1

A Major Project Report on

“REAL-TIME ROAD DAMAGE DETECTION AND

GEOSPATIAL MAPPING USING YOLOv8”

Submitted in partial fulfilment of the requirement for Degree in Bachelor of

Engineering in Information Technology

By
Asavari Bhelawe (5021106)
Vedika Pagar (5021138)
Aman Singh (5021159)

Internal Guide: Dr. Archana Shirke

External Guide: Dr. Shashikant Dugad

Indian Institute of Science Education And Research (IISER), Mohali

Department of Information Technology

Fr. Conceicao Rodrigues Institute of Technology
(An Autonomous Institute & Permanently Affiliated to the University of Mumbai)
Sector 9A, Vashi, Navi Mumbai – 400703
University of Mumbai
2024-2025
 CERTIFICATE

This is to certify that the project entitled

“REAL-TIME ROAD DAMAGE DETECTION AND

GEOSPATIAL MAPPING USING YOLOv8”

Submitted By

Asavari S Bhelawe (5021106)

Vedika S Pagar (5021138)
Aman Singh (5021159)

In partial fulfilment of the degree of B.E. in Information Technology for term work
of the Semester 8 major project is approved.

____________ ____________
External Examiner Internal Examiner

____________ ____________
External Guide Internal Guide

____________ ____________
Head Of Department Principal

Date : College Seal

Place :
 DECLARATION

We declare that this written submission represents our ideas in our own words and where
others’ ideas or words have been included, We have adequately cited and referenced the
original sources. We also declare that we have adhered to all principles of academic
honesty and integrity and have not misrepresented fabricated or falsified any idea/data/-
fact/source in our submission. We understand that any violation of the above will be
cause for disciplinary action by the Institute and can also evoke penal action from the
sources that have thus not been properly cited or from whom proper permission has not
been taken when needed.

__________
Asavari S Bhelawe (5021106)

__________
Vedika S Pagar (5021138)

__________
Aman Singh (5021159)

Date :
Place :

iii
 ABSTRACT

Modern technology has revolutionized the monitoring of urban roads using various video
sources such as smartphones, car cameras, and surveillance systems. Focusing on roads in
India and Japan, this study presents a scalable deep learning- based system for the real-
time identification, categorization, and mapping of road damage. The solution addresses
challenges such as inconsistent image quality, diverse climatic conditions, and varying
regional infrastructures by utilizing the YOLO object detection algorithm, trained on
annotated datasets from the Japan Road Association and enhanced by data augmenta-
tion. The system improves efficiency by 75.5 percent, enabling faster inspections and
helping authorities prioritize repairs for safer roads. In India, approximately 40 percent
of total road accidents annually are caused by damaged road surfaces, including potholes,
cracks, and poor maintenance. These contribute to a significant number of both fatal
and non-fatal incidents, underlining the critical need for proactive monitoring systems.

A user interface is developed, featuring a main page with live real-time detection, a road
damage map, and report generation capabilities. These reports are provided to authorities
for efficient infrastructure management, optimizing resource allocation, and prioritizing
road repairs. This system significantly contributes to improving road maintenance pro-
cesses, benefiting both nations by streamlining road monitoring and decision-making. In
summary, your system could make road condition improvement processes 70-80 percent
more efficient and 50-95 percent faster in India compared to traditional methods of road
damage detection and recovery process.

Additionally, the system’s scalability enables deployment on edge devices, ensuring real-
time analysis with minimal latency and reduced dependency on high-performance com-
puting infrastructure. By leveraging smartphone-based image capture, the solution re-
mains cost-effective and accessible, making it feasible for large-scale implementation.
With these advancements, the system aims to transform road monitoring into a more ef-
ficient, data-driven, and automated process, ultimately leading to safer and more reliable
road networks.

iv
 CONTENTS

1 INTRODUCTION 1
1.1 BACKGROUND . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 MOTIVATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 PROBLEM DEFINITION . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 SCOPE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.5 AIM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.6 OBJECTIVES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.7 LIMITATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.8 APPLICATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 LITERATURE REVIEW 5
2.1 LITERATURE SURVEY . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 EXISTING SYSTEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 REQUIREMENT ANALYSIS . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.1 FUNCTIONAL REQUIREMENTS: . . . . . . . . . . . . . . . . . 9
2.3.2 NON-FUNCTIONAL REQUIREMENTS . . . . . . . . . . . . . . 9

3 SYSTEM DESIGN 11
3.1 ARCHITECTURAL DESIGN . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2 DATAFLOW DIAGRAM . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2.1 LEVEL 0 DFD . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2.2 LEVEL 1 DFD . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2.3 LEVEL 2 DFD . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 FLOW CHART . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4 WORKING OF SYSTEM . . . . . . . . . . . . . . . . . . . . . . . . . . 14

4 IMPLEMENTATION DETAILS 16
4.1 SYSTEM REQUIREMENTS . . . . . . . . . . . . . . . . . . . . . . . . 16
4.1.1 HARDWARE REQUIREMENTS:- . . . . . . . . . . . . . . . . . 16
4.1.2 SOFTWARE REQUIREMENTS:- . . . . . . . . . . . . . . . . . 16
4.2 METHODOLOGY USED . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3 ALGORITHM USED . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

v
 4.4 GANTT CHART . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.4.1 TIMELINE CHART SEMESTER 7 . . . . . . . . . . . . . . . . . 18
4.4.2 TIMELINE CHART SEMESTER 8 . . . . . . . . . . . . . . . . . 19

5 EXPERIMENTAL DETAILS 20
5.1 DATASET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2 EXPERIMENTAL RESULT . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.3 GUI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.4 RESULT ANALYSIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

6 CONCLUSION AND FUTURE SCOPE 28

6.1 CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
6.2 FUTURE SCOPE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

REFERENCES

APPENDIX A: CODE SAMPLE

ACKNOWLEDGMENT

PLAGIARISM REPORT

CONFERENCE PRESENTATION CERTIFICATES

PRESENTED RESEARCH PAPER

PCUBE PROJECT PRESENTATION CERTIFICATES

TECHSPARKS PROJECT COMPETITION CERTIFICATES

vi
 LIST OF FIGURES

Fig No. Figure Caption Page Number

3.1 Architectural Diagram of the system 11
3.2 DFD Level-0 12
3.3 DFD Level-1 13
3.4 DFD Level-2 13
3.5 Flowchart for Working of System 14
3.6 Working of System 15
4.1 YOLOv8 Architecture for Road Damage Detection 18
4.2 Gantt Chart for semester 7 19
4.3 Gantt Chart for semester 8 19
5.1 Statistics for the number of damage instances included 21
in the underlying datasets
5.2 Road Damage Detection using trained YOLOv8 model 22
5.3 Road damage photos and classes for a model training 23
5.4 Application Interface (a) Login Page (b) Main Page 24
5.5 Application Interface (a) Geo-Mapping of damages (b) 25
Report Accident
5.6 Loss Curves for YOLOv8 Model (a) Training Loss 26
Curves (b) Validation Loss Curves
5.7 Performance Curves (a) mAP50-95 Curve (b) mAP50 27
Curve

vii
 LIST OF TABLES

Table No. Table Caption Page Number

2.1 Literature Survey 6
2.2 Literature Survey 7
5.1 Road Damage Types 20

viii
 Chapter 1
INTRODUCTION

Efficient road maintenance is crucial for transportation safety and infrastructure sustain-
ability. Traditional manual inspections are labor-intensive, costly, and often inaccurate.
Factors like weather and aging contribute to road damage, increasing accident risks and
vehicle maintenance costs. This project leverages deep learning and computer vision to
automate road damage detection using crowdsourced smartphone images. By integrating
AI-driven analysis with geospatial mapping, our approach enhances inspection accuracy
and helps authorities prioritize repairs efficiently.

1.1 BACKGROUND

Urban infrastructure, particularly road networks, plays a vital role in a city’s economic
and social development. Roads facilitate transportation, commerce, and emergency re-
sponse, making their maintenance crucial for ensuring efficiency and safety. However, fac-
tors such as weather conditions, heavy traffic loads, and aging infrastructure contribute to
road deterioration, leading to increased vehicle maintenance costs and hazardous driving
conditions. Traditional road damage detection methods, including manual inspections
and vibration-based techniques, are often labor-intensive, expensive, and inefficient for
large-scale monitoring. While laser-scanning methods offer high accuracy, they require
significant financial investment and may cause traffic disruptions. Recent advancements
in deep learning and image processing have opened new possibilities for automated, cost-
effective, and scalable road damage detection, making AI-driven approaches a promising
alternative.

1.2 MOTIVATION

The need for an efficient, scalable, and cost-effective road damage detection system is
critical for modern cities striving to enhance infrastructure management. Poor road
conditions are a major contributor to accidents, injuries, and economic losses due to
increased vehicle maintenance and transportation inefficiencies. With governments in-
vesting billions in road maintenance, optimizing resource allocation is crucial for effective

1
infrastructure management. Image-based deep learning models, such as YOLO, offer
a viable solution by enabling automated detection of road damage using crowdsourced
images. By integrating AI-driven analysis with geospatial mapping and a user-friendly
interface, this project aims to provide real-time road condition monitoring, improve repair
prioritization, and assist authorities in making data-driven infrastructure decisions.

1.3 PROBLEM DEFINITION

This research aims to reduce the inaccuracy and inefficiency of conventional road inspec-
tion techniques. These manual procedures ignore important road damage and are costly
and time-consuming. Furthermore, it is not economical to deploy specialized vehicles
with cameras and sensors for advanced inspection techniques on a large scale. The goal
of this project is to address these problems by creating a deep learning-based system that
uses smartphone photos to automatically identify and categorize road damage, allowing
for real-time analysis and prompt restoration.

1.4 SCOPE

The scope of this project includes the development and implementation of a deep learning
system capable of detecting various types of road damage, including cracks, potholes,
and surface deformations, from smartphone-captured images. The system will be trained
on annotated datasets and optimized for real-time processing on edge devices such as
smartphones or low-power computing devices. The project also aims to evaluate the
model’s accuracy and scalability, ensuring that it can be deployed in diverse environmental
conditions and regions. Future work may expand the system to detect additional types of
infrastructure damage and optimize the model for integration with cloud-based systems
for large-scale use.

