Car Parking Space Detection Using Yolo: A Project Report Submitted by

CAR PARKING SPACE DETECTION
USING YOLO
A PROJECT REPORT
Submitted by
19104047 NELSON J
19104063 SANTHOSH E
19104084 VERJIN V
19104086 VIGNESHWARAN A
in partial fulfilment for the award of the degree
of
BACHELOR OF ENGINEERING
in
COMPUTER SCIENCE AND ENGINEERING
ERODE SENGUNTHAR ENGINEERING COLLEGE

(AUTONOMOUS)
PERUNDURAI, ERODE – 638 057
MARCH 2023
ERODE SENGUNTHAR ENGINEERING COLLEGE
(AUTONOMOUS)
PERUNDURAI, ERODE – 638057
BONAFIDE CERTIFICATE
Certified that this project report “CAR PARKING SPACE DETECTION USING
YOLO” is the bonafide work of NELSON J (19104047), SANTHOSH E
(19104063), VERJIN V (19104084) and VIGNESHWARAN A (19104086) who
carried out the project work under my supervision. Certified further that to the best of
my knowledge the reported here in does not from part of any other thesis or
dissertation on the basis of which a degree or award was conferred on an earlier
occasion on this or any other candidate.
SIGNATURE SIGNATURE
Ms.R.SAVITHA, M.E., Dr.G.SIVAKUMAR, M.E., Ph.D.,
SUPERVISOR HEAD OF THE DEPARTMENT
ASSISTANT PROFESSOR, PROFESSOR AND HEAD,
Department of Artificial Intelligence and Department of Computer Science and
Data Science, Engineering,
Erode Sengunthar Engineering College, Erode Sengunthar Engineering College,
(An Autonomous Institution). (An Autonomous Institution)
Thudupathi, Erode - 638 057. Thudupathi, Erode - 638 057.
Submitted for End Semester Project Viva-Voce Examination held on _________
INTERNAL EXAMINER EXTERNAL EXAMINER

ACKNOWLEDGEMENT
First of all, we would like to convey our heartfelt thanks to our respected Founder,
“UDYOG RATTAN” DEIVATHIRU J. SUDHANANDHEN, B.Sc., who always blessed us to
give the best.
We are delighted to thank our esteemed President Thiru.V.ANNADURAI of Erode

Sengunthar Educational Trust & Correspondent Thiru.G.KAMALAMURUGAN, B.B.M., of
Erode Sengunthar Engineering College and gave useful support for carrying out our project in
adequate time and providing us with facilities.
We have immense pleasure to thank our Secretary

Thiru.S.N.THANGARAJU, B.E., M.B.A., of Erode Sengunthar Educational Trust who gave us
full support and guide in all the endeavours of our project.
We would like to express our thanks to our Principal

Dr.V.VENKATACHALAM, B.E., M.S., M.Tech., Ph.D., for forwarding us to do our project and
offering adequate duration in completing our project.
We render our sincere thanks to, Dr.G. SIVAKUMAR, M.E., Ph.D., Professor and Head,
Department of Computer Science and Engineering for his anchoring support in doing this project.
We are grateful to our project coordinator Dr.S.TAMILSELVAN, M.E., Ph.D., Assistant

Professor for his excellent ideas and encouragement.
We are hearty extend our thanks to our project supervisor Ms.R.SAVITHA, M.E., Assistant
Professor for her support, constant supervision and as well as for providing necessary information
regarding the project.
ABSTRACT
The project aim is to detect the parking spaces of cars using the YOLO (You Only Look Once)
algorithm. The camera footage is utilized to extract images which then undergo object detection to
identify cars in parking spaces. The current approach to detecting cars in parking spaces involves
the use of sensors as well as Internet of Things (IoT) technology and also uses object detection
algorithms such as Support Vector Machine (SVM) and Region-based Convolutional Neural
Networks (R-CNN). However, these methods have been found to be inadequate in terms of
efficiency and accuracy. One of the main challenges associated with these techniques is their high
cost, which can be a significant barrier to widespread adoption. In addition, their installation can be
quite complex, requiring skilled technicians to set up and configure the system correctly.
Furthermore, these methods are often required regular maintenance to ensure that they continue to
function correctly. Despite these challenges, researchers and developers are working to improve
these techniques and overcome these limitations to make car parking space detection more efficient
and accurate. We proposed a solution that aims to overcome these drawbacks by introducing a deep
learning approach that uses the YOLO algorithm. This algorithm is efficient and has proven to be
highly accurate in object detection tasks. The proposed solution is expected to outperform existing
solutions, providing a cost-effective and accurate solution to detect car parking spaces. The
advantages of using YOLO includes improved accuracy, faster processing and reduced
computational costs compared to other algorithms.
Keywords: parking spaces, YOLO algorithm, object detection, camera footage, sensors, IoT, SVM,
R-CNN, accuracy, efficiency, cost-effective, deep learning, computational costs.
i
TABLE OF CONTENTS
CHAPTER PAGE
TITLE
NO. NO.
ABSTRACT i
LIST OF TABLES v
LIST OF FIGURES vi
LIST OF ABBREVIATIONS vii
1 INTRODUCTION 1
2 LITERATURE REVIEW 2
2.1 IoT based Intelligent Parking Management System 2
2.2 Smart Parking System Based on Optical Character 3
Recognition
2.3 Smart Parking Using IoT Technology 5
2.4 A Grid Projection Method Based on Ultrasonic Sensor for 6
Parking Space Detection
2.5 A Review on Automatic Parking Space Occupancy Detection 8
2.6 Mobile Outdoor Parking Space Detection Application 9
2.7 Automatic Parking Space Detection System 11
2.8 Automatic Parking Space Detection and Tracking for 12
Underground and Indoor Environments
2.9 A Convenient Vision-Based System for Automatic Detection 14
of Parking Spaces in Indoor Parking Lots Using Wide-Angle
Cameras
2.10 Parking Space Detection with Hierarchical Dynamic 16
Occupancy Grids
2.11 Summary 17
3 SYSTEM ANALYSIS 18
3.1 Existing System 18
3.1.1 Disadvantages 18
3.2 Proposed System 19
3.2.1 Advantages 20
ii
CHAPTER PAGE
TITLE
NO. NO.
4 SYSTEM REQUIREMENTS 21
4.1 Hardware Requirements 21
4.2 Software Requirements 21
4.3 Tools and Technology 21
4.3.1 Python 21
4.3.2 Pip 22
4.3.3 NumPy 22
4.3.4 Keras 22
4.3.5 OpenCV 22
4.3.6 Jupyter Notebook 22
4.3.7 TensorFlow 22
4.3.8 PyTorch 23
5 SYSTEM DESIGN 24
5.1 System Architecture 24
5.2 List of Modules 25
5.2.1 Pre-Processing 25
5.2.2 Object Detection 25
5.2.3 Post-Processing 25
5.2.4 User Interface 26
5.3 Proposed Methodology 26
5.3.1 Data Collection 26
5.3.2 Image Annotation 27
5.3.3 YOLO Architecture 27
5.3.4 Training and Testing 28
5.3.5 Performance Evaluation Metrics 29
5.4 Workflow 30
5.5 Data-flow 30
5.6 Sequence 31
5.7 Use Case 31
6 RESULT AND ANALYSIS 32
iii
CHAPTER PAGE
TITLE
NO. NO.
7 CONCLUSION AND FUTURE ENHANCEMENT 35
7.1 Conclusion 35
7.2 Future Enhancement 35
APPENDIX 37
I. Source Code 37
II. Screenshots 43
REFERENCES 47
iv
LIST OF TABLES
TABLE PAGE
TABLE NAME
NO. NO.
6.1 Model Performance with Motion Detection Results 33
6.2 Model Performance with No Motion Detection Results 33
v
LIST OF FIGURES
FIGURE PAGE
FIGURE NAME
NO. NO.
5.1 System Architecture 24
5.2 Workflow Diagram 30
5.3 Data-flow Diagram 30
5.4 Sequence Diagram 31
5.5 Use Case Diagram 31
6.1 Evaluation Metrics 32
vi
LIST OF ABBREVIATIONS
ABBREVIATION FULL FORM OF ABBREVIATION
AP Average Precision
AVM Around View Monitor
CLI Command-Line Interface
CNN Convolutional Neural Network
GPS Global Positioning System
GUI Graphical User Interface
IoT Internet of Things
IoU Intersection over Union
mAP Mean Average Precision
NMS Non-Maximum Suppression
OCR Optical Character Recognition
RANSAC Random Sample Consensus
RFID Radio Frequency Identification
RNN Recurrent Neural Network
SSD Single-Shot Detector
SVM Support Vector Machine
YOLO You Only Look Once
vii
CHAPTER 1
INTRODUCTION
Car parking is a major issue in urban areas where limited parking spaces and high demand
lead to congested roads, increased traffic and frustrated drivers. To address this issue, various
solutions have been proposed and implemented. However, most of these solutions are still not
effective enough to provide real-time information about available parking spaces. To overcome this,
the integration of deep learning techniques with computer vision has shown promising results in the
development of effective car parking space detection systems. This new approach to car parking
space detection leverages the advances in machine learning and computer vision to provide accurate
and real-time information about the availability of parking spaces in urban areas.
The use of deep learning algorithms has proven to be a highly effective solution for car parking
space detection. These algorithms utilize complex neural networks to analyse large amounts of data
and make accurate predictions about the presence of objects in an image. In the case of car parking
space detection, deep learning algorithms can quickly and accurately identify the presence of cars
in parking spaces, making it possible to determine the availability of parking spaces in real-time.
There are several advantages to using deep learning for car parking space detection. First,
deep learning algorithms are highly accurate, providing a more reliable and effective solution than
traditional methods. Finally, deep learning algorithms are flexible and can be easily adapted to
different parking scenarios, making them a highly versatile solution for car parking space detection.
The purpose of parking space detection is to provide drivers with information about the
availability of parking spaces in a given area, allowing them to make informed decisions about
where to park their vehicle. This is achieved by using a variety of technologies, including cameras,
sensors and computer vision algorithms, to detect and track the presence of vehicles in parking
spaces. The result is an efficient and effective system that can quickly and accurately identify
available parking spaces, reducing traffic congestion and improving the overall experience for
drivers.
In recent years, deep learning algorithms, such as YOLO (You Only Look Once), have
become increasingly popular for parking space detection due to their accuracy and efficiency. These
algorithms are trained using large datasets to identify and classify objects in images, making them
well-suited for parking space detection tasks. By utilizing these cutting-edge technologies, parking
space detection systems are becoming increasingly sophisticated and effective, providing drivers
with real-time information about the availability of parking spaces in their area.
1
CHAPTER 2
LITERATURE REVIEW
2.1 IoT based Intelligent Parking Management System

