on March 26,2025 at 08:30:47 UTC from IEEE Xplore. Restrictions apply.
 VI. ANALYSIS AND INTERPRETATION 6 151 158 4.60%
During the analysis phase of the Automated Traffic 7 218 212 3.20%
Management System (ATMS), which uses deep learning- 8 240 235 3.50%
based object recognition, a number of intriguing 9 225 227 3.00%
discoveries and observations have been obtained. Deep 10 258 260 0.50%
learning models for object recognition have shown
11 330 325 2.10%
exceptional accuracy and efficiency in detecting and
12 292 280 3.70%
tracking automobiles, pedestrians, and other relevant
items in real-time traffic scenarios. This has been possible Total 2263 2233 49.40%
because to these models' use of deep learning. The Average of Hourly Errors = 49.40%
employment of these models allows for the recognition The computer once again recorded 92 in the
and tracking of many entities. Because of this, the system second hour, which was 4.40% more than the human
is now able to deliver information that is both accurate count of 88. The algorithm captured 92 in the first hour.
and up to date on the present state of the traffic. This The difference between the 108 numbers that were
enables the organisations that are responsible for traffic manually counted and the 103 numbers that the automatic
management to make choices that are better informed. counter recorded during the third hour was 5.60 percent.
The approach that was used to evaluate the data has also The computer's count of 146 in the fourth hour showed a
shown substantial breakthroughs in the optimisation of 6.50 percent accuracy disadvantage when compared to the
traffic flow. Utilising the capabilities of object human count of 141. The system displayed a count of 123
identification, the ATMS was able to correctly identify at the end of the fifth hour, which was 8.80 percentage
locations that were crowded. Because of this, it was points higher than the actual count of 112 that had been
feasible to control lanes and adjust the timing of the traffic maintained on the ground. The system's count was 151 in
signals in a timely way. Because of this, there is much the sixth hour, which was 4.60 percent less than the
less traffic on the roads, which leads to a flow of traffic humans' total of 158. The machine-generated count for the
that is both less unpredictable and more efficient in the seventh hour, which was 218 instead of the human count
regions that are being watched. Table 1 presents the of 212, was wrong by 3.20%. The system tallied 240 in
results of a comparison between the counts that were the seventh hour, a 3.50 percentage point gain over the
generated by an automatic counter that was developed by human count of 235. The automated counter read 225 in
a machine and the counts that were manually conducted the ninth hour, which was three percent lower than the
by an individual over the course of a period of twelve human count of 227 kept. The machine's tally of 258 in
hours. The table not only provides an interesting overview the tenth hour, as opposed to the human count of 260,
of both, but it also identifies the errors that occurred as a produced a minimal inaccuracy of only 0.50%. The
direct consequence of the functioning of the system. computer generated a final total of 330, with an accuracy
Within the first hour, the algorithm came up with a count error of 2.10% when compared to the human count of
of 85, which was 3.50 percent lower than the real figure 325. The system's ultimate count after the first twelve
of 87. hours was 292, which was 3.70 percentage points more
TABLE 1: THE POTENTIAL FOR HOURLY INACCURACIES IN
AN AUTOMATED COUNTING SYSTEM IS EXPLORED. than the human total of 280.On the whole Comparing the
Automated Ground automated system's 12-hour total count of 2,263 to the
Hour Counter by Truth Hourly human-completed ground truth count of 2,233, the
No Developed Counted by Error cumulative error rate was 49.40%. Table 2 shows that
system Humam
"Our architecture" significantly outperformed all other
1 85 87 3.50% models in terms of accuracy, recall, F1 score, and
2 92 88 4.40% evaluation metrics as a whole. Compared to the other
3 103 108 5.60% models, there is a sharp contrast. It is the model of
4 146 141 6.50% preference for issues concerning object identification as a
5 123 112 8.80%
result.