1.5 AIM

To design and implement an intelligent, real-time, and scalable road infrastructure as-
sessment system that automates the detection, classification, and prioritization of road
damages using advanced technologies such as machine learning, image processing, and
sensor-based analytics. This system aims to revolutionize traditional road inspection
methods by significantly improving data accuracy, reducing human intervention, opti-
mizing maintenance schedules, and enhancing road safety while minimizing repair costs.

2
1.6 OBJECTIVES

1. Improve Data Accuracy and Consistency: Minimize human error by replacing

manual road inspections with an automated system that provides consistent data.

2. Enable Real-Time Processing: Develop a real-time damage assessment solu-

tion to allow immediate identification and reporting of road defects.

3. Prioritize Critical Damages: Utilize data analytics and machine learning to

classify and prioritize road damages based on severity, ensuring timely and effective
maintenance.

4. Reduce Maintenance Costs and Improve Road Safety: Optimize road main-
tenance efforts, reduce repair costs, and enhance safety for road users by timely
identification and repair of road damages.

1.7 LIMITATIONS

Not with standing the possible advantages, this initiative has many drawbacks. Variations
in image quality caused by elements like weather, lighting, and camera specs may have
an impact on the detecting system’s accuracy. Furthermore, the system’s capacity to
process massive amounts of data rapidly may be limited by the computational capacity
of the edge devices employed for real-time processing. Furthermore, the model’s ability to
generalize to other areas or conditions may be limited by the training dataset’s probable
lack of representativeness of various road kinds and damage patterns.

1.8 APPLICATIONS

1. Real-Time Detection of Road Damage: The system uses YOLOv8 to in-

stantly detect and classify different types of road damage from video feeds. This
allows immediate identification of hazards without manual inspection.

2. Geospatial Mapping of Damage Locations: Detected road damages are geo-

tagged and visualized on an interactive map. This aids in monitoring affected areas
and planning targeted repair operations.

3. Automated Report Generation for Authorities: The system creates struc-

tured reports containing damage type, location, and severity. These reports help
road maintenance teams make informed decisions quickly.

4. Prioritization of Road Repairs: Based on detection frequency and damage

severity, the system ranks road segments needing urgent attention. This helps
authorities prioritize repairs for maximum safety impact.

3
5. Efficient Infrastructure Management: By automating detection and mapping,
the system reduces manual labor and inspection time. It improves the management
of road maintenance resources and timelines.

6. Efficient Infrastructure Management Enhanced Data-Driven Decision

Making: Accumulated detection data supports trend analysis and strategic plan-
ning. Authorities can track damage patterns and forecast maintenance needs effec-
tively.

4
 Chapter 2
LITERATURE REVIEW

2.1 LITERATURE SURVEY

Using computer vision and deep learning methods to automate the detection of road dam-
age has been the subject of numerous studies in recent years. Scholars have investigated
a number of techniques, including region-based methods like R-CNN and convolutional
neural networks (CNNs), to detect and categorize various forms of road damage, such as
potholes and cracks. The accuracy and efficiency of road inspections have been shown
to be improved by these techniques in comparison to manual methods. However, issues
remain, particularly with the scalability of these systems for real-time application on
low-power devices like smartphones and their adaptability to varied climatic conditions.
Many solutions also demand substantial processing resources, making them unfeasible for
large-scale or edge device deployment.

2.2 EXISTING SYSTEM

Several systems are currently in use for road inspection and damage detection, each
with its own strengths and limitations. These methods range from traditional manual
processes to more advanced, sensor-based approaches. Below are the details of the most
commonly used systems:

1. Manual Inspections: Traditionally, road inspections are conducted manually by

road maintenance crews who visually assess road conditions, identify damages like
cracks and potholes, and document them. This process is slow, labor-intensive, and
prone to human error, leading to inconsistent data and delayed repairs. Due to its
high cost and inefficiency, it is not scalable for large road networks.

2. Specialized Vehicles with Sensors: Advanced systems have been developed

where specialized vehicles equipped with cameras, LiDAR, and other sensors cap-
ture road surface data. These vehicles can detect and classify road damage using
image processing algorithms. However, the systems are costly to operate, require
trained personnel, and are not practical for continuous or large-scale monitoring.

5
 Table 2.1: Literature Survey

Title Details Methodology Strength Weakness

Road Computer SSD framework The ap- Manual
Damage Aided Civil that uses transfer proach is image an-
Detec- and Infras- learning to enhance highly prac- notation
tion tructure performance and a tical as it is time-
Using Engineer- pre-trained VGG-16 leverages consuming
Deep ing, 2021model for the CNN existing and labour-
Neural Hiroya architecture. To in- smartphone intensive.
Net- Maeda, crease the resilience technology, The sys-
works Yoshihide of the model, data making tem’s perfor-
with Sekimoto, augmentation tech- it cost- mance was
Images Toshikazu niques like rotation, effective constrained
Cap- Seto, random cropping, and easily by the com-
tured Takehiro and color modifica- deployable. putational
Through Kashiyama, tions were used. To power of the
a Hiroshi assess how well the smartphone,
Smart- Omata model performed, affecting
phone the dataset was real-time
[1] divided into training processing
and testing sets. capabilities.
Real- International Following the anno- Mask R- The frame-
Time Journal of tation of images to CNN’s work might
Road Advanced identify the areas of capability face chal-
Surface Computer various forms of road to provide lenges
Damage Science damage, the model pixel-level related to
Detec- and Ap- was pre-trained on segmenta- processing
tion plications the COCO dataset tion en- latency
Frame- Vol. 14, and refined on the hances both and com-
work No. 9, 2023 annotated road detection ac- putational
Based Bakhytzhan damage dataset. curacy and demands,
on Mask Kulam- Mask R-CNN, which the ability especially
R-CNN bayev, builds upon Faster to classify when imple-
[2] Magzat R-CNN by including different mented in
Nurlybek, a branch for pre- types of real-world
Gulnar As- dicting segmentation road dam- scenarios
taubayeva, masks, was used to age. with vary-
Gulnara identify and catego- ing image
Tleu- rize the damage. quality
berdiyeva, and envi-
ronmental
conditions.

Additionally, these systems often lack real-time processing capabilities, delaying

damage detection and repair.

3. Crowdsourced Applications: By taking pictures and reporting them to road

6
 Table 2.2: Literature Survey

Title Details Methodology Strength Weakness

YOLO- EURASIP 1)Dataset Expan- Improved Dataset
LRDD: Journal on sion: Added Chinese Efficiency: Imbalance:
a light Advances Road samples to the 22.3 faster Struggles
weight in Signal RDD2020 dataset. and 28.8 with unbal-
method Processing 2)YOLO-LRDD smaller than anced defect
for road 2022 Sadra Model:Backbone: YOLOv5s. sample sizes.
damage Naddaf-sh, Improved efficiency Enhanced Special-
detec- M-Mahdi with a lightweight Accuracy: ized Scope:
tion Naddaf-Sh, [Link]: En- Lightweight Mainly
based Amir R. hanced feature fusion and suitable tailored
on im- Kashani, for better detection. for mobile for Chi-
proved Hassan 3)Training Process: use while nese roads,
YOLOv Zargarzadeh Used advanced data maintaining limiting gen-
5s [3] augmentation and accuracy. eralizability.
adaptive techniques Robust De-
for better training. tection: Bal-
anced and
precise de-
tection with
efficient net-
works and
advanced
loss func-
tions.
Road Comput. The YOLOv8-PD The model Despite its
Damage Aided Civ. model is designed combines strengths,
Detec- Infrastruc- for detecting pave- Transformer the model
tion and ture Eng. ment distress types, and con- may face
Classi- (CACAIE), including longitudi- volutional challenges
fication 2023 H. nal and transverse techniques balancing
Using Maeda, Y. cracks, mesh cracks, to improve speed and
Deep Sekimoto, and [Link] detection accuracy
Neural T. Seto, T. C2fGhost block to accuracy in more
Net- Kashiyama, optimize perfor- and effi- complex en-
works and H. mance and reduce ciency. Its vironments.
with Omata computational costs, lightweight While it’s
Smart- making it suitable design, fea- optimized
phone for real-time applica- turing the for effi-
Im- tions. C2fGhost ciency.
ages[4] block, allows
real-time
performance
on edge
devices.

7
 maintenance departments, drivers can report road damage using some of the cur-
rent techniques, like crowdsourced mobile applications. For systematic road damage
evaluation, this method is inconsistent and unreliable because it still requires man-
ual input and lacks automated detection, despite being more scalable.

4. Satellite Imaging: Some municipalities use satellite images to detect changes in

road surfaces over time. While satellite imaging can cover large areas, it lacks the
resolution necessary for detecting small-scale road damage like cracks or potholes.
Additionally, satellite imaging is costly and not suitable for frequent monitoring

8
2.3 REQUIREMENT ANALYSIS

Requirement analysis is a pivotal phase in project development, aimed at comprehensively

defining the project’s needs and constraints to ensure alignment with desired objectives.
The requirement analysis is outlined as follows:

2.3.1 FUNCTIONAL REQUIREMENTS:

• Image Capture: The system should allow users to capture road surface images
through a smartphone camera. The image quality should meet the minimum reso-
lution required for accurate damage detection.

• Data Processing:The system should preprocess the captured images (e.g., resiz-
ing, normalization) before passing them to the deep learning model for analysis.
The system should handle image input in real-time or batch mode.

• Damage Classification:The system should be able to classify road damage into

predefined categories, such as minor cracks, major cracks, and potholes, based on
the analysis of the captured image. optimizing resource management.

• Data Storage: The system should store the results of road damage detection in
a database, including the image, detected damage, location, and timestamps for
future reference and analysis.

• Alert Generation: The system should generate real-time alerts or notifications

to relevant authorities or users when significant damage is detected that requires
immediate attention.