Author: Annirudh D, Arun Kumar D, Adabala Taraka Sai Raghava Kumar &
K.R.M. Vijaya Chandraka
Publication: IEEE 2nd International Conference on Control, Measurement and
Instrumenation (CMI)
Year of Publication: 2021
The biggest problem for car owners driving in urban areas is finding a parking space for their
car. In most cases, people waiting in front of parking spaces are informed at the last moment that
the parking space is not available. The purpose of this work is to provide users with real-time
information about the number of free parking spaces using the GSM messaging service and to
manage parking and calculate fees using Optical Character Recognition (OCR) and time stamps. It's
about automating. This significantly reduces customer waiting times and also reduces parking labor
costs.
In a country like India, most of the city streets are congested with cars. The majority of
vehicles that cause traffic jams are looking for efficient parking spaces. The average vehicle spends
15 minutes looking for an available parking space at a very slow speed. The main reason for this
situation is the unavailability of parking spaces in shopping malls and other parking slots.
In addition, drivers waiting for parking are notified at the last moment that the parking space
is not available. With so many cars on the street, there isn't much you can do to keep the parking
slot full. The least that a parking slot can do is to publish the information on the number of available
parking slots so that customers who want to park their car can check this information and check the
availability of experienced parking slots, that's it.
So, customers can plan or change their plans accordingly. This significantly reduces traffic
congestion outside the parking slot and shortens waiting times for customers. To do this, they need
to collect data from all parking spaces under consideration. This includes user notification systems
using either GSM messaging or cloud-based systems. Additionally, vehicles receive time-stamped
tokens at entry and exit and spend a significant amount of time paying for parking. This entire
parking fee collection process can be automated. This allows for seamless movement of vehicles
and also reduces the need for manpower in parking slots.
2
The suggested smart parking system is a dependable and cost-effective parking slot
management system. The entire system is designed to provide the user with maximum flexibility by
allowing the user to add more modules to the system as required. For example, because they employ
Automatic Number Plate Recognition, by adding an additional module, the tariff data might be
delivered to the car owner's phone numbers. They could also add an Android app for consumers to
view the slot status, pre-book a parking spot and pay the rate, which would cut customer wait time.
If the system is installed in several parking slots across the city, the aggregate data may be
evaluated to determine the pattern of individuals visiting a specific spot. The parking slot
management can utilise this data to forecast a rise in the number of cars and prepare for capacity
expansion. The data may also be used to design the capacity and placement of a new parking slot in
the city.
Despite the smart parking system's many advantages, there are few more potential drawbacks
to consider such as setting up such a system can be costly, involving hefty investments in
infrastructure, technology and more over to run and maintain the system, it may be necessary to
have specific expertise and training. Any technological issues or malfunctions could lead to
erroneous data, system unavailability and lost sales due to dissatisfied customers. This have some
privacy concerns such as the system records information about client vehicles, licence plate numbers
and parking behaviours. This poses privacy issues and might need following stringent data
protection laws.
2.2 Smart Parking System Based on Optical Character Recognition

Author: Athira A., Lekshmi S., Pooja Vijayan, Boby Kurian
Publication: IEEE 3rd International Conference on Trends in Electronics and
Informatics (ICOEI)
Parking slot detection and user notification are the two major sections of a smart parking
system. The empty parking space detection was initially done by deploying a number of sensors in
the parking slot. It is highly expensive and complicated to install. But the advancement of image
processing has enabled us to use images of the parking area to find out empty spaces.
When the drivers are in search of a parking space, the possibility for accidents increases as
they give less attention to the road. A sophisticated car parking system can only solve these
problems. That is why numerous research works are taking place in this area all around the world.
Empty parking slot detection is the first phase of any smart parking system. The second phase is
sharing this information to the drivers who are in search of parking slots.
3
Wherever there is a significant amount of traffic some form of car parking system is also
needed. Rapid rise of urban population and the increase in traffic congestion has escalated the
demand for the researches in the smart parking system domain during the past two decades. By
taking the advantage of some modern tools in information and communication engineering, the
parking space information can be shared effectively in real time.
Modern smart parking systems can be broadly divided into two based on the technology used
for identifying the empty parking spaces in a parking slot, one is Sensor based methods and another
is Image processing-based methods.
Here in this paper they make a comparative study about different parking spot detection
techniques and they also propose an advanced parking spot identification technique using OCR
algorithm of image processing.
The first section of the smart parking system is the parking slot detection. This can be done
using sensors placed at different locations of the parking slot. But this is a very costly system with
a slot of drawbacks. The image processing based empty parking slot identification has made the
system simple as well as cheap. Using suitable image processing algorithms, the accuracy of
detection can be improved.
The parking spot detection is the initial part of the smart parking system. This may be
accomplished by placing sensors throughout the parking slot. However, this is an extremely
expensive device with several downsides. The image processing-based vacant parking slot
recognition has simplified and reduced the cost of the system. They merely need to utilise photos
recorded by the parking slot surveillance camera to discover empty slots here. Future smart parking
systems will use an OCR based approach. The detection accuracy can be increased by using
appropriate image processing methods.
This can be done using sensors placed at different locations of the parking lot. But this is a
very costly system with a lot of drawbacks. The image processing based empty parking lot
identification has made the system simple as well as cheap. Here we need only to use images
captured by the surveillance camera in the parking lot for empty slot detection. OCR based solution
will be an integral part of future smart parking systems. Using suitable image processing algorithms,
the accuracy of detection can be improved.
Some of the drawbacks including the image processing-based approach is dependent on the
parking lot's cameras. The system's accuracy could be jeopardised if the cameras break down or
become obscured. The parking lot can only be partially viewed by the cameras with some
restrictions. This implies that regions not covered by the cameras may not be able to be detected by
4
the system as empty parking spaces. The system's accuracy may be impacted by things like the
weather, the illumination and obstacles. For instance, glare, reflections, or shadows can obstruct the
image processing algorithms and cause erroneous detections. Moreover, the efficiency of the image
processing algorithms utilised will determine how accurate the system is.
2.3 Smart Parking Using IoT Technology

Author: Rachapol Lookmuang, Krit Nambut, Sasiporn Usanavasin
Publication: IEEE 5th International Conference on Business and Industrial
Research (ICBIR)
The efficient and intelligent parking system management with the IoT technology provides
wireless access to the system and allows the user to keep this one. Tracking of parking space
availability. with an increase in Car Population and Road Congestion in Large Cities That's the main
problem they faced. This work aims to solve this problem. Users typically waste time and effort
Looking for specific free space availability parking. Email parking information notification. This
reduces user latency when searching. Parking spaces are kept to a minimum. Adopt Radio
Frequency Identification (RFID) technology to Prevent car theft.
The IoT can switch information thru the community without concerning human interactions.
IoT permits a person to apply for a lower-priced wi-fi era and additionally allows the person to
switch the information into the cloud. IoT allows the person to hold transparency. The concept of
IoT commenced with the identification of factors for connecting numerous gadgets. The Internet
permits the records to be sent, get hold of or maybe speak with the gadgets. The parking trouble
reasons air pollutants and site visitor congestion. In a state of an art scenario, a parking area is
difficult to look in a day after day existence for the people.
According to the latest survey, there might be a fast growth inside the vehicle's populace of
over 1.6 billion around 2035. Around 1,000,000 barrels of the world's oil are being burnt every day.
Thus, a clever parking gadget is an important thing answer to lessen the waste degree of gasoline.
The answer for the trouble is being raised. Clever parking may be an option to minimize a person's
time and performance in addition to the general value of the gasoline burnt on the lookout for the
parking area. In this, the information is accrued from the sensor and through analyzing and
processing, the output is obtained.
They created this smart parking system with hardware and software based on the IoT concept,
as well as a mobile application that allows the driver to easily check parking information and pay
5
the parking fee via mobile payment. Their study's goal is to improve the parking process by
shortening the time it takes to park a car.
To the best of their knowledge, they are the first to build this type of device architecture and
mobile application to identify vehicle/registration plate objects. They evaluated the system to
discover the optimal range sensor detection for the camera to snap a decent picture of the view. The
findings reveal that our different ranges have an influence on each element of the extraction text,
allowing our application to apply the suggested technique to give appropriate function services to
the user.
While the adoption of IoT technology in parking management has various advantages, there
are also possible disadvantages to consider. To begin, building an IoT-based parking system might
be costly, demanding major investments in hardware, software and infrastructure. This cost may be
exorbitant for certain towns or groups with limited funds, making large-scale implementation
difficult.
IoT-based parking solutions rely significantly on technology, which might be faulty at times.
Technological concerns, such as network connectivity issues or sensor breakdowns, might result in
erroneous data, causing consumers confusion and annoyance and implementing an IoT-based
parking system necessitates extensive planning and cooperation with a wide range of stakeholders,
including city authorities, parking management businesses and technology providers.
2.4 A Grid Projection Method Based on Ultrasonic Sensor for Parking Space
Detection
Author: Yunfeng Shaoa, Pengzhen Chena, Tongtong Caoa
Publication: IEEE International Geoscience and Remote Sensing
Symposium (IGARSS)
In this paper a parking space detection method is proposed. It is based on the grid map
projection and utilizes the ultrasonic sensors. Firstly, it applies a virtual grid map to quantify the
observing target space and build a coordinate system. Then the probable outline of the targets can
be deduced from the ultrasonic wave echo signal and projected into the grid map.
The boundary of the object can be obtained by comparing the overlap number of the outline
in the grid with a threshold. In this way, the size of the target space can be calculated and determined
whether it is proper for parking a vehicle. The method is simple while remaining effective.
6
Various sensors have been applied in the detection of parking space, including the vision
sensor, short range radar and laser radar to name a few of them. For instance, vision sensor such as
the camera is excellent in the scene recognition, the parking line detection.
For instance, when the parking area is located indoor with a dark surrounding environment or
the artificial light are too dazzling, the performance of the sensor will decrease. Additionally, it is
difficult detect the parking slot when the parking line is damaged or shaded from other obstacles. It
can be easily disturbed by the external condition.
In this paper, they have proposed a grid projection method to calculate the size of the parking
space. It can detect the boundary of the obstacles by counting the total overlap numbers of the
deduced outlines according to the call back ultrasonic signal. It is founded that the edge of the
obstacle, overlap the number of echo signal which is larger than the non-boundary area. The
performance of this method is only influenced by the space size, it can tell the approximate shape
of the target and has been verified according to the experimental tests.
For the parallel parking slot case in which the size of the parking space is very small, the
measurement error of the space detection is 0.2m. Moreover, for the perpendicular parking slot
situation in which the size is larger, it can precisely measure the size of the parking slot. This method
can be applied in multiple scenes. Furthermore, the proposed method has a significant advantage
regarding the calculation and implementation.
It can detect the boundary of the obstacles by counting the total overlap numbers of the
deduced outlines according to the call back ultrasonic signal. It is founded that the edge of the
obstacle, the overlap number of the echo signal is larger than the non-boundary area. The
performance of this method is only influenced by the space size, it can tell the approximate shape
of the target and has been verified according to the experimental tests. For the parallel parking slot
case in which the size of the parking space is very small. Moreover, for the perpendicular parking
slot situation in which the size is larger, it can precisely measure the size of the parking slot.
The Ultrasonic sensor used in the test works at a frequency of 20Hz and the wave beam angle
is 60 degree. The position of the vehicle where the sensor is installed can be obtained through Global
Positioning System (GPS). Considering the fact that the smaller size of the parking space is, the
bigger challenge will be detecting it accurately. Parallel parking is to park a vehicle parallel to the
road and keep in line with other parked vehicles. Therefore, the size of the parking space needed
should be larger than the length of the vehicle.
Thus, it is difficult to get the accurate edge of the target. In other words, it is easy to recognize
the target, but it is hard to tell the shape information of it using the traditional detecting method.
7
2.5 A Review on Automatic Parking Space Occupancy Detection
Author: Twinkle Singh, Safdar Sardar Khan, Surendra Chadokar
Publication: International Conference on Advanced Computation and
Telecommunication (ICACAT)
The problem of parking is also increasing due to the increase in the number of vehicles. From
the last decade, there are various researches took place with an objective to develop an ideal
automatic parking slot occupancy detection. There is an auto mechanism that can park vehicle
automatically but it is required to detect which parking slot is vacant and which one is occupied.
The systems which have been designed so far are intended to build parking slot detection
using some sensors or mechanism that can indicate whether the vehicle has been parked correctly
or not. This paper has been designed for reviewing various techniques which have been used for
automated parking slot detection till now.
There are different methods related to parking system have been employed till now such as
user interface-based approach, free space-based approach, parking slot marking-based approach,
infrastructure-based approach is the best example for parking slot marking based approach. Around
View Monitor (AVM) provides 360 degrees around the system nearby visual images.
Hierarchical tree structure is used for the marking of parking slots of various types. AVM
system is used for the assistance to the driver while parking. System consists of four cameras which
is capable to show the footage of all direction. Visual assistance on the monitor while parking helps
the driver to accurately park his vehicle.
Ultrasonic sensors, RFID and some other sensing technique were also used to develop an
Automatic Parking System. The method which has been used in this paper employs simple detector
to find the slots and merges sequentially obtained slots. The technique applied in this paper uses
four different types of parking slot marking which are rectangular, slanted rectangular, diamond and
open rectangular types of Transformation of current and previously detected images of parking slot
took place to predict the current position of earlier detected slot. Thus, obtained outcomes of slots
are clustered according to their assortment and orientation.
The method in which parking assistance system is developed to find the vacant slot for parking
and reserve, it. Fragmentation of parking area into blocks is done by using calibration process. Every
block is organized in a manner through which vehicles can be identified and process the information
with reference to the vacant and occupied slot to the driver.
8
As per the survey of many approaches most of the systems are using sensors like ultra-sonic,
AVM, proximity and fish cameras which increase the cost of the system since these components
needs to install in every vehicle to execute the system. AVM itself indicates whether the vehicle has
been parked correctly or not along with occupancy detection but vehicles can be parked manually.
Instead of using AVM or other monitoring sensors a visual based approach that can be
installed at parking slots at particular elevation which is useful and that can intelligently monitor
several slots along with vehicles transition. It would be best for vehicle security views that indicate
vehicle transitions and pedestrian occurrence. The real time parking slot detection can be used for
automatic parking management. But accuracy is often important which is required for developing
an ideal system. OpenCV based approaches can build a cost-effective system.
The proposed system is not reliable for practical execution as the sensors needs to install in
every vehicle which is not cost effective. Meanwhile, sensors are able to detect the free space only.
System is not trained in a way to detect the adjacent and slanted slots.
2.6 Mobile Outdoor Parking Space Detection Application