TABLE 2: OBJECT DETECTION MODEL EVALUATION

Class
Precision Recall Map F1 GloU objective
Roll
Pretrained
tiny- <0.01 <0.01 <0.01 0 6.7 3.21 7.26
YOLOv3
tiny-
0.654 0.821 0.832 0.71 1.19 0.51 0.102
YOLOv3
Our
0.803 0.912 0.902 0.86 1.35 0.45 0.907
architecture
The pre-trained tiny-YOLOv3 model receives an demonstrates that accurately identifying items is a
F1 score of 0 since its accuracy and recall are both lower challenging issue to deal with. In spite of the fact that the
than 0.01, making it the model with the lowest Global Utility (GloU) is relatively high, coming in at 6.7,
performance of the three. This model has the worst both the objective score and the class roll are low, which
performance in terms of its overall effectiveness. This implies that performance was below average. Tiny-

Authorized licensed use limited to: Yenepoya Institute of Technology. Downloaded on March 26,2025 at 08:30:47 UTC from IEEE Xplore. Restrictions apply.
YOLOv3, the second model, has performance that is performance in terms of recall and accuracy. It is
much superior than the pretrained version. It's possible considered to have a moderate level of utility when its
that an F1 score of 0.71 will be achieved if the accuracy is Global Utility (GloU) is 1.35, which indicates that it may
0.654 and the recall is 0.821. This suggests that the trade- be used in a number of settings. It had a remarkable
offs between accuracy and recall are being maintained at a overall performance, earning it a high objective score of
satisfactory level. Given that the model has a decent mean 0.45, which demonstrates that it performed extremely
Average Precision (mAP) of 0.832, it is possible for it to well. Another reason for concern is that there are only 907
succeed in a range of item recognition tasks. This is students registered for the session. In the following table,
proved by the fact that the model has this value. A lower you will find a detailed comparison of each significant
objective score of 0.51 was given to the pre-trained traffic flow measure for the two different approaches that
model, and the GloU was only 1.19, which is not an have been used. The terms "Method of Moments"
especially impressive value. 0.102 is the lowest possible (sometimes shortened as "MOM") and "Method of
class roll. The accuracy and recall scores of our technique Moments with Deep Learning" (often abbreviated as
are 0.803 and 0.912, respectively, which sum up to an "MOM-DL") refer to methods that are comparable to one
astounding F1 score of 0.86. These values demonstrate another. These components are necessary in order to
that our method is superior to both of the earlier models. understand and analyse the capacity of a route, as well as
The fact that the model's mean accuracy proportion the behaviour of the traffic along that route.
(mAP) is 0.902 demonstrates the model's overall high
TABLE 3: COMPARISON OF VARIOUS MEASURES OF TRAFFIC FLOW.
Free-Flow Lane Critical
Critical Speed
Method Speed Vf Capacity c Dencity Kc
Vc [Km/h]
[Km/h] [pcu/h] [pcu/Km]

MOM 71.08 1243.14 44 29.87

MOM-DL 71.83 1256.77 43 28.87
∆[%] 1 1.1 2.5 0.1
strategy for getting inside This column explains difference (%) between each parameter configuration. The
both "MOM" and "MOM-DL," which are the two critical density for "MOM" is 29.87 pcu/km, however the
methodologies for analysing traffic flow that are being critical density for "MOM-DL" is just 28.87 pcu/km,
contrasted with one another. (in km per hour) Vf which indicates that congestion starts at a somewhat lower
Unhindered Flow The term "speed" relates to the vehicle density for "MOM-DL." The percentage changes
following: It is common for automobiles to go along a in these parameters, which range from 0.1% all the way
road at a free-flowing speed, which is measured in up to 2.5%, suggest that "MOM-DL" often displays
kilometres per hour (Km/h) and occurs when there is very improvements.