2.3.2 NON-FUNCTIONAL REQUIREMENTS

• Performance Efficiency: The system must exhibit high operational efficiency,

ensuring near real-time processing of road surface images. It should guarantee a
maximum response time of 3-5 seconds per image under normal load, with minimal
latency, even when processing multiple images in parallel.

• Data Security and Privacy: For data transfer and storage, the system must
use industry-standard encryption (such AES-256) to protect sensitive information,
such as geo-located photos. Role-based access control (RBAC) and multi-factor
authentication (MFA) ought to be implemented in order to safeguard user accounts
and system access from unauthorized parties.

9
• Maintainability and Modularity: The system should adopt a modular and
loosely coupled design, allowing for easy enhancements, bug fixes, and feature up-
dates without impacting existing functionality. It should support continuous inte-
gration and deployment (CI/CD) pipelines to ensure seamless, automated updates.

• Platform Compatibility: The solution should be fully compatible across multiple

mobile and desktop platforms, including Android and iOS. It should ensure seamless
operation regardless of device specifications (camera resolution, processor speed),
supporting various image formats and resolutions for optimal flexibility.

• Detection Accuracy and Precision: The deep learning models deployed must
maintain a high level of detection precision, targeting at least 90percent accuracy
in classifying various types of road damage. The system should support adap-
tive learning capabilities to refine its detection performance as more data becomes
available, minimizing false positives and negatives.

• Elastic Scalability for Cloud Deployments:The system should be designed for

cloud-native deployment, supporting elastic scaling of resources based on real-time
demand. It should leverage containerization (e.g., Docker) and orchestration plat-
forms (e.g., Kubernetes) for efficient resource utilization and workload distribution.

10
 Chapter 3
SYSTEM DESIGN

3.1 ARCHITECTURAL DESIGN

The road damage detection system’s architecture is organized into a multi-tiered frame-
work that guarantees effective image processing, analysis, and capture. The User Interface
Layer, which is at the forefront, includes a mobile application that works with both iOS
and Android smartphones and makes it simple for users to take pictures of road surfaces
and send them in for analysis. By giving access to past data and feedback on detection
results, this layer enables real-time involvement. The mobile application and the backend
processing components may communicate easily thanks to APIs included in the Applica-
tion Layer, which also controls user interactions. The Processing Layer, which contains
the deep learning model in charge of processing collected images, lies at the heart of the
architecture. Before implementing the model inference utilizing sophisticated architecture
like YOLO to detect and categorize different kinds of road damage, this layer includes
image preprocessing operations including scaling, normalization, and augmentation.

Figure 3.1: Architectural Diagram of the system

Data Collection: The system collects data in the form of images, typically captured by
cameras mounted on vehicles or drones.

Preprocessing: The collected images are preprocessed to handle noise, inconsistencies,

and distortions to ensure the quality and accuracy of further analysis.

11
Feature Extraction: Relevant features and patterns, such as cracks, potholes, and surface
anomalies, are extracted from the preprocessed images.

Data Splitting: The data is separated into training and testing sets. The model is trained
on the training set, and its performance is assessed on the testing set.
Model Selection (YOLO): The YOLO (You Only Look Once) object detection algorithm
is selected to identify road damage from the images.

Evaluation: The trained YOLO model is evaluated using the test data to measure its
accuracy in detecting road damage.

Road Damage Analysis: The system analyzes the identified road damage based on the
results from the YOLO model, classifying the severity and type of damage.

Result Visualization: The results are visualized, typically through dashboards or reports,
for stakeholders to review. This can include graphical representations of road conditions
and identified damages.

3.2 DATAFLOW DIAGRAM

3.2.1 LEVEL 0 DFD

The Level 0 DFD shows the overall system where a user uploads an image, which is
processed by the application to extract text or features. The output is saved in the file
system. It gives a high-level view of the system’s main function.

Figure 3.2: DFD Level-0

3.2.2 LEVEL 1 DFD

The Level 1 DFD breaks down the system into smaller parts like image upload, damage
detection, geolocation tagging, data storage, and report generation. It shows how data

12
moves between these processes and the related databases.

Figure 3.3: DFD Level-1

3.2.3 LEVEL 2 DFD

The Level 2 DFD focuses on the detailed steps inside the damage detection process.
It includes image preprocessing, loading the YOLOv8 model, detecting and classifying
damage, and saving the results. It gives a deeper look into how detection works.

Figure 3.4: DFD Level-2

13
3.3 FLOW CHART

Figure 3.5: Flowchart for Working of System

The flowchart describes an automated road inspection process using the YOLO object
detection algorithm. The system starts by initializes. It then captures images of the road,
which are preprocessed to remove noise and prepare them for further analysis. The YOLO
object detection algorithm is applied to these images to detect any potential road damage.
After processing, the system analyzes the road conditions. If no damage is found, the
system continues monitoring. However, if damage is detected, the system immediately
sends notifications to the authorities responsible for road maintenance, ensuring timely
repairs. This cycle repeats continuously to ensure ongoing road inspection.

3.4 WORKING OF SYSTEM

The system illustrates an end-to-end road damage detection and reporting framework.
It begins with users capturing road images using the RoadX mobile application, which

14
 Figure 3.6: Working of System

then communicates with a FastAPI-based backend server. The server processes the image
through a road damage detection model that identifies and highlights damages such as
potholes or cracks using bounding boxes. Simultaneously, the app captures the GPS
location of the image, which, along with the detection results, is stored in a centralized
database. This data can be accessed by users for awareness and by administrators who
receive compiled reports for maintenance planning and decision-making. The system
ensures real-time, location-based monitoring of road conditions for efficient infrastructure
management.

15
 Chapter 4
IMPLEMENTATION DETAILS

4.1 SYSTEM REQUIREMENTS

This section includes all the details of how the proposed system is implemented and the
minimum system requirements for the project to run smoothly

4.1.1 HARDWARE REQUIREMENTS:-

1. User Devices:

• Smartphones (Android and iOS) with a minimum camera resolution of 12 MP

for capturing high-quality images.

2. Server Requirements

• Processor: Quad-core CPU

• Memory: Minimum of 16 GB RAM.
• Storage: SSD with at least 500 GB capacity

4.1.2 SOFTWARE REQUIREMENTS:-

1. Mobile Application:

• Required Libraries: TensorFlow Lite

• Development Framework: React Native, Flutter, or native development (Java/Kotlin
for Android, Swift for iOS).

2. Backend Development:

• Programming Language: Python, [Link], or Java.

• Frameworks: Flask or Django (for Python) to handle API requests and re-
sponses.
• Deep Learning Libraries: TensorFlow, PyTorch

16
4.2 METHODOLOGY USED

The methodology for the project encompasses the approach, tools, and techniques em-
ployed to design, develop, and implement the system. The methodology involves several
key phases:

1. Data Collection: Gather a diverse dataset of road images capturing various types
of damage, ensuring adequate representation of different conditions, lighting, and
environments. This may involve collecting images from public datasets or conduct-
ing field surveys.

2. Data Preprocessing: Clean and preprocess the collected images, including re-
sizing, normalization, and augmentation techniques to enhance model robustness.
Label the dataset with appropriate annotations for damage types.

3. Model Development: Choose a suitable deep learning architecture (e.g., YOLOv5,

R-CNN) for road damage detection. Train the model using the prepared dataset,
tuning hyperparameters for optimal performance.

4. Model Evaluation:Use metrics like precision, recall, F1-score, and mean Average
Precision (mAP) to assess the model’s performance. To make sure the model is
generalizable across many datasets, use cross-validation.

5. System Integration: Integrate the trained model into the application layer, en-
suring seamless communication between the mobile app and the processing backend.
Implement APIs to handle image submissions and return detection results.

6. Testing: Perform comprehensive system testing, including user acceptability, in-

tegration, and unit testing. Get end-user input to determine what needs to be
improved.

7. Deployment: Deploy the application on relevant app stores and set up the backend
server or cloud infrastructure for production use. Monitor system performance and
user feedback for future enhancements.

8. Maintenance and Updates:Keep an eye on user interactions and system perfor-

mance at all times, and adjust iteratively in response to feedback and fresh data.
To keep the model accurate, retrain it frequently using updated datasets.

4.3 ALGORITHM USED

YOLO (You Only Look Once) is a single-stage object detection model that si-
multaneously predicts bounding boxes and class probabilities, making it a highly

17
 Figure 4.1: YOLOv8 Architecture for Road Damage Detection

efficient and fast solution for real-time applications. Unlike two-stage models like
Faster R-CNN, YOLO divides the input image into a grid, where each grid cell
generates anchor boxes that predict object presence, bounding box coordinates,
and class [Link] there have been significant breakthroughs in de-
tection approaches, real-time processing demands continue to be a major challenge.
Due to the operational requirements of road maintenance, damage identification
must be done promptly in addition to accurately. Real-time object detection was
emphasized by architectures like YOLO (You Only Look Once) and SSD (Single
Shot MultiBox Detector), as explained by researchers like [17–18]. Even while these
frameworks aren’t specifically designed for road anomalies, their fundamental ideas
offer priceless insights. They draw attention to the complex trade-offs and balance
between accuracy and detection speed. These factors are crucial when imagining
a model that functions in dynamic real-world environments, highlighting the ne-
cessity of any potential road damage detection system to have the ideal balance of
accuracy and speed.

4.4 GANTT CHART

4.4.1 TIMELINE CHART SEMESTER 7

The Gantt chart outlines the initial stages of the project from July to October
2024, covering Planning, Requirement Gathering, and Design. It includes key ac-
tivities such as domain selection, problem research, abstract formation, and system
architecture development. These phases establish the foundation for the project’s
technical direction and deliverables.

18
 Figure 4.2: Gantt Chart for semester 7

4.4.2 TIMELINE CHART SEMESTER 8

Figure 4.3: Gantt Chart for semester 8

The chart illustrates the execution, evaluation, and documentation stages of the
project, scheduled from November 2024 to March 2025. It captures the develop-
ment of the YOLOv8-based detection model, integration with a GUI/web interface,
evaluation using performance metrics, and final report and paper submission. These
phases are critical for implementing, assessing, and presenting the project outcomes.