Author: Chin-Kit Ng, Soon-Nyean Cheong,
Erfan Hajimohammadhosseinmemar, Wen-Jiun Yap
Publication: IEEE 8th Control and System Graduate Research Colloquium
(ICSGRC)
Finding a vacant parking space in outdoor parking slots is a daily concern of most vehicle
drivers during rush hours, especially in the urban context. In this paper, an outdoor parking space
vacancy detection system is proposed, using mobile devices to improve parking space searching
experience for vehicle drivers by providing them with the location and occupancy information of
parking spaces. The system uses state-of-the-art image recognition algorithm, Convolutional Neural
Network (CNN) with a Raspberry Pi to identify vacant parking spaces from a parking slot image
retrieved in real time via an IP camera.
Due to the ever-increasing parking demand in contrast with the shortage of parking facilities
available, it has become an everyday norm for vehicle drivers to bear with the frustration of
ceaselessly circling the parking slots for prolonged time just to look for a vacant parking space. The
time-consuming process of cruising for parking is not only non-environmentally friendly due to
carbon emissions from vehicles, but also constitutes as one of the major causes of traffic congestion.
Some of these parking slots are also installed with additional display panels in their interior to
navigate drivers towards those available parking spaces by showing the number of vacant parking
9
spaces along with the corresponding direction. The time-consuming process of cruising for parking
is not only non-environmentally friendly due to carbon emissions from vehicles, but also constitutes
as one of the major causes of traffic congestion.
The proposed system has its detection mechanism realized using state-of-the-art image
recognition algorithm, namely CNN to identify vacant parking spaces within an outdoor parking
slot monitored using a stationary IP camera on real time basis. The capability of CNN classifier is
to provide high detection accuracy on vacant parking spaces serves as the basis for assuring the
reliability and practicality of the Driver App to generate smart parking service for the benefits of
vehicle drives.
As for computer vision-based system, existing methods for vacant parking space detection can
be categorized as car-driven or space-driven as proposed by Huang and Wang. The former method
uses car features in vehicle detection algorithms to determine vacant parking spaces based on the
results of vehicle detection. For instance, Wang et al. presented a vision based vacant parking space
detection system which employed a feature-based background model and foreground feature
extraction module to identify the occupancy status of outdoor parking spaces.
With this feature, vehicle drivers can easily find a vacant parking space without having to
circle around the outdoor parking slot, thus improving their parking space searching performance.
An outdoor parking space vacancy detection system was designed and implemented for real-
time detection of parking space vacancy using CNN classifier at an outdoor parking slot within a
university campus.
While the suggested outdoor parking spot vacancy monitoring system offers several benefits,
there are some possible drawbacks to consider the system's implementation can be costly because it
necessitates the installation of IP cameras, Raspberry Pi devices and software development. While
the CNN algorithm is cutting-edge image recognition technology, it is not flawless and may
occasionally misidentify parking spots as available or occupied, providing drivers with inaccurate
information. The usage of IP cameras creates privacy issues since it entails recording photos of
people and cars in public places. To solve these issues, the system would need to be built and
deployed with privacy in mind. The system is reliant on technology and any technological
malfunctions or failures can cause serious issues.
10
2.7 Automatic Parking Space Detection System
Author: Nazia Bibi, Muhammad Nadeem Majid, Hassan Dawood & Ping guo
Publication: IEEE 2nd International Conference on Multimedia and Image
Processing (ICMIP)
In this paper, a system is proposed that will detect the total number of available parking spaces
and displays the information to the drivers so that they can easily parked their cars. A web camera
is used to get the images of the parking area and image processing techniques are used to detect the
presence or absence of cars to count and locate the available parking spaces. The status of the
parking slot is updated whenever a car enters or leaves the parking slot.
Various methods and techniques have been proposed to overcome the problem of parking in
the congested areas. A method for counting the vehicles at the checkpoint from which the number
of available parking spaces can be counted. The counting is performed by installation of the
induction loop sensors under the road surface. Although the usage of sensors was less costly, not
easily affected by environmental conditions and it detects accurately however, its installation was
difficult and cause damage to roads. It was also difficult to maintain it in case of malfunction.
Moreover, the exact locations of free parking area cannot be determined because the counting
method is not able to give the detail information, it just records the number of vehicles passing the
checkpoints.
The other kinds of detection methods are presented based on vision-based methods. Through
vision-based methods, the whole parking area available for parking can be examined though the
camera, the data is than processed and the result generated will determine the exact number and
location of the free parking spaces. A proposal that vision-based parking space detection methods
are very easy to install, low in cost and the detector can be easily adjusted according to requirements.
Moreover, the data obtained from images is very rich. However, the defects in the vision method
are that the accuracy is highly dependent upon the position of the camera.
An image-based method for detection of free slots in the outdoor parking area. A low-
resolution web camera is utilized for acquiring images of the parking slot that reduces the cost
greatly. The images acquired are preprocessed and then a pair of Region of Interests is applied on
every division of the parking slot, which increases the reliability of detecting vehicles.
To strengthen the recognition capacity of system video data was captured at different
environmental conditions and temporal shifts. Video is segmented into frames. Then from each
segment a key frame is extracted and further processing is applied on this key frame, to reduce
11
computational complexity. When radio-controlled toy car enter or leave the parking lot from parking
arena, motion of car is estimated by key frame subtraction.
The main contribution of this study is to optimize the identification of available parking slots
to possibly reduce the congestion in parking arena. Due to advancement in machine learning and
vision base technology cost effective automatic parking systems facilitate the drivers to locate
available spaces at parking arena. Future researchers can focus on allocation specific location to
customers already registered from online parking management system.
The disadvantage of this method was that the sensors are easily affected by weather conditions
like rain, temperature, snow and fast air breeze. Another method was presented based on wireless
sensor nodes. This method was less costly and it uses the wireless sensor nodes implemented at the
critical places like the lane turns, entrance and exit positions of the parking slot. The total number
of cars in the parking area can be determined by the difference of incoming and outgoing cars.
2.8 Automatic Parking Space Detection and Tracking for Underground and
Indoor Environments
Author: Jae Kyu Suhr and Ho Gi Jung
Publication: IEEE Transactions on Industrial Electronics
Although many public parking slots in heavily populated countries are located underground
or indoors, most existing methods have difficulty handling such scenarios. These parking slots also
consist mostly of narrow perpendicular parking slots along with numerous pillars, which causes the
most commonly used ultrasonic sensor-based parking systems to frequently fail to detect available
parking spaces due to inaccurate range data. Also, dim lighting and reflections cause imaging
sensor-based methods to work improperly. It can be seen that they contain various severe situations
such as amplified noise, road surface reflections, low contrast markings and pillars. In addition, the
reflections on the road surfaces and the existence of pillars partially hide parking slot markings in
the AVM images.
The proposed method detects vacant parking spaces by combining two complementary
approaches- free space-based and parking slot marking-based. However, its performance depends
on the existence and positions of adjacent vehicles. The parking slot marking-based approach finds
parking spaces by recognizing slot markings on road surfaces. While its performance is independent
of adjacent vehicles, it can be degraded under severe illumination and slot marking conditions.The
proposed method takes advantage of these two approaches in order to achieve robustness against
illumination, slot marking and adjacent vehicle conditions. It presents a method that reliably
12
recognizes parking spaces in underground and indoor environments based on a high-level fusion of
free space-based and parking slot marking-based approaches. It proposes a method that efficiently
detects pillars via a low-level fusion of AVM images and ultrasonic sensor data and utilizes them to
increase both parking space detection and tracking performances. It suggests a method that robustly
recognizes parking slot markings under severe illumination conditions using Random Sample
Consensus (RANSAC) and chamfer matching.
An automatic parking system consists of target position designation, path planning and path
tracking by active steering. This paper only deals with target position designation, as both path
planning and path tracking have already been adopted by mass produced vehicles.
Target position designation methods can be categorized into four approaches: free space-
based, parking slot marking-based, user interface-based and infrastructure-based. Since this paper
is concerned with free space-based and parking slot marking-based approaches, this section focuses
on these two approaches.
The free space-based approach finds vacant parking spaces by recognizing adjacent vehicles.
This is the most popular approach as it can be implemented using various range-finding sensors.
However, this approach has a fundamental drawback in that it cannot find free spaces when there is
no adjacent vehicle and its accuracy depends on the positions of adjacent vehicles. Among a variety
of range-finding sensors, an ultrasonic sensor is most widely used as it is easy to attach to vehicles
at low cost.
The parking slot marking-based approach finds parking spaces by recognizing markings on
road surfaces. Unlike the free space-based approach, performance of this approach does not depend
on the existence and positions of adjacent vehicles. However, it cannot be used in cases where
parking slot markings are not present or are severely damaged. All methods in this approach utilize
imaging sensors and can be categorized into semi-automatic and full-automatic methods. The
methods in semi-automatically detect parking slot markings.
Parking slots are detected by combining two separating lines. To reduce the number of false
detections, the proposed method utilizes two constraints: one is that one of the two separating lines
should contain both positive and negative lines; the other is that a parking slot width should be
between 200 cm and 400cm.
In addition, it is hard to separate a pillar and a parked vehicle if they are located next to each
other, thus deteriorating the obstacle position estimation accuracies of the ultrasonic sensor-based
method. To overcome this problem, this paper proposes a method that detects pillars based on a
low-level fusion of AVM images and ultrasonic sensor data and generates free spaces according to
13
the positions of the detected pillars. It is notable that the pillar detection results are not only used to
enhance the free space detection performance but also improve the parking slot tracking accuracy.
This paper has proposed a parking space detection and tracking method for underground and
indoor parking slots. It has shown that are parking spaces can reliably be detected by taking
advantage of two complementary approaches: free space-based and parking slot marking-based,
pillars can effectively be detected by fusing AVM images and ultrasonic sensor data and are useful
to enhance both parking space detection and tracking performance and parking slot markings can
robustly be recognized by RANSAC and chamber matching-based parallel line detection.
Nevertheless, a few potential issues or restrictions that could surface are dependence on
certain sensors. The suggested technique could largely rely on particular kinds of sensors, including
image or ultrasonic sensors, which might have limits in particular situations or circumstances. For
instance, noisy or reflective environments may make ultrasonic sensors less effective, while dim
lighting may have an impact on image sensors. Depending on how complicated the suggested
solution is, it can need a lot of processing power or hardware, which would raise the implementation
cost. This could make it less useful in specific circumstances or for particular users. How effectively
the suggested strategy will work in practice cannot be predicted without extensive testing in real-
world settings.
2.9 A Convenient Vision-Based System for Automatic Detection of Parking