little to no traffic congestion. According to the
information in the table, the performance of "MOM-DL" VII. RESULT AND DISCUSSION
is 1% better than that of "MOM" in terms of the free-flow Positive outcomes have been seen as a consequence
speed. The capacity of the lane is denoted by the symbol of the application of a deep learning strategy to the task of
c, which stands for "pcu/h." The number of passenger object identification inside an automated traffic control
vehicle units that may be moved through a lane in one system. The method's high degree of accuracy was proved
hour is denoted with the notation pcu/h (passenger car by metrics such as mean average precision (mAP) scores
units per hour), which stands for "passenger car units per that were consistently higher than 90% and intersection
hour." According to the findings of this study, "MOM- over union (IoU) scores that were higher than 0.7. The
DL" is capable of accommodating 1.1% more traffic lanes ability of the system to manage real-time video streams at
than "MOM." Vc is the abbreviation for the critical speed, frame rates that are higher than 30 frames per second is
which is expressed in kilometres per hour. The critical another indication of the utility of the system in busy
speed, which is measured in kilometres per hour (Km/h), environments. The flexibility of the model to changing
is the velocity at which one may hypothesise that the environmental factors such as weather, traffic, and
onset of traffic congestion occurred. With a higher critical illumination further highlighted the model's endurance.
speed, there is a larger possibility that there will be more The system was fairly good at recognising times of
traffic flow before congestion begins. The critical speed congestion and forecasting the length of time that
of the vehicle designated as "MOM" is 44 kilometres per passengers would have to wait, which resulted in an
hour, but the critical speed of the vehicle designated as improvement in traffic flow and a large drop in the
"MOM-DL" is 43 kilometres per hour. The critical density frequency of traffic jams. The outcome was an overall
is represented by the symbol Kc, which is written as reduction in the amount of time spent waiting in traffic.
pcu/Km. When there are a particular number of The accurate detection of individuals at crosswalks and
automobiles per km, the likelihood of their being traffic the identification of obstacles on routes that are used
backups increases. This particular value has been given mostly by pedestrians made it feasible to achieve a
the label of the critical density. This row provides a significant improvement in the safety of pedestrians.
straightforward way to compare the results of "MOM" Figure 7 provides a concise review of the statistical
with "MOM-DL" by determining the percentage factors that contribute to the efficiency of an automated

Authorized licensed use limited to: Yenepoya Institute of Technology. Downloaded on March 26,2025 at 08:30:47 UTC from IEEE Xplore. Restrictions apply.
counting system in comparison to human ground truth Developed System: In the next column, the results of the
counts. This table contains the statistics for any other automated counting system, which had to be developed
relevant components as well as other important specifically for the task at hand, are shown. The counts
information such as the mean, standard deviation, that were gathered via the use of manual counting, which
minimum, and maximum values for each key parameter. was carried out by a human operator, are shown in this
In addition, this table contains the statistics for any other column as the benchmark or point of reference for
relevant components. The Computerised Counter of the determining the degree of accuracy that was achieved.

FIGURE 7: ERRORS IN AN AUTOMATED COUNTING SYSTEM: AN HOURLY STATISTICAL ANALYSIS

Every Day, Errors are Committed in In order to how closely the data and the model matched up with one
calculate the hourly error, the counts that were generated another. These tests are of the utmost significance in order
by the automated system are contrasted with the counts to ascertain whether or not the data correspond to a certain
that were determined using the ground truth. After then, theoretical model or probability distribution. The table
the variance is shown in the form of a percentage for each presents the results of four different statistical tests,
hour. The following table provides a statistical overview including the Shapiro-Wilk test, the Anderson-Darling
of the hourly error levels that were measured. I mean this test, the Kolmogorov-Smirnov test, and the Kolmogorov-
by that. The mean (or average) number is provided for Smirnov test with the Lilliefors adjustment.
both the automated counts and the counts obtained using
ground truth. The average number obtained from the
Automated Counter is 348.15, whereas the average
number obtained from the Ground Truth Count is 343.54.
The difference lies in the fact that the standard deviation
may be used as a measurement tool to determine whether
the data are uniformly distributed or randomly scattered.