19
 Chapter 5
EXPERIMENTAL DETAILS

5.1 DATASET

The RDD2020 dataset[19] is a crucial resource for developing smartphone-based road

damage detection systems. It provides an affordable solution for road condition moni-
toring, making it valuable for municipalities and road agencies. This dataset facilitates
the creation of new deep convolutional neural network (CNN) architectures and the en-
hancement of existing ones to improve detection algorithms. It includes road images from
multiple countries, such as India, Japan, and the Czech Republic, and offers opportuni-
ties for expansion to other regions. The dataset is designed to support the detection
and classification of various road damage types, including longitudinal cracks, transverse
cracks, alligator cracks, and potholes, with potential for incorporating more categories in
the future. Researchers can use RDD2020 for benchmarking the performance of machine
learning algorithms in image classification and object detection tasks. Furthermore, it
played a significant role in the Global Road Damage Detection Challenge (GRDDC 2020),
part of the IEEE Big Data Cup, which evaluated the effectiveness of road damage detec-
tion models. The dataset contains annotations for four types of road damage(Table II):
longitudinal cracks (D00), transverse cracks (D10), alligator cracks (D20), and potholes
(D40).

Table 5.1: Road Damage Types

The RDD2020 image dataset comprises 26,336 road images from India, Japan, and
the Czech Republic, representing over 31,000 instances of road damage. The dataset is

20
divided into training, test1, and test2 subsets. The training set includes subdirectories
for India, Japan, and the Czech Republic, with images and annotations specific to each
country. Images from Japan and the Czech Republic have resolutions of 600 × 600
pixels, while those from India are 720 × 720 pixels. The test1 and test2 subsets follow
the same resolution patterns and contain images from all three countries. Specifically, the
test1 subset includes 1,313 images from Japan, 969 from India, and 349 from the Czech
Republic, while the test2 subset contains 1,314, 990, and 360 images from the respective
countries.

Figure 5.1: Statistics for the number of damage instances included in the underlying
datasets

The Figure 1 shows how the dataset is distributed among the different countries. The
graph illustrates the distribution of dataset components across Japan, India, and Czech,
focusing on train images, test images, total images, and train labels. Japan has the
highest number of train images (10,506), test images (2,627), and total images (13,133),
followed by India with 7,706 train images, 1,959 test images, and 9,665 total images, while
Czech has the least in all categories (2,829 train, 709 test, and 3,538 total). Similarly,
train labels are most abundant in Japan (16,470), compared to India (6,831) and Czech
(1,745). This distribution highlights Japan as the dominant dataset contributor, while

21
Czech has the smallest dataset, affecting model training balance. The combined dataset
from these three countries is used for training, testing, and analyzing the performance of
the road damage detection model, ensuring its applicability across different regions and
road conditions.

5.2 EXPERIMENTAL RESULT

The road damage detection system is implemented using the YOLO (You Only Look
Once) architecture, which effectively identifies and classifies various types of road damage,
such as potholes and cracks, in real time. The initial results indicate satisfactory accuracy,
facilitating timely feedback for maintenance decisions. To further enhance detection
precision, an additional model focusing on improving accuracy will be integrated into the
system. Metrics like precision, recall, F1-score, and mean Average Precision (mAP) will
be used to compare the performance of this new model and the YOLO implementation.
The goal is to create a robust system that balances speed and accuracy, providing valuable
insights for efficient road maintenance and management.

Figure 5.2: Road Damage Detection using trained YOLOv8 model

Figure 5.2 compares the model’s performance on unannotated and annotated frames
for three road damage types. The first row detects a transverse crack (Confidence: 0.64)
in yellow. The second row highlights an alligator crack (Confidence: 0.81). The third
row identifies a pothole (Confidence: 0.79) in green. The annotations demonstrate the

22
model’s effectiveness in detecting and classifying road damages.

Figure 5.3: Road damage photos and classes for a model training

Figure 5.3 showcases the detection and classification of various road damages using
the proposed model. Subfigure (a) identifies a longitudinal crack (D00) in red, while (b)
highlights potholes (D40) in green. Subfigure (c) presents a transverse crack (D01) in
yellow, and (d) detects multiple potholes (D40), demonstrating the model’s ability to
identify multiple instances. In (e), both a transverse crack (D10) and an alligator crack
(D20) are marked in yellow. Subfigure (f) detects a transverse crack (D01) on a straight
road section. The bounding boxes, labels, and confidence scores highlight the model’s
accuracy in classifying different road damage types.

23
5.3 GUI

Using React Native, a sophisticated application for identifying road damage was created,
combining several pages to provide a thorough and intuitive user experience. The Main
Page(Fig.5.4 (a)) serves as the main navigation hub, giving users easy access to impor-
tant features such a real-time damage detection module, a manual reporting interface
for users to submit problems, and a comprehensive map that highlights discovered road
damages. The application’s user-friendly layout makes it simple for users to navigate,
which expedites the road monitoring and management process. Real-time road damage
detection is made possible by the Live Real-Time Detection Screen(Fig.5.4 (b)), which
provides state-of-the-art functionality. It gives consumers trustworthy and useful infor-
mation by precisely identifying the kind of damage, such as potholes, cracks, or faded
road markings, and by displaying confidence ratings for each detection. The highest level
of accuracy is guaranteed by this real-time feedback, enabling stakeholders to prioritize
road maintenance and repairs with knowledge. The application is a major advancement
in using technology for proactive road monitoring and safety improvement because of its
integration of cutting-edge capabilities and user-friendly design.

(a) Login Page (b) Main Page

Figure 5.4: Application Interface

Using React Native, a sophisticated application for identifying road damage was cre-

24
ated, combining several pages to provide a thorough and intuitive user experience. The
Login Page(Fig.5.4 (a)) of the RoadX application allows users to enter their email, pass-
word, and role (e.g., Citizen) to access the system for reporting road issues..The Main
Page(Fig.5.4 (b)) serves as the main navigation hub, giving users easy access to impor-
tant features such a real-time damage detection module, a manual reporting interface
for users to submit problems, and a comprehensive map that highlights discovered road
damages. The application’s user-friendly layout makes it simple for users to navigate,
which expedites the road monitoring and management process.

(a) Geo-Mapping of damages (b) Report Accident

Figure 5.5: Application Interface

Real-time road damage detection is made possible by the Live Real-Time Detec-
tion Screen(Fig.5.5 (a)), which provides state-of-the-art functionality. It gives consumers
trustworthy and useful information by precisely identifying the kind of damage, such as
potholes, cracks, or faded road markings, and by displaying confidence ratings for each
detection. The highest level of accuracy is guaranteed by this real-time feedback, enabling
stakeholders to prioritize road maintenance and repairs with knowledge. The application
is a major advancement in using technology for proactive road monitoring and safety im-
provement because of its integration of cutting-edge capabilities and user-friendly design.

25
5.4 RESULT ANALYSIS

The training and evaluation was carried out using the dataset. The following results were
observed Figure (5.6 (a) & (b)) illustrates the training and validation loss curves for the

(a) Training Loss Curves (b) Validation Loss Curves

Figure 5.6: Loss Curves for YOLOv8 Model

model over 50 epochs. The chart on the left represents the training losses, while the chart
on the right shows the validation losses. Three distinct components of the loss are plot-
ted: Box Loss (blue), Class Loss (red), and DFL Loss (green). These losses demonstrate
the gradual reduction in error as the training progresses, indicating effective learning
by the model. In the training loss curve (Fig 5.6(a)), all three components exhibit a
significant decrease during the initial epochs, followed by a slower but steady decline as
the model converges. Notably, the Class Loss starts at a higher value compared to the
other losses but consistently reduces, eventually stabilizing below 1.6. Similarly, Box Loss
and DFL Loss also exhibit a downward trend, stabilizing at approximately 1.5 and 1.4,
respectively. In the validation loss curve (Fig 5.6(b)), a similar pattern is observed, with
all loss components initially fluctuating before stabilizing. The Class Loss shows slightly
higher values compared to the training curve, peaking around epoch 5 before gradually
decreasing. The Box Loss and DFL Loss follow a smoother decline, indicating consis-
tency between training and validation phases. These curves highlight the model’s ability
to generalize effectively, as evidenced by the convergence of training and validation losses.
However, the minor fluctuations in the validation losses suggest opportunities for further
optimization to reduce overfitting or variance.

The performance metric curves (Fig. 5.7) provide further evidence of the model’s suc-
cess. The mAP@50 (Mean Average Precision at an IoU threshold of 0.5) demonstrates a

26
 (a) mAP50-95 Curve (b) mAP50 Curve

Figure 5.7: Performance Curves

consistent upward trend, showcasing improved alignment between the model’s predictions
and ground truth. This metric, which evaluates the balance between precision and recall,
reflects the model’s growing accuracy in detecting objects. Similarly, the mAP@50-95, a
more stringent metric averaging precision across a range of IoU thresholds (0.5 to 0.95),
also increases steadily, emphasizing the robustness of the model’s performance under
varying levels of overlap criteria. Overall, the model achieved a mean Average Precision
(mAP50) of 0.547 and mAP50-95 of 0.254 across all damage categories, demonstrating
robust performance in real-world conditions.

27
 Chapter 6
CONCLUSION AND FUTURE SCOPE

This section includes the Conclusion of the project through various Findings, Research
and the Success in implementing the desired project, the Future Scope of this project
and how this project can be made more dynamic.

6.1 CONCLUSION

The road damage detection system leverages the YOLO (You Only Look Once) architec-
ture for real-time identification and classification of road damage types with promising
initial accuracy. Future advancements aim to refine detection algorithms to handle di-
verse conditions and surfaces, enhance precision with additional models like semantic or
instance segmentation, and establish a comprehensive road condition database. Collabo-
rating with maintenance authorities and conducting field tests will validate its real-world
effectiveness, paving the way for smarter, data-driven infrastructure management solu-
tions.