Spaces in Indoor Parking Lots Using Wide-Angle Cameras
Author: Shen-En Shih and Wen-Hsiang Tsai
Publication: IEEE Transactions on Vehicular Technology
In this paper, they investigate the use of cameras for parking slot management. Most cameras
used in parking slot systems are perspective cameras; wide-angle cameras, such as fisheye-lens or
catadioptric cameras, are not commonly adopted yet. No matter what type of camera systems is
used, one problem in using them is the complicated system setup procedure, including camera
calibration, environment learning, object modelling. A convenient indoor vision-based parking lot
system has been proposed in this study, which is easy to set up by a typical user with no technical
background and can detect vacant parking spaces automatically. The system uses wide-angle
cameras, such as fisheye-lens or catadioptric cameras and analyses parking space boundary lines
based on a new camera model proposed in this study. In addition, finding vacant parking spaces is
often a trouble for people, particularly in large indoor parking slots, where one has to spend much
14
time with concentration to drive through the entire parking area to visually search for a vacant space
if the parking slot is nearly fully occupied.
This study aims to develop an intelligent vision-based system using wide-angle cameras for
parking slot management, which has merits:
 The camera system can be set up easily by a common user with no technical background,
Parking spaces can be detected precisely,
 Vacant parking spaces can be identified automatically for convenient car parking.
 In these methods, the regions of the parking spaces were segmented out manually
On the contrary, the method proposed in this study aims to conduct this major step
automatically to construct an easy-to-set-up system. The indoor parking slot system proposed in this
study utilizes multiple wide-angle cameras affixed on the ceiling and looking downward vertically.
In this paper, a simplified camera model is proposed for this aim. To design an easy-to-set-up
vision-based parking slot system, the camera calibration process must be easy to carry out by users
with no technical knowledge.
The proposed vision-based parking slot system has at least the following merits are wide-
angle cameras are used so that less cameras are needed to cover the area of a given parking slot, the
camera system can be calibrated using only one line in the environment, so that the system can be
set up easily by a user with no technical background, the proposed Hough transform with the new
cell accumulation scheme generates equal-width curves.
So that the proposed system is capable of directly dealing with distorted images captured by
wide-angle cameras and the space line detection results are more precise than those yielded by
conventional methods, unlike many previous studies, which specify parking spaces manually, the
proposed method detects them automatically for a convenient system setup as well.
While the suggested vision-based parking lot management system has numerous advantages,
there are also possible drawbacks to consider. To begin, the suggested system is strongly reliant on
the precision of the camera calibration procedure. If the camera is not correctly calibrated, it might
result in erroneous space recognition, causing confusion for cars seeking for parking places.
The suggested system employs wide-angle cameras, which can result in visual distortion,
particularly around the image's edges. While the suggested Hough transform with the new cell
accumulation strategy is intended to cope with distorted pictures, its usefulness may be limited in
extreme situations of distortion. As a result, it is critical to test the system under various scenarios
to guarantee that it can function properly.
15
2.10 Parking Space Detection with Hierarchical Dynamic Occupancy Grids
Author: Matthias R. Schmid, S. Ates, J. Dickmann, F. von Hundelshausen and
H.- J. Wuensche
Publication: IEEE Intelligent Vehicles Symposium (IV)
An automatic parking system relies on precise estimation of parking space geometry. Today,
most parking aid systems use active sensors like ultrasonic or radar sensors to determine the parking
space. The field of view of these sensors is limited and quite often they have noticeable measurement
noise. For those sensors, the accumulation of measurements over time is helpful to increase
accuracy. To gain a comprehensive map of the environment, one thus needs to accumulate sensor
data.
A dense representation of three-dimensional space would be costly, so a hierarchical data

structure is used. As a secondary effect of the hierarchical data structure, the applications can control
the level of detail of the environment representation dynamically. While it is easy to detect free
parking space in a two-dimensional dense grid, it is challenging to detect free space in a hierarchical
three-dimensional data structure.
The approach proposed in this paper is able to detect spots for parallel parking as well as
perpendicular parking. This paper describes two contributions. First, the extraction of parking
spaces from a hierarchical three-dimensional occupancy grid and secondly the control of the grid
resolution based on the detected parking space.
The hierarchical data structure allows dynamic control of the level of detail in the grid based
on the requirements of the applications. Further, the data structure makes three-dimensional grids
feasible because of the sparse representation of occupied cells. Areas with a homogenous occupancy
probability are grouped together into larger cells. This reduces memory consumption and computing
power requirements.
For the estimation of a free parking space, the contour of the parking space has to be modelled
accurately. However, it is not practicable to represent the whole surrounding in a high level of detail.
Therefore, a hierarchical grid structure is used. This approach increases the level of detail only at
the borders of the parking space, whereas all other objects are represented in a low level of detail
for the parking function.
For a full parking aid system, other components such as trajectory calculation and execution
are necessary. A set of parameters of the detected parking space is therefore defined as an abstract
16
interface to other components. A big improvement in the proposed approach can be achieved using
even smaller grid cells in future evaluations and by using radar sensors with a higher angular
resolution. Future work should focus on sensors with real three-dimensional data such as a multi-
layer laser scanner or a stereo camera.
Here the full potential of the three-dimensional grid will unfold. For high reliability, a fusion
of sensors with different measuring principles should be considered. Possible sensors could be short
range radar sensors and cameras or short-range radar sensors and a laser scanner.
While the suggested solution for automatic parking systems based on a hierarchical three-
dimensional occupancy grid data structure offers some advantages, it also has certain drawbacks
that should be considered. To store and analyses the hierarchical data structure, the technique needs
a large amount of computer power and memory. While the sparse representation of occupied cells
saves memory usage, the system still requires a substantial quantity of data to be processed,
especially when employing smaller grid cells. This might be difficult for some embedded systems
or low-power devices. The proposed method may not be appropriate for all parking circumstances.
It may not operate effectively, for example, in busy parking lots with numerous parked cars or on
small roadways with little moving space. The method assumes that the parking place is available.
2.11 Summary
In an overview of this Literature research, we learned through that IoT and Sensors play a
significant part in the present System. Furthermore, we discovered that challenge in the approaches.
As a result, it becomes sense to consider Deep Learning Strategies. So that the laborious effort and
severe workload of the prior analysis may be eased.
In furthermore, we were inspired by the technique used in wide angle camera views with
embedded systems to detect planes for space identification, extract features from image or video
frames from the trained model using YOLO and remove conflicts using Non-Maximum Suppression
(NMS) strategy to select the single entity (Bounding boxes) out of overlapping entities. The above
literature analysis leads us to the conclusion that we should construct a deep learning model utilizes
an image processing technique to improve the recognition and correct the faults in the problem to
solve it swiftly.
In addition, we compared the research findings to a variety of methodologies in order to

