IRJET - An Intelligent Pothole Detection System using Deep Learning

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 02 | Feb 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 1591
An Intelligent Pothole Detection System using Deep Learning
Harun Joe1, Joseph Blessingh2, Joel Cherian3
1,2,3Student, Dept. of Computer Engineering, Fr. C. Rodrigues Institute of Technology, Vashi, Maharashtra, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract -Potholes on roads constituteamajorproblem for
citizens acting as pedestrians as well as vehicular drivers.
Government bodieswhichconsistofengineersand workers are
responsible to detect damages on roads. Manually assessing
every single part of the road is highly time-consuming,
requires a lot of manpower and hence it cannot be done
efficiently. The method to fix this issue by automating the
detection. The study focuses on collecting and analyzing the
dataset of potholes to train a convolutional neural network.
The object detection system tiny YOLOv3 is used for detecting
the potholes. The design of a system is identified which can be
used for developing a mobile application for detection and
presenting a visualized view of the potholes.
Key Words: YOLOv3, Deep Learning, Pothole, Object
Detection, API, NoSQL
1. INTRODUCTION
Transportation by roads has beenfairlyeasyandcostsaving,
although the condition of road decides the comfort of the
ride. Potholes on roads are a major cause of inconvenience
for the people travelling by vehicular modes of transport.
They are caused by the expansion and contractionofground
water after the water has entered into the ground under the
pavement. Potholes are therefore becoming a sizable threat
to drivers as they face accidents and damage to the vehicle.
In 2017, Union Ministry of Road Transport and Highway in
India reported 2,000 pothole-related deaths [1].
The potholes have a need to be repaired to make the ride
comfortable and rule out dangers. The responsibility to
repair these potholes lies under the Municipal Authoritiesof
the city, but the current method of inspecting the roads for
potholes and other irregularities are manual or through
complaints registered by citizens affected [2]. This makes it
difficult to comprehend the severity and urgency of roads
that require immediate attention if it is not detected and
informed to the authorities.
The paper shows the work done on the dataset used by
analyzing and preparing itforthetrainingphase.Thedataset
consists of images from majorly two sources, first being an
open dataset available in Kaggle [3] and the second source
being images of the current state of roads captured from
Mumbai city. Images from both these sources give a diverse
set of data for training the model for detection. YOLOv3, an
object detection model to detectthepotholesistrainedusing
the dataset. The focus is also on creating a system design
which is capable of operating on mobile to detect potholes.
The aim is to provide a solution to automate the manual
challenges of finding the potholes by the municipal body. It
constitutes a Deep learning model which can be used to
detect potholes and visualize them on map using a mobile
application. The road to settingupthe entiresystemincludes
a mobile being mounted on the dashboard inside a vehicle
belonging to the municipal authority. When the vehicle is in
motion, the mobile application begins recording thevideoof
the road ahead of the vehicle. The video feed is
simultaneously processed to capture images based on
distance intervals. Images are fed to the pothole detection
model to detect potholes. The data pertaining to the pothole
is stored, which is then retrieved to be able to visualize it on
a Map. This information can be used by the Municipal
Authority to repair the potholes.
2. RELATED WORKS
Different real timepothole detectionsystemswerestudiedto
gain an insight over the various algorithms and techniques
used in it. The location of pothole is an important aspect. In
[4], space interpolation method is used to obtain exact
location of the pothole, the distance between the start point
and the pothole followed by the distance between end point
and the pothole is obtained by using a GPS module on the
mobile. A system to detect the potholeusinganopticaldevice
mounted on the vehicle is explained by [5]. This approach
usescandidateregion extractiontoidentifythepotholeandit
is first segmented by various techniqueslikeOtsu,Histogram
shape-based thresholding. This proposed system gave an
accuracy of 73.5 percentage with a precision of 80
percentage, the location and severity of the pothole is given
to the center i.e. pothole alert service and road management
system.
A simulation-based method was developed by Sudarshan et
al. [6] to construct mobile nodes equipped with WiFi and a
single access point between the roads to alert drivers i.e.
mobile nodes about pothole on roads is achieved by a
software called Qualnet 4.0. The alert is sent by packet
transmission from the access point to the nearby nodes and
collision avoidance by RTS/CTS or CA. Youngtae et al. [7]
employed a Cascade detector method to distinguish pothole
from similar types of shades and road patches. This method
makes use of a black box camera fitted on top of a vehicle
which removes objects like shades that changes its shape
irrespective of a pothole. As our proposed system uses tiny
YOLOv3, the basic understanding of the model was studied
from this paper [8]. This model makes use anchor boxeswith
a score for each object identified with the help of logistic
regression.