6.2 FUTURE SCOPE

There are a number of sectors that need further research and development in the fu-
ture. Future advancements will involve improving the identification algorithms’ accuracy
and resilience to a range of road surfaces and environmental conditions. Putting other
deep learning models into practice, including those that concentrate on instance or se-
mantic segmentation, might yield more information about the features of road damage.
More educated maintenance choices may also be made possible by the incorporation of
an extensive database of road conditions. Field testing and ongoing cooperation with
road maintenance authorities will also be essential to confirming the system’s efficacy in
practical uses. The ultimate goal of this field’s continued study and development is to
produce more intelligent, data-driven infrastructure management systems.

28
 REFERENCES

[1] Road Damage Detection Using Deep Neural Networks with Images Captured Through
a Smartphone Hiroya Maeda , Yoshihide Sekimoto, Toshikazu Seto, Takehiro Kashiyama,
Hiroshi Omata University of Tokyo, 4-6-1 Komaba, Tokyo, Japan
[2] Kulambayev, Bakhytzhan Nurlybek, Magzat Astaubayeva, Gulnar Tleuberdiyeva,
Gulnara Zholdasbayev, Serik Tolep, Abdimukhan. (2023). Real-Time Road Surface
Damage Detection Framework based on Mask R-CNN Model. International Journal of
Advanced Computer Science and Applications. 14. 10.14569/IJACSA.2023.0140979.
[3] Wan, F., Sun, C., He, H. et al. YOLO-LRDD: a lightweight method for road damage
detection based on improved YOLOv5s. EURASIP J. Adv. Signal Process. 2022, 98
(2022). [Link]
[4] H. Maeda, Y. Sekimoto, T. Seto, T. Kashiyama, and H. Omata, “Road Damage
Detection and Classification Using Deep Neural Networks with Smartphone Images,”
Comput. Aided Civ. Infrastructure Eng. (CACAIE), vol. 33, no. 12, pp. 1127–1141,
2023.
[5] J. Redmon and A. Farhadi, “Yolo9000: Better, faster, stronger,” in The IEEE Confer-
ence on Computer Vision and Pattern Recognition (CVPR),2022, pp. 6517–6525
[6] Alexey Bochkovskiy, Chien-Yao Wang and Hong-Yuan Mark Liao,“YOLOv4: Optimal
Speed and Accuracy of Object Detection,” 2020. [Online]. Available: arXiv:2004.10934
[7] K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale
Image Recognition,” 2019. [Online]. Available:arXiv:1409.1556.
[8] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,Z. Huang, A. Karpathy,
A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei, “ImageNet Large Scale Visual
Recognition Challenge,” International Journal of Computer Vision (IJCV), vol. 115, no.
3, pp.211–252, 2023.
[9]Smita Rukhande, Prachi G, Archana S, Dipa D, “Implementation of GPS enabled
carpooling System” at International Journal of Advances in Engineering and Technol-
ogy(IJAET) November 2011 ISSN: 2231-1963.
[10]K. He, X. Zhang, S. Ren and J. Sun, “Deep Residual LearningforImage Recognition,”
2020 IEEE Conference on Computer Vision andPattern Recognition (CVPR), Las Vegas,
NV, 2016, pp. 770-778,doi:10.1109/CVPR.2020.90.
[11] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-timeobject
detection with region proposal networks,” in Advances in neural information processing
systems, 2019, pp. 91–99.
[12]R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich FeatureHier-archies for Accu-
rate Object Detection and Semantic Segmentation,”in2020 IEEE Conference on Computer
Vision and Pattern Recogni-tion(CVPR), 2020, vol. 0, pp. 580–587.
[13] Abushakra, A., Subhi, H., and Abdelkarim, S. (2023). Real-Time Road Surface
Damage Detection Framework Using Deep Learning. International Journal of Advanced
Computer Science and Applications (IJACSA), 14(9), 641-649.
[14] M. Maniat, C. Camp and A. Kashani, —Deep learning-based visual crack detection
using Google Street View images,‖ Neural Computing andApplications, vol. 33, no. 21,
pp. 14565-14582, 2021.
[15] J. Y. Lee, H. Lee, and Y. H. Kim, "Deep road crack detection with mobile applica-
tion," 2019 IEEE International Conference on Consumer Electronics (ICCE), Las Vegas,
NV, USA, 2019, pp. 354-359, doi: 10.1109/ICCE.2019.8621899.
[16] D. Quang and S. Bae. "A hybrid deep convolutional neural network approach for
predicting the traffic congestion index," Promet-Traffic and Transportation, vol. 33, no.
3, pp. 373-385, 2021.
[17] N. Safaei, O. Smadi, B. Safaei and A. Masoud, —Efficient road crack detection based
on an adaptive pixel-level segmentation algorithm,‖ Transportation Research Record, vol.
2675, no. 9, pp. 370-381, 2021.
[18] Maeda, Hiroya Sekimoto, Yoshihide Seto, Toshikazu Kashiyama, Takehiro Omata,
Hiroshi. (2018). Road Damage Detection Using Deep Neural Networks with Images
Captured Through a Smartphone. Arxiv. 10.48550/arXiv.1801.09454.
[19] V. Tran, T. Tran, H. Lee, K. Kim, J. Baek et al., —One stage detector (RetinaNet)-
based crack detection for asphalt pavements considering pavement distresses and surface
objects,‖ Journal of Civil Structural Health Monitoring, vol. 11, no. 1, pp. 205-222, 2021
[20] Z. Lingxin, S. Junkai and Z. Baijie, —A review of the research and application of deep
learning-based computer vision in structural damage detection,‖ Earthquake Engineering
and Engineering Vibration, vol. 21, no. 1, pp. 1-21, 2022.
[21] F. Andika and Y. Bandung, "Road Damage Classification using SSD Mobilenet with
Image Enhancement," 2023 International Conference on Computer Science, Information
Technology and Engineering (ICCoSITE), Jakarta, Indonesia, 2023, pp. 540-545, doi:
10.1109/ICCoSITE57641.2023.10127763.
 APPENDIX A: CODE SAMPLE

from xml . dom import minidom

import bs4 as bs
import os
from p a t h l i b import Path
import g l o b
from tqdm import tqdm
import random
import s h u t i l

def convertPascal2YOLOv8 ( f i l e P a t h ) :

class_mapping = {
"D00" : 0 ,
"D10" : 1 ,
"D20" : 2 ,
"D40" : 3 ,
"D01" : 4 ,
"D11" : 5 ,
"D43" : 6 ,
"D44" : 7 ,
"D50" : 8
}

f i l e = open ( f i l e P a t h , " r " )

c o n t e n t s = f i l e . read ( )
soup = bs . B e a u t i f u l S o u p ( c o n t e n t s , ’ xml ’ )
image_size = soup . f i n d _ a l l ( " s i z e " ) [ 0 ]
image_width = int ( image_size . f i n d _ a l l ( " width " ) [ 0 ] . g e t _ t e x t ( ) )
image_height = int ( image_size . f i n d _ a l l ( " h e i g h t " ) [ 0 ] . g e t _ t e x t ( ) )
o b j e c t s = soup . f i n d _ a l l ( " o b j e c t " )
bounding_box_list = [ ]
 class_list = []

for object in o b j e c t s :
_ c l a s s = object . f i n d _ a l l ( "name" ) [ 0 ] . g e t _ t e x t ( )
_ c l a s s = class_mapping . g e t ( _class , 1 0)
c l a s s _ l i s t . append ( _ c l a s s )

_xmin = f l o a t ( object . f i n d _ a l l ( "xmin" ) [ 0 ] . get_text ( ) )

_ymin = f l o a t ( object . f i n d _ a l l ( "ymin" ) [ 0 ] . get_text ( ) )
_xmax = f l o a t ( object . f i n d _ a l l ( "xmax" ) [ 0 ] . get_text ( ) )
_ymax = f l o a t ( object . f i n d _ a l l ( "ymax" ) [ 0 ] . get_text ( ) )
w = (_xmax − _xmin )
h = (_ymax − _ymin )
cx = _xmin + (w/2 )
cy = _ymin + ( h / 2)

w = round ( (w / image_width ) , 4 )
h = round ( ( h / image_height ) , 4 )
cx = round ( ( cx / image_width ) , 4 )
cy = round ( ( cy / image_height ) , 4 )

_bbox = [ cx , cy , w, h ]
bounding_box_list . append ( _bbox )

outputFilename = os . path . s p l i t ( f i l e P a t h ) [ 1 ]
outputFilename = outputFilename . r e p l a c e ( " . xml" , " . t x t " )

o u t pu tD ir = Path ( f i l e P a t h ) . p a r e n t s [ 2 ]
o u t pu tD ir = o ut put Dir / " l a b e l s "
i f not os . path . e x i s t s ( out put Di r ) :
os . makedirs ( o utp utD ir )

outputPath = out put Di r / outputFilename

with open ( outputPath , ’w ’ ) as f :
for i in range ( len ( c l a s s _ l i s t ) ) :

if class_list [ i ] < 4:
anno = s t r ( c l a s s _ l i s t [ i ] ) + "␣" + s t r ( bounding_box_list [ i ]
s t r ( bounding_box_list [ i ] [ 1 ] ) + "␣" + s t r ( bounding_box_list [ i ] [ 2 ] ) + "␣" +
s t r ( bounding_box_list [ i ] [ 3 ] ) + "\n"
f . w r i t e ( anno )
ROOTDIR = "/home/ o r a c l 4 / p r o j e c t / rdd / d a t a s e t /RDD2022/"

C o u n t r y L i s t D i r = [ " RDD2022_all_countries / Japan / t r a i n / a n n o t a t i o n s / xmls " ,