determine the best answer for a dynamic approach to determining car space occupancy.
17
CHAPTER 3
SYSTEM ANALYSIS
This chapter gives the complete outlook of what the existing system will do and how the
proposed system will work. This gives the advantage and importance of using this new system.
3.1 Existing System
 Ultrasonic sensors: These sensors use high-frequency sound waves to detect the presence
of a vehicle in a parking space.
 Infrared sensors: These sensors use infrared radiation to detect the presence of a vehicle
in a parking space.
 License plate recognition (LPR) systems: These systems use cameras and OCR
technology to identify and track vehicles as they enter and exit a parking space.
 Magnetic field-based systems: These systems use magnetic sensors to detect the presence
of a vehicle in a parking space.
 Wireless communication systems: These systems use wireless communication
technology, such as Bluetooth or Wi-Fi, to communicate information about the availability
of parking spaces.
 RFID systems: These systems use RFID tags attached to vehicles to track their location
and determine the availability of parking spaces.
 Pressure-sensitive systems: These systems use pressure sensors embedded in the
pavement to detect the weight of a vehicle and determine the availability of parking spaces.
Moreover, each of these systems has its own advantages and disadvantages and the choice of
system will depend on the specific requirements and constraints of the parking environment.
3.1.1 Disadvantages
 Ultrasonic sensors: These sensors can be affected by obstacles such as walls or other
vehicles, reducing their accuracy.
 Infrared sensors: These sensors may not work well in direct sunlight, as the intense light
can interfere with the sensor's ability to detect vehicles.
 License plate recognition (LPR) systems: These systems can be affected by low-light
conditions, making it difficult to accurately read license plate numbers. They may also raise
privacy concerns, as they capture images of vehicles and their license plates.
18
 Magnetic field-based systems: These systems can be affected by metal objects in the
surrounding environment, which can interfere with the magnetic field and reduce accuracy.
 Wireless communication systems: These systems may have limitations in terms of range
and reliability, as they depend on the availability of wireless networks.
 RFID systems: These systems require RFID tags to be attached to each vehicle, which can
be inconvenient for users and may not be feasible for all types of vehicles.
 Pressure-sensitive systems: These systems can be affected by environmental factors such
as rain or snow, which can change the pressure readings and reduce accuracy. Additionally,
they can be damaged by heavy vehicles, reducing their lifespan.
However, the system comes with several disadvantages includes high cost of installation,
maintenance costs, technical difficulties, privacy concerns, dependence on technology, vulnerability
to hacking, inaccurate results and resistance to change. It is important to consider the pros and cons
of each system before deciding on the best approach for a particular parking environment.
3.2 Proposed System
 The proposed system uses the YOLO object detection algorithm to detect and classify
parking spaces in real-time video frames.
 The system was trained on a dataset of parking scenes to improve its accuracy in detecting
parking spaces.
 The performance of the system is evaluated and compared with other existing systems and
the results showed high accuracy in detecting and classifying parking spaces.
 It can detect and classify parking spaces in various lighting and weather conditions, making
it a versatile solution for different parking scenarios.
 It can also reduce the need for manual inspection of parking lots, saving time and resources.
 It has potential applications in intelligent transportation systems, as it can provide real-time
information about parking availability and optimize parking space utilization.
 It also has potential for additional features, such as license plate recognition and payment
processing, to further improve the efficiency and convenience of car parking.
 It has the potential for further improvement and integration with other technologies, such
as machine learning and artificial intelligence, to enhance its performance and capabilities.
Overall, the proposed system has several advantages including its accuracy, speed and ease of
integration. It is a highly accurate and efficient algorithm for object detection, making it an ideal
choice for detecting parking space occupancy. The system can be easily integrated into existing car
parking systems, allowing for a seamless and cost-effective upgrade to the system.
19
3.2.1 Advantages
 Increased Efficiency: By providing real-time parking space detection, the system can
reduce the time and effort required for drivers to find available parking spaces, ultimately
increasing the efficiency of the parking facility.
 Cost-Effective: The integration of the YOLO model into existing parking systems is a
cost-effective way to upgrade the system, as it does not require significant changes to the
infrastructure.
 Accurate and Reliable: The YOLO algorithm is known for its high accuracy in object
detection, ensuring that the system can accurately detect parking space occupancy and
reduce the likelihood of errors.
 Fast Processing: YOLO is also known for its speed in processing images, making it
possible for the system to detect parking space occupancy quickly and efficiently.
 Easy Integration: The YOLO model can be easily integrated into existing car parking
systems, allowing for a seamless and cost-effective upgrade to the system.
 User-Friendly: The system is user-friendly, with clear indicators showing available
parking spaces and reducing the need for drivers to search for available spots.
 Customizable: The system is highly customizable, with the ability to adjust the sensitivity
of the system to fit the specific needs of different parking environments.
 Potential for Additional Features: The system has potential for additional features, such
as license plate recognition and payment processing, which can further improve the
efficiency and convenience of car parking.
Overall, the proposed system offers several advantages that can greatly benefit parking
facilities and their customers, ultimately improving the efficiency, convenience and overall
experience of car parking.
20
CHAPTER 4
SYSTEM REQUIREMENTS
This chapter exhibits the hardware and software requirements for the proper running of the
project. Understanding the system requirements is important as it ensures that the software will run
smoothly on the user's device and meet their expectations for performance and functionality.
4.1 Hardware Requirements

The hardware requirements for using YOLO models depend on the specific architecture and
size of the model.
CPU : At least 2 cores.

GPU : At least 4GB memory for small YOLO models.
At least 8GB for larger YOLO models.
RAM : Minimum 8 GB.
Hard Disk : 500 GB.
Monitor : 15'' LED.
Input Devices : Keyboard, Mouse.
4.2 Software Requirements

The software requirements for using YOLO models include operating systems, necessary
platforms and frameworks on which the software requires support to run.
Operating System : Windows 10 or Linux.

Programming Language : Python 3.5 or higher.
IDE’s : Visual Studio, PyCharm, Anaconda.
Tools : Jupyter Notebook, OpenCV, NumPy, Keras.
Deep Learning Frameworks : TensorFlow.
4.3 Tools and Technology

4.3.1 Python
Python is a high-level, general-purpose programming language. Its design philosophy
emphasizes code readability with the use of significant indentation. Python is dynamically typed
and garbage-collected. It supports multiple programming paradigms, including structured, object-
oriented and functional programming.
21
4.3.2 Pip
Pip is a package-management system written in Python and is used to install and manage
software packages. The Python Software Foundation recommends using pip for installing Python
applications and its dependencies during deployment.
4.3.3 NumPy
NumPy is a library for the Python programming language, adding support for large, multi-
dimensional arrays and matrices, along with a large collection of high-level mathematical functions
to operate on these arrays.
4.3.4 Keras
Keras is an open-source software library that provides a Python interface for artificial neural
networks. Keras acts as an interface for the TensorFlow library. Up until version 2.3, Keras
supported multiple backends, including TensorFlow, Microsoft Cognitive Toolkit, Theano and
PlaidML.
4.3.5 OpenCV
OpenCV is an open-source computer vision library which consists of programming functions
mainly for real-time. It provides a variety of functions and algorithms for image and video
processing. It can be used with several programming languages, including C++, Python and Java.
Its features include object detection, tracking and recognition, as well as machine learning and deep
learning capabilities.
4.3.6 Jupyter Notebook

Jupyter Notebook is an open-source web-based interactive computing environment used to
create and share documents that contain live code, equations, visualizations and narrative text. It
supports a wide range of programming languages, including Python, R, Julia and many others. It is
often used for data exploration, data cleaning, machine learning and scientific computing, but they
can also be used for general-purpose programming, teaching and collaboration.
It also provides a rich ecosystem of extensions and plugins that can be used to add new
functionality and improve your productivity. It integrates with many popular data science libraries
and tools, such as NumPy, Pandas, Scikit-learn and Matplotlib, making it a popular choice for data
analysis and visualization.
4.3.7 TensorFlow
TensorFlow is a free and open-source software library for machine learning and artificial
intelligence. It can be used across a range of tasks but has a particular focus on training and inference
22
of deep neural networks. It also provides a wide range of pre-built algorithms and models, such as
CNN, Recurrent Neural Network (RNN) and more.
It also provides a range of tools for model visualization, including TensorBoard, which
enables users to monitor and visualize the progress of their training runs.
4.3.8 PyTorch
PyTorch is a machine learning framework based on the Torch library, used for applications
such as computer vision and natural language processing. It provides a platform for building and
training deep learning models in a various application including natural language processing,
computer vision. It provides several modules and functions that can help developers build and train
neural networks, including tensor computation, gradient computation and neural network modules.
Additionally, it allows for easy integration with popular Python libraries such as NumPy, SciPy and
Pandas.
23
CHAPTER 5
SYSTEM DESIGN
5.1 System Architecture
Fig. 5.1 System Architecture
24
5.2 List of Modules
 Pre-Processing
 Object Detection
 Post-Processing
 User Interface
5.2.1 Pre-Processing
The Image Pre-processing module is responsible for enhancing the input images or video
frames to improve the accuracy of object detection. The module may perform operations such as
noise removal, image filtering, image enhancement, or resizing of images to suit the requirements
of the object detection module. These operations help remove any unwanted information from the
images that may interfere with the detection process. Furthermore, the module may also perform
data normalization, which makes the input data consistent in terms of image size and format. The
aim is to enhance their quality and make them suitable for object detection. The steps may include
resizing the image to a standard size, normalizing the pixel values to a certain range, applying colour
correction, or removing noise from the image. It is essential to improve the accuracy of the object
detection results.
5.2.2 Object Detection

The Object Detection module is the core of the system and is responsible for detecting and
classifying parking spaces in the input data. The module uses the YOLO algorithm, which is a deep
learning-based object detection algorithm. The algorithm divides the input image into a grid of cells
and applies a CNN to each cell to predict the class and bounding box of any object present. In the
case of parking space detection, the output of the object detection module will include the location
and size of all detected parking spaces. The YOLO algorithm is efficient and fast, allowing real-
time object detection.
5.2.3 Post-Processing
The Post-Processing module receives the output from the object detection module and
performs additional processing to refine the accuracy of the detection results. The module filters out
any false detections or noise in the output and may perform morphological operations such as
dilation or erosion to improve object segmentation. Additionally, the module may also perform
thresholding to segment the detected parking spaces from the input image accurately.
This module is responsible for refining and improving the accuracy of the output of the object
detection module. The output of the object detection module may include false positives or false
negatives and the post-processing module aims to remove these errors. The post-processing steps
25
may include NMS to eliminate overlapping bounding boxes, applying filters to remove small or
irrelevant detections, or re-scoring the detected objects to improve their confidence scores.
5.2.4 User Interface

This module provides a user-friendly interface to interact with the system and display the
output results. The user interface may include a Graphical User Interface (GUI) or a Command-Line
Interface (CLI). The user interface allows the user to input the image or video file, adjust the
parameters of the object detection algorithm and view the output results. The user interface module
is essential to make the system accessible to users with varying levels of technical expertise.
The User Interface module is responsible for providing a user-friendly interface to interact
with the system. It displays the input and output data of the system in a format that is easy to
understand for users. The module also enables users to adjust system settings or parameters to fine-
tune the performance of the system.
5.3 Proposed Methodology