3. POTHOLE DETECTION SYSTEM
The pothole detection system aims to provide a solution for
both municipal authority and the citizens to reduce the
number of potholes by developing an application for
mobiles. The entire design of the system is based on the
integration of various libraries and frameworks. A mobile
application, APIs to provide insights from data gathered, an
object detection model to precisely detectthepotholesanda
cloud storage solution to store and handle various forms of
data. The application has two major functionalities:First,for
detection and gathering of data mainly for the local
municipal body’s usage and the other is the view of the
gathered data on a map which is helpful for both the
authorities responsible for maintenanceandalsoforcitizens
of the locality to be aware of their surrounding roads.
3.1 Design of the System
The Fig -1 shows the basic structure of the system. The flow
of the working is represented through the diagram.
Fig -1: Block Diagram
3.1.1 Mobile Application
With the processing capability of mobile devices sky
rocketing compared to its preceding years, computational
workload that can be handled by the device has also
increased. This makes the mobile devices to be able to
handle machine learning task easily. With the help of
libraries for integration of CNN to an application it is thus
made possible to run an object detectionmodel inthemobile
device itself. Along with that, a constant listener can run in
the background that updates the user’s location. With the
location updated, the distance is calculated and the image is
captured. The captured image is then fed into to the Neural
Network and the potholes are recognized as co-ordinates in
the images.
Fig -2: Use Case Diagram
A constant listener runs on the background that updates the
location. With the locationupdated,thedistanceiscalculated
and the image is captured. The captured image is then fed
into to the Neural Network and the potholes are recognized
as co-ordinates in the images.
The use case diagram of the application is as shown inFig -2.
3.1.2 Database
Data gathered through all the mentioned proceduresare not
in a structured format, i.e. data can be in a variety of forms
and thus cannot be predefined and inserted in a set of rows
and columns. This make the traditional form of databases
like MySQL not useful in this type of scenario. NoSQLform of
databases can deal with huge amount of data faster than
MySQL [9]. Firebase is a cloud Storagesolutionprovided and
hosted by Google which is NoSQL form of database that
stores data in a JSON file format [10]. To have reliable
synchronization within the servertofeedinlivedata,suchas
the images of roads with potholes, number of potholes,
locations as latitude and longitude etc. APIs provided from
Firebase itself allow to make HTTP requests to the cloud
storage which helps in performing various operation on the
database such as insert, update,delete.Also,integrationwith
Firebase using Flutter or other mobile development
platforms becomes effortless since both the platforms are
created by the same parent organization.

3.2 Study of Dataset
The main engine behind this application and database is the
dataset, the dataset is trained using the YOLOv3 model. The
study of dataset is divided into three parts:
1) About the Dataset
2) Description of Dataset
3) Analyzing the Dataset
4) Object Detection using YOLOv3 and Darknet
3.2.1 About the Dataset
The dataset for training constitutes1,500imagesfortraining
and 350 images for testing the model developed. Kaggle
provides an open dataset of roads which contain potholes
called Nienabar Pothole dataset [3]. This dataset consists of
repetitive images which needs to be filtered out such that
images in the final dataset is uniqueandofdifferentfeatures.
But, certain factors in this dataset is constant such as colors
of roads, illumination, and pothole vicinity do not help in
making the model well versedwithvariousreal-lifescenario.
Using python and OpenCV, we also reduced the resolution of
images to remove the constant part of images in the dataset
like dashboard and the sky.
Fig -3: Captured Image for Dataset
The Fig -3 shows the self-captured image from the streets of
Mumbai containing pothole which is combined with the
Nienabar Pothole dataset as shown in Fig -4 to give the
dataset a wide spectrum of images. To make the dataset
unique and versatile, web scraping is performed. Using
various keywords for search, 1000 images were extracted.
But, since searches using different keywords provides same
images sometimes, manual identification and removal of
such images were performed. The scraped images were
merged with the dataset. LabelImg [11] wasusedtolabelthe
potholes from images according to the YOLO requirements.
Fig -4: Image from Nienabar Dataset
3.2.2 Description of Dataset
There are 2 categories in the dataset namely positive and
negative. Positive denotes the fact thattheroadcontains one
of multiple potholes whereas negative will denote that the
image doesn’t contain any kind of potholes in it. But training
a model in YOLO does not require partition of images in the
dataset and hence a 70-30 ratio of images from the positive
and negative images were taken from the files into the final
dataset.
3.2.3 Analyzing the Dataset
Python is used to collect and process the final pothole
dataset for our system. The scraping of images from theweb
is done as the images were in different file formats. A
common file format is used to convert all thescrapedimages
into same extension. Since, the dataset had scraped images
from web it consisted of multiple repetitive images in a row,
images in the final dataset have been taken from an interval
of 17 images. This ensures that the final dataset is free from
redundancy. The images selected has a common
characteristic in them i.e. the dashboard and the sky from
the Nienabar Pothole dataset. So, to filter out the other
unwanted parts of the image, cropping of images is done
using a simple python program.
Fig -5: Pothole Detected using YOLOv3