" RDD2022_all_countries / I n d i a / t r a i n / a n n o t a t i o n s
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣/ xmls " ,
" RDD2022_all_countries /China_Drone/ t r a i n
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣/ a n n o t a t i o n s / xmls " ,
" RDD2022_all_countries / China_MotorBike / t r a i n
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣/ a n n o t a t i o n s / xmls " ,
" RDD2022_all_countries / Czech / t r a i n / a n n o t a t i o n s
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣/ xmls " ,
" RDD2022_all_countries /Norway/Norway/ t r a i n /
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣ a n n o t a t i o n s / xmls " ,
" RDD2022_all_countries / United_States
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣/ United_States / t r a i n / a n n o t a t i o n s / xmls " ,
]

for CountryDir in C o u n t r y L i s t D i r :

CountryDir = ROOTDIR + CountryDir

f i l e L i s t = sorted ( g l o b . g l o b ( CountryDir + " / ∗ . xml" ) )
for f i l e in tqdm ( f i l e L i s t ) :
convertPascal2YOLOv8 ( f i l e )

def C o p y D a t a s e t S p l i t ( b a s e D i r ) :
random . s e e d ( 1 3 3 7 )
baseOutputDir = "/home/ o r a c l 4 / p r o j e c t / RoadDamageDetection / t r a i n i n g /
␣␣␣␣ d a t a s e t / r d d J a p a n I n d i a F i l t e r e d /"
countryName = Path ( b a s e D i r ) . p a r e n t s [ 0 ]
countryName = os . path . s p l i t ( countryName ) [ 1 ]

baseImageDir = b a s e D i r + " images /"

baseAnnotDir = b a s e D i r + " l a b e l s /"

i m a g e _ l i s t _ a l l = sorted ( g l o b . g l o b ( baseImageDir + " /∗ " ) )

a n n o t _ l i s t _ a l l = sorted ( g l o b . g l o b ( baseAnnotDir + " /∗ " ) )
backgroundImages_Percentage = 0 . 1
image_list = [ ]
annot_list = [ ]

d a t a s e t _ l e n g t h _ a l l = len ( i m a g e _ l i s t _ a l l )
max_background_image = int ( d a t a s e t _ l e n g t h _ a l l ∗ backgroundImages_Percenta
_counter = 0

for i in range ( len ( a n n o t _ l i s t _ a l l ) ) :

with open ( a n n o t _ l i s t _ a l l [ i ] ) as f :
_annot = f . read ( )
i f _annot :
i m a g e _ l i s t . append ( i m a g e _ l i s t _ a l l [ i ] )
a n n o t _ l i s t . append ( a n n o t _ l i s t _ a l l [ i ] )
e l i f _counter < max_background_image :
i m a g e _ l i s t . append ( i m a g e _ l i s t _ a l l [ i ] )
a n n o t _ l i s t . append ( a n n o t _ l i s t _ a l l [ i ] )
_counter = _counter + 1
d a t a s e t _ l e n g t h = len ( i m a g e _ l i s t )
split_ratio = 0.9
middle_point = round ( s p l i t _ r a t i o ∗ d a t a s e t _ l e n g t h )
numberList = l i s t ( range ( 0 , d a t a s e t _ l e n g t h ) )
random . s h u f f l e ( numberList )
t r a i n N u m b e r L i s t = numberList [ : middle_point ]
v a l i d N u m b er L i s t = numberList [ middle_point : ]
print ( " T r a i n i n g / V a l i d a t i o n ␣ Samples ␣ : " , len ( t r a i n N u m b e r L i s t ) , len ( v a l id N
print ( " Copying ␣ t r a i n i n g ␣ images ␣and␣ l a b e l s ␣ f o r " , countryName )
for i in tqdm ( t r a i n N u m b e r L i s t ) :

outputImagesDir = baseOutputDir + countryName + "/ images / t r a i n /"

i f not os . path . e x i s t s ( outputImagesDir ) :
os . makedirs ( outputImagesDir )

s h u t i l . copy2 ( i m a g e _ l i s t [ i ] , outputImagesDir )
outputAnnotDir = baseOutputDir + countryName + "/ l a b e l s / t r a i n /"
i f not os . path . e x i s t s ( outputAnnotDir ) :
os . makedirs ( outputAnnotDir )

s h u t i l . copy2 ( a n n o t _ l i s t [ i ] , outputAnnotDir )
 print ( " Copying ␣ v a l i d a t i o n ␣ images ␣and␣ l a b e l s ␣ f o r " , countryName )
for i in tqdm ( v a l i d Nu m b e r Li s t ) :

outputImagesDir = baseOutputDir + countryName + "/ images / v a l /"

i f not os . path . e x i s t s ( outputImagesDir ) :
os . makedirs ( outputImagesDir )

s h u t i l . copy2 ( i m a g e _ l i s t [ i ] , outputImagesDir )
outputAnnotDir = baseOutputDir + countryName +
"/ l a b e l s / v a l /"
i f not os . path . e x i s t s ( outputAnnotDir ) :
os . makedirs ( outputAnnotDir )

s h u t i l . copy2 ( a n n o t _ l i s t [ i ] , outputAnnotDir )

ROOTDIR = "/home/ o r a c l 4 / p r o j e c t / rdd / d a t a s e t /RDD2022/"

C o u n t r y L i s t D i r = [ " RDD2022_all_countries / Japan / t r a i n /" ,

" RDD2022_all_countries / I n d i a / t r a i n /" , ]
 ACKNOWLEDGMENT

The making of the project “REAL-TIME ROAD DAMAGE DETECTION AND GEOSPA-
TIAL MAPPING USING YOLOv8” involves the contribution of many people. We would
like to convey our sincere thanks to Dr. S.M. Khot, Principal, Fr. C. Rodrigues Institute
of Technology, Vashi for giving us the opportunity to showcase our skills and providing us
with the necessary resources. We would also like to convey our heartfelt gratitude to the
Head of Department of Information Technology, Dr. Shubhangi Vaikole for her constant
support and motivation. We express deep gratitude to our external project guide Dr.
Shashikant Dugad Indian Institute of Science Education And Research (IISER), Mohali
and project guide and mentor Dr. Archana Shirke for her constant motivation to think
out of the box and immense contribution throughout to this project. Last but not the
least, we convey our heartfelt thanks to the project coordinator, Prof. Lakshmi Gad-
hikar for supporting and guiding us throughout the process. We also extend our heartfelt
thanks to our families and well-wishers.

__________
Asavari S Bhelawe (5021106)

__________
Vedika S Pagar (5021138)

__________
Aman Singh (5021159)
 The Report is Generated by DrillBit Plagiarism Detection Software

Submission Information

Author Name Asavari B

Title Blackbook
Paper/Submission ID 3518640
Submitted by [Link]@[Link]
Submission Date 2025-04-17 [Link]
Total Pages, Total Words 28, 5681
Document type Project Work

Result Information

Similarity 8%
1 10 20 30 40 50 60 70 80 90

Sources Type Report Content

Journal/
Publicatio
n 2.71%

Internet
5.29%
Words <
14,
4.21%

Exclude Information Database Selection

Quotes Not Excluded Language English

References/Bibliography Not Excluded Student Papers Yes
Source: Excluded < 14 Words Not Excluded Journals & publishers Yes
Excluded Source 0% Internet or Web Yes
Excluded Phrases Not Excluded Institution Repository Yes

A Unique QR Code use to View/Download/Share Pdf File

CONFERENCE PRESENTATION CERTIFICATES
 Real-Time Road Damage Detection and Geospatial
Mapping Using YOLOv8
[1]
Asavari S Bhelawe, [2]Aman Singh, [3]Vedika Pagar, [4]Archana Shirke, [5]Shashikant Dugad
[1]
Fr. C. Rodrigues Institute of Technology, [2]Fr. C. Rodrigues Institute of Technology, [3]Fr. C. Rodrigues
Institute of Technology, [4]Fr. C. Rodrigues Institute of Technology, [5]Indian Institute of Science Education And
Research (IISER)
[1]
basavari2003@[Link], [2]amansingh0550@[Link], [3]vedikasp2003@[Link],
[4]
[Link]@[Link], [5]shashi@[Link]

Abstract: Modern technology has revolutionized the monitoring of urban roads using various video
sources such as smartphones, car cameras, and surveillance systems. Focusing on roads in India
and Japan, this study presents a scalable deep learning-based system for the real-time identification,
categorization, and mapping of road damage. The solution addresses challenges such as inconsistent
image quality, diverse climatic conditions, and varying regional infrastructures by utilizing the
YOLO object detection algorithm, trained on annotated datasets from the Japan Road Association
and enhanced by data augmentation. A user interface is developed, featuring a main page with live
real-time detection, a road damage map, and report generation capabilities. These reports are
provided to authorities for efficient infrastructure management, optimizing resource allocation, and
prioritizing road repairs. This system significantly improves road maintenance processes, benefiting
both nations by streamlining road monitoring and decision-making. In summary, your system could
make road condition improvement processes 70-80% more efficient and 50-95% faster in India
compared to traditional methods of road damage detection and recovery process.