5.3.1 Data Collection
The first step in the methodology is to collect a dataset of images taken from a multi-level car
park. The images were taken at different times of the day and in different weather conditions to
ensure that the system could handle variations in lighting and weather. It used for testing YOLO
model for car parking space detection. The dataset is used to train YOLO algorithm how to recognize
different types of parking spaces in different lighting and weather conditions. The data collection
process involved the following steps:
 Identifying parking lots: The first step in the data collection process is to identify parking
lots in different locations. These locations included public parking lots, private parking lots
and parking garages.
 Capturing images and videos: Once the parking lots were identified, images and videos
were captured using various cameras and recording devices. The images and videos were
captured from different angles and distances to ensure that the dataset represented real-
world parking lot scenes.
 Labelling parking spaces: The captured images and videos were manually labelled to
identify the parking spaces within the scenes. This labelling process involved drawing
bounding boxes around each parking space and assigning a unique label to each box.
 Pre-processing the dataset: The labelled images and videos were pre-processed by
resizing them to a standard size and applying data augmentation techniques to increase the
size of the dataset. The data augmentation techniques included flipping the images
26
horizontally, vertically, or both, adding random noise to the images and changing the
brightness and contrast levels.
 Splitting the dataset: The pre-processed dataset is split into a training set and a testing set.
The training set is used to train the YOLO algorithm, while the testing set is used to
evaluate the performance of the algorithm.
5.3.2 Image Annotation

The next step is to annotate the images by marking the locations of the parking spaces and
vehicles. This is done using a bounding box annotation tool, which allows the user to draw a box
around the object of interest. The annotated images were then used to train and test the YOLO
algorithm. It is an essential step as it involves the manual labelling of the images with bounding
boxes to identify the parking spaces within the images. The process of image annotation involves
drawing a bounding box around each parking space in the image and assigning a unique label to the
box. The labelled images are then used to train the YOLO algorithm to recognize parking spaces in
real-world scenarios. The annotations were used to create ground truth data that the algorithm could
use to learn and improve its accuracy. The image annotation process involved the following steps:
 Selection of annotation tool: The first step in the image annotation process is to select an
appropriate annotation tool. There are several annotation tools available, including
LabelImg, RectLabel and CVAT. For this project, LabelImg is used as the annotation tool.
 Defining the classes: The next step is to define the classes of objects that need to be
annotated. In this project, the classes were defined as "parking space" and "non-parking
space."
 Annotation process: The annotation process involved opening an image in the annotation
tool and drawing a bounding box around each parking space. The bounding box is adjusted
to cover the entire parking space and a unique label is assigned to each box.
 Reviewing the annotations: After annotating the images, the annotations were reviewed
to ensure that all parking spaces were correctly labelled and that there were no false
positives or false negatives.
 Exporting the annotations: Once the annotations were reviewed and finalized, they were
exported in the YOLO format, which includes a text file for each image that contains the
coordinates of the bounding boxes and the class labels.
5.3.3 YOLO Architecture

The YOLO algorithm consists of a CNN that is trained to detect objects in an image. The
CNN is made up of several layers, including convolutional layers, pooling layers and fully
27
connected layers. The final layer of the CNN outputs the bounding box coordinates and class
probabilities for each object in the image. The YOLO architecture is a popular deep learning
algorithm used for object detection in images. In this proposed methodology, the YOLO algorithm
is used to detect parking spaces in images. It has the following key features:
 Single convolutional network: YOLO is a single neural network that takes the entire
image as input and outputs the predicted bounding boxes and class probabilities. This is in
contrast to other object detection algorithms that require multiple stages and separate
classifiers for each object class.
 Grid system: The YOLO algorithm divides the input image into a grid system and predicts
bounding boxes and class probabilities for each grid cell. This approach enables the
algorithm to detect objects of different sizes and aspect ratios.
 Bounding box prediction: The YOLO algorithm predicts bounding boxes that are defined
by their centre coordinates, width and height. This approach is different from other
algorithms that predict the corners of the bounding box.
 Class probabilities: In addition to predicting bounding boxes, the YOLO algorithm also
predicts the class probabilities for each bounding box. This enables the algorithm to
identify the type of object in each bounding box.
The YOLOv3 architecture uses a series of convolutional layers, followed by a set of detection
layers that predict bounding boxes and class probabilities for each grid cell. The YOLOv3
architecture is trained on a dataset of annotated images to learn how to detect parking spaces in real-
world scenarios. The trained YOLOv3 model is then used to detect parking spaces in new images
by analysing the input image and outputting the predicted bounding boxes and class probabilities.
5.3.4 Training and Testing

The annotated images were used to train the YOLO algorithm. The training process involves
adjusting the weights of the CNN so that it can accurately detect the objects in the images. Once the
training is complete, the algorithm is tested on a set of images taken from the same car park. The
following are the steps involved in training and testing the algorithm:
 Collection of Data and annotation: The first step is to collect a dataset of images
containing parking spaces and annotate them with bounding boxes around the parking
spaces. This annotated dataset is used for both training and testing the algorithm.
 Training the YOLOv3 model: The annotated dataset is used to train the YOLOv3 model.
The training process involves optimizing the weights of the neural network to minimize
the difference between the predicted bounding boxes and the ground-truth bounding boxes.
28
 Validation: The trained model is then validated on a separate validation dataset to ensure
that it is not overfitting to the training dataset.
 Testing the model: The final step is to test the trained YOLOv3 model on a separate test
dataset to evaluate its performance in detecting parking spaces in new images. The
performance of the model is evaluated using metrics such as precision, recall and F1 score.
The training and testing process is iterative and the model is retrained and tested multiple
times until the desired performance is achieved. The proposed methodology uses of transfer learning
to train the YOLOv3 model on a pre-trained model, such as the COCO dataset, to improve the
accuracy of the model.
5.3.5 Performance Evaluation Metrics

The performance of the YOLO algorithm is evaluated using several metrics. Precision
measures number of true positive detections out of all positive detections, recall measures the
number of true positive detections out of all the ground truth objects. The following metrics are
most commonly used performance evaluation:
 Precision: Precision is a measure of the accuracy of the positive predictions made by the
model. It is defined as the ratio of true positives (TP) to the total number of positive
predictions (TP + false positives (FP)). A high precision value indicates that the model has
a low rate of false positives.
 Recall: Recall is a measure of the completeness of the positive predictions made by the
model. It is defined as the ratio of true positives (TP) to the total number of actual positives
(TP + false negatives (FN)). A high recall value indicates that the model has a low rate of
false negatives.
 F1 score: The F1 score is a harmonic mean of precision and recall and it is used to evaluate
the overall performance of the model. It is defined as 2 * (precision * recall) / (precision +
recall). A high F1 score indicates that the model has both high precision and high recall.
 Intersection over Union (IoU): IoU is a measure of the overlap between the predicted
bounding boxes and the ground-truth bounding boxes. It is defined as the ratio of the
intersection of the two boxes to the union of the two boxes. A high IoU value indicates that
the predicted bounding box is accurately aligned with the ground-truth bounding box.
 Mean Average Precision (mAP): mAP is a commonly used metric for evaluating object
detection models. It is calculated by averaging the precision values at different recall levels.
A high mAP value indicates that the model has high precision across different recall levels.
29
In the proposed methodology, the performance is evaluated using a combination of these
evaluation metrics. The metrics are calculated on a separate test dataset and the model is iteratively
refined until the desired performance is achieved.
5.4 Workflow
Image Acquisition
Image
Pre-Processing
Parking Space
Detection
Parking Space
Occupancy Detection
Highlighting Slot
Status
Fig. 5.2 Workflow Diagram
5.5 Data-flow
Fig. 5.3 Data-flow Diagram
30
5.6 Sequence
Fig. 5.4 Sequence Diagram
5.7 Use Case
Fig. 5.5 Use Case Diagram
31
CHAPTER 6
RESULT AND ANALYSIS
The results of the experiments show that the YOLO algorithm is able to accurately detect the
parking spaces and vehicles in the images taken from the car park. The Average Precision (AP) for
the YOLO algorithm is found to be 0.92, which indicates that the algorithm is able to accurately
detect 92% of the parking spaces and vehicles in the images. The precision and recall values were
also found to be high, with precision and recall values of 0.926 and 0.92, respectively. These results
demonstrate the effectiveness of the YOLO algorithm in detecting parking spaces and vehicles in
real-world images.
Fig. 6.1 Evaluation Metrics
The results of the experiments were compared to those of other object detection algorithms,
including the faster R-CNN and Single-Shot Detector (SSD) algorithms, where YOLO
outperformed these algorithms in terms of speed and accuracy. The YOLO algorithm is faster than
the other algorithms because it uses a single forward pass of the network to detect multiple objects,
whereas other algorithms require multiple passes.
32
The table below shows the summary of findings gotten when motion detection had been turned on.
Table. 6.1: Model Performance with Motion Detection Results
No Video With
Actual Predicted Actual Predicted Wrong Model
File Motion
Empty Empty Occupied Occupied Predictions Accuracy
Name Detection
1 Parking Yes 4 4 1 1 0 100%
Slot
2 Parking Yes 3 3 1 1 0 100%
Slot 2
3 Parking Yes 6 5 5 5 1 83%
Slot 3
4 Parking Yes 5 4 0 0 1 80%
Slot 4
5 Parking Yes 3 3 2 2 0 100%
Slot 5
Average Accuracy 92.6%
The table below shows the summary of findings gotten when motion detection had been turned off.
Table. 6.2: Model Performance with No Motion Detection Results
No Video With
Actual Predicted Actual Predicted Wrong Model
File Motion
Empty Empty Occupied Occupied Predictions Accuracy
Name Detection
1 Parking Yes 4 4 1 1 0 100%
Slot
2 Parking Yes 3 3 1 1 0 100%
Slot 2
3 Parking Yes 6 5 5 4 1 80%
Slot 3
4 Parking Yes 5 4 0 0 1 80%
Slot 4
5 Parking Yes 3 3 2 2 0 100%
Slot 5
Average Accuracy 92%
33
For this thesis we have used 10% of the total images for validating our model so that we can
calculate the mAP of our model. As mentioned earlier we have two classes in our model which are
represented by class_id 0,1. In the case of mAP, precision and recall are the two ways to calculate
the predictions. It can be calculated by the below formula
Precision = TP / (TP+FP)
TP (True Positive): In this case which the network predicted correct and it is True.
TN (True Negative): In this case which the network predicted wrong and it is False.
FP (False Positive): In this case which the network predicted correct but it is False.
FN (False Negative): In this case which the network predicted wrong but it is True.
In terms of accuracy, the YOLO algorithm has evaluation metrics compared to other algorithms,
indicating that it is able to detect more objects in the images with higher accuracy.
34
CHAPTER 7
CONCLUSION AND FUTURE ENHANCEMENT
7.1 Conclusion
In this study, a car parking space detection system is developed using the YOLO algorithm.
The system is trained and tested on a dataset of images taken from a multi-level car park. The results
of the experiments showed that the YOLO algorithm is able to accurately detect the parking spaces
and vehicles in the images, with an AP. The results were compared with those of other object
detection algorithms and the results showed that YOLO outperformed the other algorithms in terms
of speed and accuracy.
The results of this study have important implications for the development of intelligent
parking systems. The use of YOLO in this study has demonstrated that it is possible to develop a
fast, accurate and reliable car parking space detection system that can be used in real-world
applications. These results have the potential to improve the efficiency and convenience of car
parking.
In conclusion, the car parking space detection system developed in this study provides a
promising approach for detecting parking spaces and vehicles in real-world images. The use of
YOLO has shown that it is possible to develop an accurate and efficient system that can be used in
real-world applications. Future work could focus on improving the performance of the system by
using more advanced deep learning techniques and incorporating additional information.
7.2 Future Enhancement