When compared to the counts based on the ground truth,
the automated counts have a standard deviation of 580.76,
while the counts based on the ground truth have a
standard deviation of 572.98. These values illustrate the
degree to which the counts stray from the mean. There is
an illustration of the minimal values for both the
automated counts and the ground truth counts. The
Automated Counter recorded a value of 85 during the FIGURE 8: TESTS OF PHYSICAL FITNESS AND ITS OUTCOMES
period of time that was being examined, but the Ground The p-values for each test are listed in the table's
Truth Count reported a value of 87 for the same time second column, which can be found in the table. If the
period. These values reflect the lowest counts that were data follow the hypothesised distribution—which is what
ever reported during those eras. Both the automated is expected to be the case—then these p-values imply that
counts and the ground truth counts indicate their upper it is feasible to construct test statistics that are just as
boundaries in their respective descriptions. This is true for severe as those that were observed. This is because it
both the automated counts and the ground truth counts. suggests that it is possible to make test statistics that are
The Automated Counter reached its maximum possible just as severe as those that were seen. If the p-value is low
value of 2263, while the Ground Truth Count reached its (usually less than 0.05), there is sufficient evidence to
highest possible value of 223. The statistics that are support the conclusion that the data does not follow the
provided here are a representation of the hours during distribution that was specified as the null hypothesis. In
which the greatest counts that were witnessed were this particular scenario, the p-values produced by each of
recorded. Figure 8 depicts the findings of many the four tests were very low, with the Kolmogorov-
evaluations of a dataset's goodness-of-fit that were carried Smirnov test producing the highest result of the bunch at
out on the data. These assessments were carried out to see 0.44. The results suggest that the data are significantly
different from the distribution that was anticipated, which

Authorized licensed use limited to: Yenepoya Institute of Technology. Downloaded on March 26,2025 at 08:30:47 UTC from IEEE Xplore. Restrictions apply.
leads to the conclusion that the null hypothesis need to be pedestrians, and other objects under real-time traffic
rejected because of this finding. Researchers may need to circumstances. As a result, it has provided accurate
take into consideration a variety of distributional theories information that can be used in an effective manner to the
or do further study on the underlying data in order to management of traffic. The unexpected benefits of this
completely appreciate the properties of the data. new technique include a reduction in the amount of traffic
Alternatively, the researchers may need to investigate the congestion, an increase in the safety of pedestrians, and an
data further. Several goodness-of-fit tests were executed improvement in the flow of traffic. The system's
on a dataset in order to determine how closely it flexibility to different climatic circumstances and its
corresponds to a certain probability distribution or capability to handle video inputs in real time are two
theoretical model. This was done in order to evaluate the additional aspects that highlight the system's utility and
accuracy of the dataset. Figure 9 illustrates the results of efficiency. Both of these traits are highlighted below.
the study. These tests were carried out to see how well the Despite the fact that the research found a number of
dataset and the model correspond to one another. The challenges and deviations from the distributions that were
table presents the results of four different statistical tests, anticipated, it is hard to deny that the ATMS has had a
including the Shapiro-Wilk test, the Anderson-Darling positive overall impact on the management of urban
test, the Kolmogorov-Smirnov test, and the Kolmogorov- traffic. This is an indisputable fact. As a consequence of
Smirnov test with the Lilliefors adjustment. this, we are in a position to assert that the proposed
YOLO V7 architectural traffic light system is capable of
controlling and regulating the flow of traffic with a degree
of precision that is equivalent to 92%. It has been shown
that the YOLO V7 works better than all prior generations
of the item detector in any scenario that may possibly
arise. In comparison to other object detector models, this
one has a speed range that starts at 5 frames per second
and goes all the way up to 160 frames per second, all
while maintaining the highest possible degree of accuracy.
As a direct result of this, there is the possibility that it will
be used in the years to come. 56% of the data was
employed for testing and training purposes in the YOLO
v7 system. In most cases, the detection of vehicles is
FIGURE 9: ERROR TEST RESULTS FROM AUTOMATED accurate to 92%.
TRAFFIC MANAGEMENT SYSTEMS PROVIDE STATISTICAL
DATA. REFERENCES
The p-values for each of these tests may be found 1. S. A. Meshram and R. S. Lande, "Traffic surveillance by using
in the second column of the table, which can be seen image processing," 2018 International Conference on Research in
above. These p-values show that it is possible to produce Intelligent and Computing in Engineering (RICE), 2018, pp. 1-3,
test statistics that are just as severe as those that were doi: 10.1109/RICE.2018.