Index terms: Deep Learning, Road Damage Detection, Object Detection, Geospatial Visualization,
Urban Street Analysis, AI-Based Maintenance Planning, Convolutional Neural Networks (CNN),
Smartphone Image Processing, Urban Street Analysis.

deep learning approaches to automate road damage detection,

I. INTRODUCTION identifying diverse crack types and structural road issues with
The infrastructure of cities significantly affects their eco increased precision. Among these, image-based techniques
nomic performance, with roads being a critical element of pub have gained widespread acceptance due to their simplicity and
lic facilities. Various factors, such as weather conditions and scalability for large-scale monitoring. This paper introduces
aging, lead to road damage, impacting not only road an AI-driven solution that integrates deep learning for
efficiency and vehicle maintenance but also posing serious automated road damage detection and geospatial
safety risks to drivers. In fact, poor road conditions are a visualization. Using state-of-the-art algorithms such as
primary contributor to traffic accidents, with an estimated 50 YOLO, combined with smartphone images crowdsourced
million people injured annually in Europe alone due to road from city crews and the public, our approach not only
conditions. Govern ments worldwide, including the United identifies various types of road damage but also integrates
States, have invested heavily in road maintenance, allocating geospatial mapping to provide a comprehensive visualization
billions annually to construct and maintain road networks of road conditions across large areas. Additionally, we
Traditional methods of road damage detection, such as developed a user interface that provides real-time damage
manual visual inspections, are labor-intensive, costly, and detection, an interactive road damage map, and automatic
time-consuming. Researchers have explored alternatives like report generation. These reports are provided to authorities for
vibration-based, laser-scanning, and image-based detection efficient infrastructure management, optimizing resource
systems. While vibration methods are limited to direct contact allocation, and prioritizing road repairs. This interface enables
with the road surface, laser-scanning techniques, though city planners to prioritize road repairs and maintenance
accurate, come with high costs and often require road efficiently, making the system an essential tool for improving
closures. Image-based methods, by contrast, are cost-effective infrastructure management.
but have historically struggled with accuracy. Recent
breakthroughs in deep learning and image process ing have II. RELATED WORK
significantly improved the potential of image-based systems.
Significant breakthroughs and game-changing inventions
Leveraging these advancements, researchers have developed
have characterized the development of automated road
damage detection. Road surface damage mapping and Deep learning’s introduction transformed the area by making
identification have moved from labor-intensive manual it possible for more automated and flexible solutions. While
procedures to ad vanced automated systems with the deep neural network-based approaches provided increased
introduction of AI-driven approaches. This study investigates efficiency and scalability, techniques such as CrackNet and
the use of cutting-edge AI methods, with an emphasis on sliding window classifiers achieved great accuracy in crack
models such as YOLO, to accomplish real-time mapping and identification. Nevertheless, a lot of methods concentrated
detection of road damage. This part offers a thorough analysis
mostly on identifying damage without categorizing it into
of similar works to place this study in the larger framework of
other kinds. Newer models have attempted to differentiate
current advancements. It highlights the technical turning
points that have influenced the field and the difficulties that between categories including vertical, horizontal, and
automated solutions in road infrastructure management have crocodile cracks in light of the practical need for road
handled. administration to implement particular repair techniques
based on damage kinds. Faster R-CNN, YOLO, R-FCN, and
A. Traditional Image Processing Techniques SSD are examples of recent developments in end-to-end
Traditional image processing approaches have served as a object detection systems that have surpassed conventional
foundational step in creating automated methods for detecting classification-based techniques. These methods greatly
road surface damage. To find abnormalities like cracks and improve speed and accuracy by processing complete images
potholes, early attempts used algorithms like edge detection, in a single pass. The absence of standardized datasets for
thresholding, and morphological operations[9]. A notable detecting road damage has presented difficulties in spite of
shift from manual inspection techniques was made when these advancements. Researchers have started producing
ground breaking investigations incorporated wavelet
publicly accessible road damage statistics using less complex
transforms to enhance fracture detection. These early
techniques, such as smartphone-captured photos from
methods, however, had serious drawbacks, most notably their
susceptibility to ambient influences such changing lighting passenger automobiles, while existing datasets, like KITTI,
and shadows, which fre quently resulted in large false positive concentrate on autonomous driving.
[Link] the same time, despite their widespread use, manual
C. R-CNN
road inspection tech niques have shown themselves to be
unreliable and resource intensive. Human-led visual An enormous advancement in object detection tasks was
inspections increase the hazards associated with aging heralded by the release of Region-based Convolutional
infrastructure since they require skilled staff, take a lot of Neural Networks (R-CNN). By integrating region proposal
time, and frequently produce inconsistent results. Maintaining networks with CNNs, R-CNN, and its evolutionary offspring,
road safety is made more difficult for municipalities with Fast RCNN and Faster R-CNN, enabled accurate object
limited resources since they find it difficult to carry out localization within images [12–14]. This resulted in a better
adequate and timely inspections. These drawbacks highlight capacity to recognize and distinguish particular damaged
the necessity for sophisticated, automated systems that can areas within a larger road image in the context of road damage
offer dependable, expandable, and reasonably priced detection. The works of [15] stand testament to the efficacy of
solutions for mapping and detecting road degradation. Faster R-CNN in detecting and segmenting road damages.
B. Artificial Intelligence and Pattern Analysis D. SSD
Road damage identification underwent a paradigm change An object identification approach called SSD [16] does not
with the move from basic image processing techniques to require a second stage per-proposal classification operation
machine learning. In the past, methods like edge detection, because it employs a single feed-forward convolutional
thresholding, and texture-based feature extraction were used network to forecast classes and anchor offsets directly. This
to extract observable features from pictures. Notable framework’s primary characteristic is the usage of multi-scale
developments that greatly increased detection accuracy convolutional bounding box outputs connected to numerous
included merging these traits with Support Vector Machines feature maps at the network’s top.
(SVM)[10]. These techniques were labor-intensive and prone
E. YOLO
to inconsistencies, though, because they needed a great deal
of manual feature engineering. Road damage identification YOLO (You Only Look Once) is a single-stage object
was further improved by quantitative inspections employing detection model that simultaneously predicts bounding boxes
mobile measuring systems (MMS) and laser scanning, which and class probabilities, making it a highly efficient and fast
provided accurate geographical data. Although MMS systems solution for real-time applications. Unlike two-stage models
with high-resolution cameras, laser scanners, and GPS like Faster R-CNN, YOLO divides the input image into a grid,
devices offered precise assessments, they were still too where each grid cell generates anchor boxes that predict
expensive for smaller towns. As a result, scientists started object presence, bounding box coordinates, and class
looking at less expensive options by utilizing in-car cameras probabilities. Although there have been significant
and image-processing software. For example, black-box breakthroughs in detection approaches, real-time processing
cameras with naive Bayes classifiers showed promise in demands continue to be a major challenge. Due to the
detecting irregularities in the road. operational requirements of road maintenance, damage
identification must be done promptly in addition to accurately.
Real-time object detection was emphasized by architectures subset includes 1,313 images from Japan, 969 from India, and
like YOLO (You Only Look Once) and SSD (Single Shot 349 from the Czech Republic, while the test2 subset contains
MultiBox Detector), as explained by researchers [17–18]. 1,314, 990, and 360 images from the respective countries. The
Even while these frameworks aren’t specifically designed for Fig.1 shows how the dataset is distributed among the different
road anomalies, their fundamental ideas offer priceless countries.
insights. They draw attention to the complex trade-offs and
balance between accuracy and detection speed. These factors
are crucial when imagining a model that functions in dynamic
real-world environments, highlighting the necessity of any
potential road damage detection system to have the ideal
balance of accuracy and speed.
III. PROPOSED SYSTEM
A. Dataset
The RDD2020 dataset[17] is a crucial resource for developing
smartphone-based road damage detection systems. It provides
an affordable solution for road condition monitoring, making
it valuable for municipalities and road agencies. This dataset
facilitates the creation of new deep convolutional neural
network (CNN) architectures and the enhancement of existing
ones to improve detection algorithms. It includes road images Fig. 1: Statistics for the number of damage instances
from multiple countries, such as India, Japan, and the Czech included in the underlying datasets
Republic, and offers opportunities for expansion to other
regions. The dataset is designed to support the detection and B. Flowchart
classification of various road damage types, including The figure(Fig.2) represents the architecture of a YOLO
longitudinal cracks, transverse cracks, alligator cracks, and model used for road damage detection. It illustrates the flow
potholes, with the potential for incorporating more categories of data through three main components: the Backbone, Neck,
in the future. Researchers can use RDD2020 for and Prediction Head. The process begins with an input image
benchmarking the performance of machine learning of a road, which is analyzed for visible damages such as
algorithms in image classification and object detection tasks. cracks or potholes.
Furthermore, it played a significant role in the Global Road The feature extraction process begins with the Backbone, a
Damage Detection Challenge (GRDDC 2020), part of the key component responsible for identifying essential features
IEEE Big Data Cup, which evaluated the effectiveness of road from input images. It employs convolutional layers to detect
damage detection models. The dataset contains annotations basic elements, such as edges and textures, which are further
for four types of road damage(Table II): longitudinal cracks refined through C2f blocks to capture complex patterns
(D00), transverse cracks (D10), alligator cracks (D20), and specific to road damage. Additionally, the Spatial Pyramid
potholes (D40). Pooling-Fast (SPPF) module aggregates features at multiple
TABLE I: Road Damage Types scales, enabling the detection of damages ranging from small
cracks to large potholes. The Neck component processes
features extracted by the Backbone and enhances the model’s
ability to detect objects across different scales. This involves
up sampling layers that scale up feature maps to improve the
detection of smaller damages and concatenation layers that
combine features from various stages of the Backbone for
richer information. The Neck also includes additional C2f
blocks to refine features, ensuring a clear distinction between
road damage and its surrounding environment.
The RDD2020 image dataset comprises 26,336 road images Finally, the Prediction Head generates outputs by detecting
from India, Japan, and the Czech Republic, representing over and classifying road damages. It employs detection layers to
31,000 instances of road damage. The dataset is divided into produce bounding boxes, assign confidence scores, and
training, test1, and test2 subsets. The training set includes classify damages such as” crack” or ”pothole.” The resulting
subdirectories for India, Japan, and the Czech Republic, with output is an annotated image with bounding boxes and class
images and annotations specific to each country. Images from labels, indicating the location and type of damage, along with
Japan and the Czech Republic have resolutions of 600 × 600 associated confidence scores.
pixels, while those from India are 720 × 720 pixels. The test1
and test2 subsets follow the same resolution patterns and
contain images from all three countries. Specifically, the test1
 Fig. 2: YOLOv8 Architecture for Road Damage Detection