 Multi-camera support: Currently, the proposed system is designed to detect parking
spaces in a single camera feed. In future work, the system can be enhanced to support
multiple camera feeds, enabling more extensive coverage of a parking lot. This would
require integrating a camera network and developing algorithms for camera calibration and
object tracking across multiple feeds. This can increase the coverage area of the system and
improve its accuracy.
 Integration with parking guidance systems: The proposed system could be integrated
with parking guidance systems to direct drivers to available parking spaces in real-time.
By providing drivers with real-time parking availability information, the system could
reduce traffic congestion and help drivers save time and fuel. This integration would
require the development of an interface that can communicate with parking guidance
systems.
35
 Improved object detection accuracy: While the YOLO algorithm is highly accurate,
there is always room for improvement. In future work, the system's accuracy could be
further enhanced by incorporating more advanced object detection techniques, such as deep
learning and machine learning algorithms. This would require a larger dataset and more
computing resources, but it could result in even higher accuracy for car parking space
detection.
 Edge computing: Implementing edge computing can improve the speed and efficiency of
car parking space detection. This can be achieved by deploying the object detection model
on edge devices, such as Raspberry Pi or NVIDIA Jetson, to perform real-time object
detection at the edge of the network, without relying on cloud computing. This can also
reduce network latency and save cloud storage space.
36
APPENDIX
I. SOURCE CODE
main.py
import argparse
import cv2
from model import YOLOv3
import utils
# Parse arguments
parser = argparse.ArgumentParser(description='Car parking space detection using YOLOv3.')
parser.add_argument('--image', type=str, help='Path to input image.')
parser.add_argument('--video', type=str, help='Path to input video.')
parser.add_argument('--weights', type=str, default='weights/yolov3.weights', help='Path to weights
file.')
parser.add_argument('--config', type=str, default='config/yolov3.cfg', help='Path to model config
file.')
parser.add_argument('--classes', type=str, default='classes.txt', help='Path to class names file.')
parser.add_argument('--conf-thresh', type=float, default=0.5, help='Confidence threshold for
detections.')
parser.add_argument('--nms-thresh', type=float, default=0.4, help='NMS threshold.')
args = parser.parse_args()
# Load YOLOv3 model

model = YOLOv3(args.config, args.weights)
# Load class names

with open(args.classes, 'r') as f:
classes = [line.strip() for line in f.readlines()]
if args.image:
# Detect parking spaces in single image
image = cv2.imread(args.image)
boxes, confidences, class_ids = model.detect(image, conf_thresh=args.conf_thresh,
nms_thresh=args.nms_thresh)
utils.draw_boxes(image, boxes, confidences, class_ids, classes)
cv2.imshow('Parking Spaces', image)
cv2.waitKey(0)
cv2.destroyAllWindows()
elif args.video:
# Detect parking spaces in video stream
cap = cv2.VideoCapture(args.video)
while cap.isOpened():
ret, frame = cap.read()
if not ret:
break
boxes, confidences, class_ids = model.detect(frame, conf_thresh=args.conf_thresh,
nms_thresh=args.nms_thresh)
utils.draw_boxes(frame, boxes, confidences, class_ids, classes)
cv2.imshow('Parking Spaces', frame)
if cv2.waitKey(1) & 0xFF == ord('q'):
37
break
cap.release()
cv2.destroyAllWindows()
model.py
import cv2
import numpy as np
class YOLOv3:
def __init__(self, config_path, weights_path):
self.net = cv2.dnn.readNetFromDarknet(config_path, weights_path)
self.output_layer_names = self.net.getLayerNames()
self.output_layer_names = [self.output_layer_names[i[0] - 1] for i in
self.net.getUnconnectedOutLayers()]
def detect(self, image, conf_thresh=0.5, nms_thresh=0.4):

h, w = image.shape[:2]
blob = cv2.dnn.blobFromImage(image, 1 / 255.0, (416, 416), swapRB=True, crop=False)
self.net.setInput(blob)
layer_outputs = self.net.forward(self.output_layer_names)
boxes = []
confidences = []
class_ids = []
for output in layer_outputs:
for detection in output:
scores = detection[5:]
class_id = np.argmax(scores)
confidence = scores[class_id]
if confidence > conf_thresh:
center_x = int(detection[0] * w)
center_y = int(detection[1] * h)
width = int(detection[2] * w)
height = int(detection[3] * h)
left = int(center_x - width / 2)
top = int(center_y - height / 2)
boxes.append([left, top, width, height])
confidences.append(float(confidence))
class_ids.append(class_id)
# Perform NMS to remove redundant overlapping boxes
indices = cv2.dnn.NMSBoxes(boxes, confidences, conf_thresh, nms_thresh)
if len(indices) > 0:
boxes = [boxes[i[0]] for i in indices]
confidences = [confidences[i[0]] for i in indices]
class_ids = [class_ids[i[0]] for i in indices]
return boxes, confidences, class_ids
dataset.py
import os
import cv2
import numpy as np
import xml.etree.ElementTree as ET
38
class YOLODataset:
def __init__(self, data_dir, image_size):
self.data_dir = data_dir
self.image_size = image_size
self.image_paths = []
self.labels = []
self._read_data()
def __len__(self):
return len(self.image_paths)
def __getitem__(self, idx):

image = cv2.imread(self.image_paths[idx])
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = cv2.resize(image, (self.image_size, self.image_size))
boxes = self.labels[idx]
targets = np.zeros((len(boxes), 5))
for i, box in enumerate(boxes):
x, y, w, h, class_id = box
x /= self.image_size
y /= self.image_size
w /= self.image_size
h /= self.image_size
targets[i, :] = [x, y, w, h, class_id]
return image, targets
def _read_data(self):
annotations_dir = os.path.join(self.data_dir, 'annotations')
for filename in os.listdir(annotations_dir):
if filename.endswith('.xml'):
tree = ET.parse(os.path.join(annotations_dir, filename))
root = tree.getroot()
image_path = os.path.join(self.data_dir, 'images', root.find('filename').text)
boxes = []
for obj in root.findall('object'):
class_name = obj.find('name').text
class_id = 0 # Change this to match your own class IDs
bbox = obj.find('bndbox')
x_min = int(bbox.find('xmin').text)
y_min = int(bbox.find('ymin').text)
x_max = int(bbox.find('xmax').text)
y_max = int(bbox.find('ymax').text)
w = x_max - x_min
h = y_max - y_min
x = x_min + w / 2
y = y_min + h / 2
boxes.append([x, y, w, h, class_id])
self.image_paths.append(image_path)
self.labels.append(boxes)
39
train.py
import torch
import torch.optim as optim
import torch.nn as nn
from tqdm import tqdm
from model import YOLOv3
from dataset import YOLODataset
from config import Config
def train(model, optimizer, dataloader, device):

# set model to train mode
model.train()
# define loss function

criterion = nn.MSELoss()
# train the model for one epoch

train_loss = 0.0
for images, targets in tqdm(dataloader):
# move data to device
images = images.to(device)
targets = targets.to(device)
# forward pass
outputs = model(images)
# calculate loss
loss = 0.0
for i in range(Config.S*Config.S):
# get target and output boxes for current grid cell
target_boxes = targets[i, :, :4]
output_boxes = outputs[i, :, :4]
# get objectness score for current grid cell
objectness = targets[i, :, 4]
# get class probabilities for current grid cell
class_probs = targets[i, :, 5:]
# calculate loss for box coordinates
loss_coord = criterion(output_boxes * objectness.unsqueeze(-1), target_boxes *
objectness.unsqueeze(-1))
# calculate loss for objectness score
loss_obj = criterion(outputs[i, :, 4], objectness)
# calculate loss for class probabilities
loss_cls = criterion(outputs[i, :, 5:], class_probs)
# add up all losses for current grid cell
loss += loss_coord + loss_obj + loss_cls
# normalize loss by number of grid cells
loss /= Config.S*Config.S
# backpropagation and weight update

optimizer.zero_grad()
loss.backward()
optimizer.step()
40
# update train loss
train_loss += loss.item()
# calculate average train loss for epoch

train_loss /= len(dataloader)
return train_loss
config.py
class Config:
# general settings
IMAGES_DIR = 'images'
ANNOTATIONS_DIR = 'annotations'
CLASSES_FILE = 'classes.txt'
BATCH_SIZE = 8
NUM_WORKERS = 4
# model settings
MODEL_NAME = 'yolov3'
BACKBONE = 'darknet53'
S = 13 # grid size
B = 3 # number of boxes per grid cell
C = 5 # number of classes
IN_CHANNELS = 3 # number of input channels (RGB)
PRETRAINED = False
# training settings
LR = 0.001
MOMENTUM = 0.9
WEIGHT_DECAY = 0.0005
NUM_EPOCHS = 10
utils.py
import torch
def iou(boxes1, boxes2):

"""Calculate the IOU of two sets of bounding boxes.
Args:
boxes1 (Tensor): Tensor of shape (N, 4) representing N bounding boxes.
boxes2 (Tensor): Tensor of shape (M, 4) representing M bounding boxes.
Returns:
iou (Tensor): Tensor of shape (N, M) representing the IOU between each pair of boxes from
boxes1 and boxes2.
"""
# get coordinates of intersection boxes
tl = torch.max(boxes1[:, None, :2], boxes2[:, :2]) # top-left coordinates
41
br = torch.min(boxes1[:, None, 2:], boxes2[:, 2:]) # bottom-right coordinates
# calculate area of intersection boxes

area_i = torch.prod(br - tl, dim=2) * (tl < br).all(dim=2)
# calculate area of each box

area1 = torch.prod(boxes1[:, 2:] - boxes1[:, :2], dim=1)
area2 = torch.prod(boxes2[:, 2:] - boxes2[:, :2], dim=1)
# calculate IOU
iou = area_i / (area1[:, None] + area2 - area_i)
return iou
def non_max_suppression(boxes, scores, iou_threshold):