2. Khushi, "Smart Control of Traffic Light System using Image
observed, assuming that the data follow the recommended Processing," 2017 International Conference on Current Trends in
distribution. This is based on the assumption that the data Computer, Electrical, Electronics and Communication (CTCEEC),
indeed match the suggested distribution. A p-value that is 2017, pp. 99-103, doi: 10.1109/CTCEEC.2017.
low (usually less than 0.05) suggests that there is 3. Ansari, K.R., Quraishi, M.A., Singh, A., “Isatin derivatives as a
non-toxic corrosion inhibitor for mild steel in 20% H2SO4.”
significant evidence against the null hypothesis and also Corrosion Science 95 (2015), pp.62-70.doi:
signals a divergence from the distribution that would be 10.1016/j.corsci.2015.02.010
anticipated. In this particular scenario, the Shapiro-Wilk 4. Panda, S.K., Aggarwal, I., Kumar, H., Prasad, L., Kumar, A.,
test has the highest p-value out of the four available Sharma, A., Vo, D.-V.N., Van Thuan, D., Mishra, V. (2021),
“Magnetite nanoparticles as sorbents for dye removal: a review”,
options, coming in at 0.54. Due to the fact that these Environmental Chemistry Letters 19(3), pp. 2487-2525. doi:
results suggest that the data considerably differ from the 10.1007/s10311-020-01173-9
distribution that was predicted, we are justified in drawing 5. Yuvaraj, N., Srihari, K., Dhiman, G., Somasundaram, K., Sharma,
the conclusion that the null hypothesis should not be A., Rajeskannan, S., Soni, M., Gaba, G.S., Alzain, M.A., Masud,
M.. “Nature-Inspired-Based Approach for Automated
accepted. It's possible that researchers may need to look at Cyberbullying Classification on Multimedia Social Networking.”
a number of different distributional theories or do further Mathematical Problems in Engineering, 2021(2021), Doi:
study in order to acquire a better grasp of the properties of 10.1155/2021/6644652
the data that lay under the surface. 6. A. Menon and B. Omman, "Detection and Recognition of Multiple
License Plate From Still Images," 2018 International Conference
VIII. CONCLUSIONS on Circuits and Systems in Digital Enterprise Technology
(ICCSDET), 2018, pp. 1-5, doi: 10.1109/ICCSDET.2018.
The incorporation of a deep learning-based 7. S. M. Sunny, T. Rahman, S. M. Zohurul Islam, A. Mujtaba, K. F.
object identification strategy into an Automated Traffic Ahmed and S. Saha, "Image Based Automatic Traffic Surveillance
System Through Number-Plate identification And Accident
Management System (ATMS) has shown significant Detection," 2021 2nd International Conference on Robotics,
promise as a means of resolving the problems that are Electrical and Signal Processing Techniques (ICREST), 2021, pp.
now evident in urban traffic. The system has been found 467-47doi: 10.1109/ICREST51555.2021.
to have a high level of accuracy when it comes to 8. Prakash, C., Uddin, M.S.,”Surface modification of β-phase Ti
implant by hydroaxyapatite mixed electric discharge machining to
recognising and keeping track of automobiles, enhance the corrosion resistance and in-vitro bioactivity”, Surface

Authorized licensed use limited to: Yenepoya Institute of Technology. Downloaded on March 26,2025 at 08:30:47 UTC from IEEE Xplore. Restrictions apply.
 and Coatings Technology (2017), 326, pp. 134-145. doi: Edge Computing Techniques for Industrial Applications," 2023 3rd
10.1016/j.surfcoat.2017.07.040 International Conference on Pervasive Computing and Social
9. N. P. Ap, T. Vigneshwaran, M. S. Arappradhan and R. Madhanraj, Networking (ICPCSN), Salem, India, 2023, pp. 1280-1285, doi:
"Automatic Number Plate Detection in Vehicles using Faster R- 10.1109/ICPCSN58827.2023.00215.