3.2.4 Object Detection using YOLOv3 and Darknet
With multiple images being captured, it becomes essential
that the object detection model detects potholes in the least
amount of time. YOLOv3 (You Only Look Once ver.3)
improves operation speed which meets real time
requirements for detection [8]. Darknet is a framework for
YOLO implemented using C/CUDA. To make the object
detection model mobile friendly, the model is convertedtoa
TensorFlow lite model, whichcompressestheneuronsinthe
neural network. This will make the model run efficiently in
the mobile devices. Fig -5 shows the detection of pothole
which has been implemented.
4. CONCLUSION
The paper showcases our work performed to train an object
detection model capableofdetectingpotholes.Thecollection
of images for dataset to train the model consists of a mixture
of Nienabar Pothole Dataset and self-captured images of
pothole. Tiny YOLOv3, a CNN which is capable tofunctionon
a mobile device is trained to be used as the object detection
model. Furthermore, a designwasintroducedforthePothole
Detection System which uses a mobile application to detect
potholes and represent on maps using markers with APIs
from Google and other third-party resources. Using the
application with the plotted information, the responsible
civic body can repair the pothole with the gained insights.
5. FUTURE SCOPE
The system’s continued work includes the development of
the mobile application and the integration of the object
detection model with the application. Also, integrating the
system to the cameras present in the new generation
vehicles to give a real-time analysis of roads. With the data
gathered through this analysis, providing thevehicledrivers
on roads with an alert or prompt regarding the current state
of the roads they are travelling on.
ACKNOWLEDGEMENT
We take this opportunity to expressoursincereappreciation
to Mr. Amroz K. Siddiqui who has been instrumental in
guiding and providing us the technical advice throughout
this project.
REFERENCES
[1] The Economic Times, “Supreme Court takes note of
3597 deaths due to pothole-related accidents in 2017”,
https://economictimes.indiatimes.com/news/politics-and-
nation/supreme-court-takes-note-of-3597-deaths-due-to-
pothole-related-accidents-in-
2017/articleshow/65858401.cms, 2018.
[2] Pankaj Upadhyay, India Today, “BMC launches new app
for potholes”, https://www.indiatoday.in/india/story/bmc-
potholes-app-mumbai-1601320-2019-09-20, 2019.
[3] Felipe Muller, Nienaber potholes 2 complex.
https://www.kaggle.com/felipemuller5/nienaber-potholes-2-
complex, 2019.
[4] Hsiu-Wen Wang, Chi-Hua Chen,Ding-YuanCheng,Chun-
Hao Lin, and Chi-Chun Lo, “Areal-time potholedetection
approach for intelligent transportation system”,
Mathematical Problems in Engineering, 2015.
[5] Seung-Ki Ryu, Taehyeong Kim, and Young-Ro Kim,
“Image-based pothole detection system for its service
and road management system”, Mathematical Problems
in Engineering, 2015.
[6] Sudarshan S Rode, Shonil Vijay, Prakhar Goyal,
Purushottam Kulkarni, and Kavi Arya, “Pothole
detection and warning system: infrastructure support
and system design”, International Conference on
electronic computer technology, pages 286–290. IEEE,
2009.
[7] Youngtae Jo and Seungki Ryu, “Potholedetectionsystem
using a blackbox camera”, Sensors, 15(11):29316–
29331, 2015.
[8] Joseph Redmon and Ali Farhadi, “Yolov3: An
incremental improvement”, arXiv, preprint
arXiv:1804.02767, 2018.
[9] Ahmad Abdullah and Qingfeng Zhuge, “From relational
databases to nosql databases: Performance evaluation”,
Research Journal of Applied Sciences, Engineering and
Technology, 11(4):434–439, 2015.
[10] Pratik Lahudkar, Sarita Sawale, Vijay Deshmane, and
Krishna Bharambe, “Nosql database-googles firebase:A
review”, International Journal of InnovativeResearchin
Science, Engineering and Technology, 7(3), 2018.
[11] Labelimg,
https://awesomeopensource.com/project/tzutalin/labe
lImg.

IRJET - An Intelligent Pothole Detection System using Deep Learning

More Related Content

What's hot (20)

Similar to IRJET - An Intelligent Pothole Detection System using Deep Learning (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET - An Intelligent Pothole Detection System using Deep Learning