"""Perform NMS on a set of bounding boxes and scores.
Args:
boxes (Tensor): Tensor of shape (N, 4) representing N bounding boxes.
scores (Tensor): Tensor of shape (N,) representing the scores for each box.
iou_threshold (float): The IOU threshold for suppression.
Returns:
keep (Tensor): Tensor of shape (M,) representing the indices of the boxes to keep after NMS.
"""
# sort boxes by score
_, idxs = scores.sort(descending=True)
keep = []
while idxs.numel() > 0:
i = idxs[0]
keep.append(i)
if idxs.numel() == 1:
break
# calculate IOU with remaining boxes

j = idxs[1:]
iou_vals = iou(boxes[i:i+1], boxes[j])[0]
# filter out boxes with IOU above threshold

idxs = idxs[1:][iou_vals < iou_threshold]
return torch.tensor(keep)
42
II. SCREENSHOTS
43
44
45
46
REFERENCES
1. Annirudh D, Arun Kumar D, Adabala Taraka Sai Raghava Kumar, & K.R.M. Vijaya
Chandrakala. (2021). IoT based Intelligent Parking Management System. 2021 IEEE Second
International Conference on Control, Measurement and Instrumentation (CMI). Kolkata, India:
IEEE. doi:10.1109/CMI50323.2021.9362845
2. Athira A, Lekshmi S, Pooja Vijayan, & Boby Kurian. (2019). Smart Parking System Based On
Optical Character Recognition. 2019 3rd International Conference on Trends in Electronics
and Informatics (ICOEI). Tirunelveli, Inida: IEEE. doi:10.1109/ICOEI.2019.8862517
3. Chin-Kit Ng, Soon-Nyean Cheong, Erfan Hajimohammadhosseinmemar, & Wen-Jiun Yap.

(2017). Mobile outdoor parking space detection application. 2017 IEEE 8th Control and System
Graduate Research Colloquium (ICSGRC). Shah Alam, Malaysia: IEEE.
doi:10.1109/ICSGRC.2017.8070573
4. Daniel Padilla Carrasco, Hatem A. Rashwan, Miguel Ángel García, & Domènec Puig. (2021).
T-YOLO: Tiny Vehicle Detection Based on YOLO and Multi-Scale Convolutional Neural
Networks. IEEE Access. 11, pp. 22430 - 22440. IEEE.
doi:https://doi.org/10.1109/ACCESS.2021.3137638
5. Jae Kyu Suhr, & Ho Gi Jung. (2016). Automatic Parking Space Detection and Tracking for
Underground and Indoor Environments. IEEE Transactions on Industrial Electronics. 63, pp.
5687 - 5698. IEEE. doi:10.1109/TIE.2016.2558480
6. Jiahui Yu, Hongwei Gao, Jian Sun, Dalin Zhou, & Zhaojie Ju. (2021). Spatial Cognition-Driven
Deep Learning for Car Detection in Unmanned Aerial Vehicle Imagery. IEEE Transactions on
Cognitive and Developmental Systems. 14, pp. 1574 - 1583. IEEE.
doi:https://doi.org/10.1109/TCDS.2021.3124764
7. Li Wang, Xinyu Zhang, Weijia Zeng, Wei Liu, Lei Yang, Jun Li, & Huaping Liu. (2022).
Global Perception-based Robust Parking Space Detection Using a Low-cost Camera. IEEE
Transactions on Intelligent Vehicles (pp. 1 - 10). IEEE.
doi:https://doi.org/10.1109/TIV.2022.3186035
8. Long Qin, Yi Shi, Yahui He , Junrui Zhang, Xianshi Zhang, Yongjie Li, . . . Hongmei Yan.
(2022). ID-YOLO: Real-Time Salient Object Detection Based on the Driver’s Fixation Region.
IEEE Transactions on Intelligent Transportation Systems. 23, pp. 15898 - 15908. IEEE.
doi:https://doi.org/10.1109/TITS.2022.3146271
47
9. Matthias R. Schmid, S. Ates, J. Dickmann, F. von Hundelshausen, & H.-J. Wuensche. (2011).
Parking space detection with hierarchical dynamic occupancy grids. 2011 IEEE Intelligent
Vehicles Symposium (IV). Baden-Baden, Germany: IEEE. doi:10.1109/IVS.2011.5940476
10. Mohamed ElMikaty, & Tania Stathaki. (2017). Detection of Cars in High-Resolution Aerial
Images of Complex Urban Environments. IEEE Transactions on Geoscience and Remote
Sensing. 55, pp. 5913 - 5924. IEEE. doi:https://doi.org/10.1109/TGRS.2017.2716984
11. Nazia Bibi, Muhammad Nadeem Majid, Hassan Dawood, & Ping Guo. (2017). Automatic
Parking Space Detection System. 2017 2nd International Conference on Multimedia and Image
Processing (ICMIP). Wuhan, China: IEEE. doi:10.1109/ICMIP.2017.4
12. Qunying He, Jingjing Liu, & Zhicheng Huang. (2022). WSRC: Weakly Supervised Faster
RCNN Toward Accurate Traffic Object Detection. IEEE Access. 11, pp. 1445 - 1455. IEEE.
13. Rachapol Lookmuang, Krit Nambut, & Sasiporn Usanavasin. (2018). Smart parking using IoT
technology. 2018 5th International Conference on Business and Industrial Research (ICBIR).
Bangkok, Thailand: IEEE. doi:10.1109/ICBIR.2018.8391155
14. Shen-En Shih, & Wen-Hsiang Tsai. (2014). A Convenient Vision-Based System for Automatic
Detection of Parking Spaces in Indoor Parking Lots Using Wide-Angle Cameras. IEEE
Transactions on Vehicular Technology. 63, pp. 2521 - 2532. IEEE.
doi:10.1109/TVT.2013.2297331
15. Twinkle Singh, Safdar Sardar Khan, & Surendra Chadok. (2018). A Review on Automatic
Parking Space Occupancy Detection. 2018 International Conference on Advanced
Computation and Telecommunication (ICACAT). Bhopal, India: IEEE.
doi:10.1109/ICACAT.2018.8933644
16. Xiaoming Sun, Xinchun Jia, Yuqian Liang, Meigang Wang, & Xiaobo Chi. (2022). A Defect
Detection Method for a Boiler Inner Wall Based on an Improved YOLO-v5 Network and Data
Augmentation Technologies. IEEE Access. 10, pp. 93845 - 93853. IEEE.
17. Xiaoming Xi, Zhilou Yu, Zhaolei Zhan, Yilong Yin, & Cuihuan Tian. (2019). Multi-Task Cost-
Sensitive-Convolutional Neural Network for Car Detection. IEEE Access . 7, pp. 98061 -
98068. IEEE. doi:https://doi.org/10.1109/ACCESS.2019.2927866
18. Yunfeng Shao, Pengzhen Chen, & Tongtong Cao. (2018). A Grid Projection Method Based on
Ultrasonic Sensor for Parking Space Detection. IGARSS 2018 - 2018 IEEE International
48
Geoscience and Remote Sensing Symposium. Valencia, Spain: IEEE.
doi:10.1109/IGARSS.2018.8519022
19. Zhi Xu, Jingzhao Li, & Mei Zhang. (2021). A Surveillance Video Real-Time Analysis System
Based on Edge-Cloud and FL-YOLO Cooperation in Coal Mine. IEEE Access. 9, pp. 68482 -
68497. IEEE. doi:https://doi.org/10.1109/ACCESS.2021.3077499
20. Zhongxiang Zhou, Yifei Yang, Yue Wang, & Rong Xiong. (2023). Open-Set Object Detection
Using Classification-Free Object Proposal and Instance-Level Contrastive Learning. IEEE
Robotics and Automation Letters. 8, pp. 1691 - 1698. IEEE.
doi:https://doi.org/10.1109/LRA.2023.3242169
49
50

Car Parking Space Detection Using Yolo: A Project Report Submitted by

Uploaded by

Copyright:

Available Formats

Car Parking Space Detection Using Yolo: A Project Report Submitted by

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Car Parking Space Detection Using Yolo: A Project Report Submitted by

Uploaded by

Copyright:

Available Formats

CAR PARKING SPACE DETECTION

in partial fulfilment for the award of the degree

COMPUTER SCIENCE AND ENGINEERING

ERODE SENGUNTHAR ENGINEERING COLLEGE

PERUNDURAI, ERODE – 638 057

Submitted for End Semester Project Viva-Voce Examination held on _________

INTERNAL EXAMINER EXTERNAL EXAMINER

We are delighted to thank our esteemed President Thiru.V.ANNADURAI of Erode

We have immense pleasure to thank our Secretary

We would like to express our thanks to our Principal

We are grateful to our project coordinator Dr.S.TAMILSELVAN, M.E., Ph.D., Assistant

6 RESULT AND ANALYSIS 32

6.1 Model Performance with Motion Detection Results 33

6.2 Model Performance with No Motion Detection Results 33

5.1 System Architecture 24

5.2 Workflow Diagram 30

5.3 Data-flow Diagram 30

5.4 Sequence Diagram 31

5.5 Use Case Diagram 31

6.1 Evaluation Metrics 32

ABBREVIATION FULL FORM OF ABBREVIATION

AVM Around View Monitor

CLI Command-Line Interface

CNN Convolutional Neural Network

GPS Global Positioning System

GUI Graphical User Interface

IoT Internet of Things

IoU Intersection over Union

mAP Mean Average Precision

NMS Non-Maximum Suppression

OCR Optical Character Recognition

RANSAC Random Sample Consensus

RFID Radio Frequency Identification

RNN Recurrent Neural Network

SSD Single-Shot Detector

SVM Support Vector Machine

YOLO You Only Look Once

2.1 IoT based Intelligent Parking Management System

2.2 Smart Parking System Based on Optical Character Recognition

2.3 Smart Parking Using IoT Technology

2.6 Mobile Outdoor Parking Space Detection Application

2.9 A Convenient Vision-Based System for Automatic Detection of Parking

A dense representation of three-dimensional space would be costly, so a hierarchical data

In addition, we compared the research findings to a variety of methodologies in order to

3.1 Existing System

3.2 Proposed System

4.1 Hardware Requirements

CPU : At least 2 cores.

4.2 Software Requirements

Operating System : Windows 10 or Linux.

4.3 Tools and Technology

4.3.6 Jupyter Notebook

5.1 System Architecture

Fig. 5.1 System Architecture

5.2.2 Object Detection

5.2.4 User Interface

5.3 Proposed Methodology

5.3.2 Image Annotation

5.3.3 YOLO Architecture

def getitem(self, idx):