CNN," 2020 International Conference on System, Computation, 21. Bani Ahmad, A. Y. A. ., William, P. ., Uike, D. ., Murgai, A. .,
Automation and Networking (ICSCAN), 2020, pp. 1-6, doi: Bajaj, K. K. ., Deepak, A. ., & Shrivastava, A. . (2023). Framework
10.1109/ICSCAN49426.2020.9262400. for Sustainable Energy Management using Smart Grid Panels
10. M. M. Gandhi, D. S. Solanki, R. S. Daptardar and N. S. Baloorkar, Integrated with Machine Learning and IOT based
"Smart Control of Traffic Light Using Artificial Intelligence," 2020 Approach. International Journal of Intelligent Systems and
5th IEEE International Conference on Recent Advances and Applications in Engineering, 12(2s), 581–590. Retrieved from
Innovations in Engineering (ICRAIE), 2020, pp. 1-6, doi: https://ijisae.org/index.php/IJISAE/article/view/3670
10.1109/ICRAIE51050.2020.9358334. 22. Deepak, A. ., William, P. ., Dubey, R. ., Sachdeva, S. ., Vinotha, C.
11. M. Prakash, C. Saravanakumar, S. K. Lakshmi, J. D. Rose and B. ., Masand, S. ., & Shrivastava, A. . (2023). Impact of Artificial
Praba, "Automatic Feature Extraction and Traffic Management Intelligence and Cyber Security as Advanced Technologies on
Using Machine Learning and Open CV Model," 2021 International Bitcoin Industries. International Journal of Intelligent Systems and
Conference on Innovative Computing, Intelligent Communication Applications in Engineering, 12(3s), 131–140. Retrieved from
and Smart Electrical Systems (ICSES), 2021, pp. 1-5. https://ijisae.org/index.php/IJISAE/article/view/3669
12. M. I. Uddin, M. S. Alamgir, M. M. Rahman, M. S. Bhuiyan and M.
23. William, P. ., Bani Ahmad, A. Y. A. ., Deepak, A. ., Gupta, R. .,
A. Moral, "AI Traffic Control System Based on Deepstream and Bajaj, K. K. ., & Deshmukh, R. . (2023). Sustainable
IoT Using NVIDIA Jetson Nano," 2021 2nd International Implementation of Artificial Intelligence Based Decision Support
Conference on Robotics, Electrical and Signal Processing System for Irrigation Projects in the Development of Rural
Techniques (ICREST), 2021, pp. 115- 119, doi: Settlements. International Journal of Intelligent Systems and
10.1109/ICREST51555.2021.9331256. Applications in Engineering, 12(3s), 48–56. Retrieved from
13. J. F. Gilmore and K. J. Elibiar, "Traffic Management Applications https://ijisae.org/index.php/IJISAE/article/view/3660
of Neural Networks," in Proceedings of the IEEE International
Conference on Neural Networks (ICNN), 2022. 24. D. R., G. ., P., Biradar, V. S. ., M., V. ., Singh, C. ., Deepak, A. .,
& Shrivastava, A. . (2023). Energy-Efficient Resource Allocation
14. Bi, F.; Yang, J. Target Detection System Design and FPGA
and Relay-Selection for Wireless Sensor Networks. International
Implementation Based on YOLO v2 Algorithm, International
Journal of Intelligent Systems and Applications in
Conference on Imaging, Signal Processing and Communication.