C. Training Results
The training and evaluation were carried out using
the dataset. The following results were observed.

(a) mAP50-95 curve (b) mAP50 curve

Fig. 4: Performance Curves

The performance metric curves (Fig. 4) provide further

evidence of the model’s success. The mAP@50 (Mean
(a) Training Loss Curves (b) Validation Loss Curves Average Precision at an IoU threshold of 0.5) demonstrates a
Fig. 3: Loss Curves for YOLOv8 model
consistent upward trend, showcasing improved alignment
between the model’s predictions and ground truth. This
Figure (3 (a) & (b)) illustrates the training and validation loss
metric, which evaluates the balance between precision and
curves for the model over 50 epochs. The chart on the left recall, reflects the model’s growing accuracy in detecting
represents the training losses, while the chart on the right
objects. Similarly, the mAP@50-95, a more stringent metric
shows the validation losses. Three distinct components of the
averaging precision across a range of IoU thresholds (0.5 to
loss are plotted: Box Loss (blue), Class Loss (red), and DFL 0.95), also increases steadily, emphasizing the robustness of
Loss (green). These losses demonstrate the gradual reduction
the model’s performance under varying levels of overlap
in error as the training progresses, indicating effective criteria. Overall, the model achieved a mean Average
learning by the model. In the training loss curve (Fig 3(a)), all
Precision (mAP50) of 0.547 and mAP50-95 of 0.254 across
three components exhibit a significant decrease during the
all damage categories, demonstrating robust performance in
initial epochs, followed by a slower but steady decline as the real-world conditions.
model converges. Notably, the Class Loss starts at a higher
The table (Table II) provides a summary of the training
value compared to the other losses but consistently reduces, metrics for a YOLO model over 50 epochs, highlighting
eventu ally stabilizing below 1.6. Similarly, Box Loss and
performance improvements. The mAP50 increases from
DFL Loss also exhibit a downward trend, stabilizing at
0.212 in the first epoch to 0.548 by the 50th epoch, while the
approximately 1.5 and 1.4, respectively. In the validation loss mAP50-95 improves from 0.086 to 0.254, showing enhanced
curve (Fig 3(b)), a similar pattern is observed, with all loss
detection accuracy across different IoU thresholds. Precision
components initially fluctuating before stabilizing. The Class
rises from 0.315 to 0.583, indicating fewer false positives, and
Loss shows slightly higher values compared to the training
Recall improves from 0.272 to 0.523, demonstrating better
curve, peaking around epoch 5 before gradually decreasing. detection of true positives. The train box loss decreases from
The Box Loss and DFL Loss follow a smoother decline, 2.232 to 1.566, and the validation box loss reduces from 2.004
indicating consistency between training and validation
to 1.807, reflecting improved bounding box prediction
phases. accuracy and generalization to validation data.
These curves highlight the model’s ability to generalize
effectively, as evidenced by the convergence of training and
validation losses. However, the minor fluctuations in the
validation losses suggest opportunities for further
optimization to reduce overfitting or variance.
 TABLE II: Training metrics table

IV. RESULTS AND ANALYSIS

A. Experimental Result
Figure [5] compares the model’s performance on unanno tated
and annotated frames for three road damage types. The f irst
row detects a transverse crack (Confidence: 0.64) in yel low.
The second row highlights an alligator crack (Confidence:
0.81). The third row identifies a pothole (Confidence: 0.79) in
green. The annotations demonstrate the model’s effectiveness
in detecting and classifying road damages..
Fig. 6: Road damage photos and classes for a model training
B. Application Interface
Using React Native, a sophisticated application for identi
fying road damage was created, combining several pages to
provide a thorough and intuitive user experience. The Main
Page(Fig.7 (a)) serves as the main navigation hub, giving
users easy access to important features such a real-time
damage detection module, a manual reporting interface for
users to submit problems, and a comprehensive map that
highlights dis covered road damages. The application’s user-
friendly layout makes it simple for users to navigate, which
expedites the road monitoring and management process. Real-
time road damage detection is made possible by the Live
Real-Time Detection Screen(Fig.7 (b)), which provides state-
of-the-art functionality. It gives consumers trustworthy and
useful information by precisely identifying the kind of
damage, such as potholes, cracks, or faded road markings, and
by displaying confidence ratings for each detection. The
highest level of accuracy is guaranteed by this real-time
feedback, enabling stakeholders to prioritize road
Fig. 5: Road Damage Detection using trained YOLOv8 maintenance and repairs with knowledge. The application is a
model major advancement in using technology for proactive road
Figure (6) showcases the detection and classification of monitoring and safety improvement because of its integration
various road damages using the proposed model. Subfigure of cutting-edge capabilities and user-friendly design.
(a) identifies a longitudinal crack (D00) in red, while (b) Additionally, the Road Damage Map Screen offers a thorough
highlights potholes (D40) in green. Subfigure (c) presents a and user-friendly display of identified road damages by
transverse crack (D01) in yellow, and (d) detects multiple highlighting their locations on an interactive map. Each
potholes (D40), demonstrating the model’s ability to identify damage marker indicates specific information, such as the
multiple instances. In (e), both a transverse crack (D10) and type of damage (e.g., potholes, cracks, or blurring of road
an alligator crack (D20) are marked in yellow. Subfigure (f) markings)(Fig.8) and the corresponding confidence level of
detects a transverse crack (D01) on a straight road section. the detection.
The bounding boxes, labels, and confidence scores highlight
the model’s accuracy in classifying different road damage.
 [2] Kulambayev, Bakhytzhan, Nurlybek, Magzat, Astaubayeva,
Gulnar ,Tleuberdiyeva, Gulnara Zholdasbayev, Serik Tolep,
Abdimukhan. (2023). Real-Time Road Surface Damage Detection
Framework based on Mask R-CNN Model. International Journal of
Advanced Computer Science and Applications. 14.
10.14569/IJACSA.2023.0140979.
[3] M. Maniat, C. Camp and A. Kashani, —Deep learning-based
visual crack detection using Google Street View images,|| Neural
Computing and Applications, vol. 33, no. 21, pp. 14565 14582, 2021.
[4] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards
real-time object detection with region proposal networks,” in
Advances in neural information processing systems, 2019, pp. 91–
99.
[5] Smita Rukhande, Prachi G, Archana S, Dipa D, “Implementation
of GPS enabled carpooling System” at International Journal of
Advances in Engineering and Technology(IJAET) November 2011
ISSN: 2231 1963.
[6] Abushakra, A., Subhi, H., and Abdelkarim, S. (2023). Real-Time
Road Surface Damage Detection Framework Using Deep Learning.
International Journal of Advanced Computer Science and
Applications (IJACSA), 14(9), 641-649.
(a) Main Page (b) Live Detection [7] M. Maniat, C. Camp and A. Kashani,Neural Computing
Fig. 7: Application GUI andApplications, vol. 33, no. 21, pp. 14565-14582, 2021.
[8] J. Y. Lee, H. Lee, and Y. H. Kim, ”Deep road crack detection
with mobile application,” 2019 IEEE International Conference on
Consumer Electronics (ICCE), Las Vegas, NV, USA, 2019, pp. 354-
359, doi: 10.1109/ICCE.2019.8621899.
[9] D. Quang and S. Bae. ”A hybrid deep convolutional neural
network approach for predicting the traffic congestion index,”
Promet-Traffic and Transportation, vol. 33, no. 3, pp. 373-385, 2021.
[10] ]N. Safaei, O. Smadi, B. Safaei and A. Masoud, Transportation
Research Record, vol. 2675, no. 9, pp. 370-381, 2021.
[11] Maeda, Hiroya Sekimoto, Yoshihide Seto, Toshikazu
Kashiyama, Takehiro Omata, Hiroshi. (2018). Road Damage
Detection Using Deep Neural Networks with Images Captured
Through a Smartphone. Arxiv. 10.48550/arXiv.1801.09454.
[12] V. Tran, T. Tran, H. Lee, K. Kim, J. Baek et al., —One stage
detector (RetinaNet)-based crack detection for asphalt pavements
considering pavement distresses and surface objects,|| Journal of
Civil Structural Health Monitoring, vol. 11, no. 1, pp. 205-222, 2021
[13] Z. Lingxin, S. Junkai and Z. Baijie, Earthquake Engineering and
Engineering Vibration, vol. 21, no. 1, pp. 1-21, 2022.
[14] K. Gopalakrishnan, S. Khaitan, A. Choudhary and A. Agrawal,
Deep convolutional neural networks with transfer learning for
computer vision-based data-driven pavement distress detection,||
Fig. 8: Geo-mapping of damages Construction and building materials, vol. 157, no. 1, pp.322-330,
2017.
CONCLUSION [15] S. Patra, A. Middya and S. Roy, —PotSpot: Participatory
The road damage detection system leverages the YOLO (You sensing based monitoring system for pothole detection using deep
Only Look Once) architecture for real-time identification and learning,|| Multime dia Tools and Applications, vol. 80, no. 16, pp.
25171-25195,2021.
classification of road damage types with promising initial [16] F. Andika and Y. Bandung, ”Road Damage Classification using
accuracy. Future advancements aim to refine detection SSD Mobilenet with Image Enhancement,” 2023 International
algorithms to handle diverse conditions and surfaces, enhance Conference on Computer Science, Information Technology and
precision with additional models like semantic or instance Engineering (IC CoSITE), Jakarta, Indonesia, 2023, pp. 540-545,
segmentation, and establish a comprehensive road condition doi: 10.1109/IC CoSITE57641.2023.10127763.
database. Collaborating with maintenance authorities and [17] Y. Shi, L. Cui Z. Qi, F. Meng and Z. Chen, —Automatic road
conducting field tests will validate its real-world crack detection using random structured forests,|| IEEE Transactions
effectiveness, paving the way for smarter, data-driven on Intelligent Transportation Systems, vol. 17, no. 12, pp. 3434-
infrastructure management solutions. 3445, 2016.
[18] Deeksha Arya, Hiroya Maeda, Sanjay Kumar Ghosh, Durga
Toshniwal, Yoshihide Sekimoto, RDD2020: An annotated image
REFERENCES dataset for automatic road damage detection using deep learn ing,
Data in Brief, Volume 36, 2021, 107133, ISSN 2352-3409,
[1] S. Wu, J. Fang, X. Zheng and X. Li, —Sample and structure- [Link]
guided network for road crack detection,|| IEEE Access, vol. 7, no.
1, pp. 130032-130043, 2019 
PCUBE PROJECT PRESENTATION
CERTIFICATES
TECHSPARKS PROJECT COMPETITION
CERTIFICATES