Engineering, 12(5s), 113–121. Retrieved from
ICISPC 2019, 10–14.
https://ijisae.org/index.php/IJISAE/article/view/3871
15. P. William, M. Gupta, N. Chinthamu, A. Shrivastava, I. Kumar and
A. K. Rao, "Novel Approach for Software Reliability Analysis 25. Venkatesh, S. ., Kori, S. P. ., William, P. ., Meena, M. L. .,
Controlled with Multifunctional Machine Learning Approach," Deepak, A. ., Hasan, D. S. ., & Shrivastava, A. . (2023). Data
2023 4th International Conference on Electronics and Sustainable Reduction Techniques in Wireless Sensor Networks with Internet
Communication Systems (ICESC), Coimbatore, India, 2023, pp. of Things. International Journal of Intelligent Systems and
1445-1450, doi: 10.1109/ICESC57686.2023.10193348. Applications in Engineering, 12(8s), 81–92. Retrieved from
https://ijisae.org/index.php/IJISAE/article/view/4098
16. Kumar, A., More, C., Shinde, N. K., Muralidhar, N. V.,
Shrivastava, A., Reddy, C. V. K., & William, P. (2023). 26. P. William, Poornashankar, A. Shrivastava, N. Tripathi, Anil and
Distributed Electromagnetic Radiation Based Renewable Energy A. Singh, "Secure Authentication Protocols For Internet Of Things
Assessment Using Novel Ensembling Approach. Journal of Nano- (Iot) Devices," 2023 6th International Conference on
and Electronic Physics, 15(4). Contemporary Computing and Informatics (IC3I), Gautam Buddha
Nagar, India, 2023, pp. 1569-1574, doi:
17. P. William, O. J. Oyebode, G. Ramu, M. Gupta, D. Bordoloi and 10.1109/IC3I59117.2023.10397626.
A. Shrivastava, "Artificial Intelligence based Models to Support
Water Quality Prediction using Machine Learning Approach," 27. P. William, A. K. Rai, P. Madan, C. P. Kumar, A. Shrivastava and
2023 International Conference on Circuit Power and Computing A. Rana, "Analysis of Blockchain Technology to Protect Data
Technologies (ICCPCT), Kollam, India, 2023, pp. 1496-1501, doi: Access Using Intelligent Contract Mechanism for 5G Networks,"
10.1109/ICCPCT58313.2023.10245020. 2023 6th International Conference on Contemporary Computing
and Informatics (IC3I), Gautam Buddha Nagar, India, 2023, pp.
18. P. William, G. Ramu, L. R. Gupta, P. Sing, A. Shrivastava and A. 1651-1657, doi: 10.1109/IC3I59117.2023.10397700.
P. Srivastava, "Hybrid Temperature and Humidity Monitoring
System using IoT for Smart Garden," 2023 Second International 28. P. William, S. Kumar, A. Gupta, A. Shrivastava, A. L. N. Rao and
Conference on Augmented Intelligence and Sustainable Systems V. Kumar, "Impact of Green Marketing Strategies on Business
(ICAISS), Trichy, India, 2023, pp. 1514-1518, doi: Performance Using Big Data," 2023 4th International Conference
10.1109/ICAISS58487.2023.10250538. on Computation, Automation and Knowledge Management
(ICCAKM), Dubai, United Arab Emirates, 2023, pp. 1-6, doi:
19. P. William, M. G. Yadav, S. Lakhanpal, A. Shrivastava, K. S. 10.1109/ICCAKM58659.2023.10449560.
Rawat and S. Chaudhary, "Artificial Intelligence based
Performance Models to Support Hydrologic Rainfall Conditions 29. P. William, A. Agrawal, N. Rawat, A. Shrivastava, A. P. Srivastava
using Ensembling Approach," 2023 Second International and Ashish, "Enterprise Human Resource Management Model By
Conference on Augmented Intelligence and Sustainable Systems Artificial Intelligence Digital Technology," 2023 4th International
(ICAISS), Trichy, India, 2023, pp. 223-229, doi: Conference on Computation, Automation and Knowledge
10.1109/ICAISS58487.2023.10250561. Management (ICCAKM), Dubai, United Arab Emirates, 2023, pp.
20. P. William, N. Chinthamu, I. Kumar, M. Gupta, A. Shrivastava and 01-06, doi: 10.1109/ICCAKM58659.2023.10449624.
A. P. Srivastava, "Schema Design with Intelligent Multi Modelling

Authorized licensed use limited to: Yenepoya Institute of Technology. Downloaded on March 26,2025 at 08:30:47 UTC from IEEE Xplore. Restrictions apply.