2023 2nd International Conference on Futuristic Technologies (INCOFT)

Karnataka, India. Nov 24-26, 2023

Pothole Detection Using Deep Learning

1
Kurra Kaushik Sai, 2D Deekshith Vardhan Kumar, 3Ande Sahrudhay, 4Krishna Dharavath,
Department of Electronics and Communication Engineering,
Vardhaman College of Engineering, Hyderabad, India
Email: {1k.kaushiksai2003, [Link], 3saichinnunetha1, [Link]}@[Link]

Abstract— Potholes are a significant road infrastructure YOLOv8’s real-time processing capabilities enable quick
problem that poses safety risks to drivers and causes detection of potholes, ensuring that maintenance crews can
substantial damage to vehicles. Accurate and real time be dispatched promptly to repair the road before the potholes
detection of potholes is crucial for timely repairs and road worsen..
maintenance. This paper presents a novel approach for pothole
detection using YOLOv8, an advanced version of the You Only
A. Pothole Formation
Look Once (YOLO) object detection algorithm. The proposed
method leverages the advantages of YOLOv8, which combines Potholes are common road defects that form due to a
state-of-the-art object detection techniques with improved combination of natural and human-made factors. The sample
2023 2nd International Conference on Futuristic Technologies (INCOFT) | 979-8-3503-0884-6/23/$31.00 ©2023 IEEE | DOI: 10.1109/INCOFT60753.2023.10425461

network architecture for enhanced accuracy and efficiency. A pothole images of Indian roads is illustrated in Figure 1(a)
comprehensive dataset of road images containing various and Figure 1(b).
pothole instances is collected and used to train the YOLOv8
model. Transfer learning techniques are employed to fine-tune
the model for pothole detection specifically. To evaluate the
performance of the proposed method, extensive experiments
are conducted on different road scenarios, including varying
lighting conditions, road types, and pothole sizes. The results
demonstrate that the YOLOv8- based pothole detection system
achieves high accuracy and real-time processing capability,
making it suitable for deployment in smart city infrastructure
and vehicle navigation systems. The contributions of this work
lie in the development of an effective and efficient pothole
detection system using YOLOv8, capable of identifying
potholes accurately in real-world scenarios. This system can
significantly aid road maintenance authorities in identifying
potholes promptly, facilitating timely repairs and ultimately
leading to safer and smoother road conditions for drivers and
pedestrians alike. Fig. 1. (a): Sample image of pothole 1

Keywords— Pothole detection; deep learning; YOLO model;

Convolutional Neural Network; object detection; road
maintenance.

I. INTRODUCTION
The number of deaths due to potholes on road is
increasing year by year. The statistics related to the accidents
due to potholes on road are as follows A total of 3,564 road
accidents took place across the country during 2020 and
increased to 3,625 during 2021, the Parliament was told on
Wednesday. Therefore, we should make use of technology to
reduce such accidents due to potholes on road. Autonomous
Vehicles: The rise of autonomous vehicles necessitates
robust pothole detection systems to ensure safe navigation. Fig. 1(b): Sample image of pothole 2
YOLOv8’s real, time performance and accuracy are vital for
self-driving cars to detect and avoid potholes, preventing The main causes of pothole formation include:
potential accidents or vehicle damages. Road Safety:
Potholes pose significant safety hazards to drivers and x Water: Water is the primary culprit behind pothole
pedestrians. They can cause accidents, damage vehicles, and formation. Rainwater and melting snow can seep into
lead to severe injuries or even fatalities. By employing the cracks and joints of the road surface. During
YOLOv8 for pothole detection, we can develop a real-time colder weather, this water can freeze and expand,
solution that alerts drivers about potholes ahead, al- lowing putting pressure on the road materials and weakening
them to take necessary precautions and reduce the risk of them.
accidents. Timely Maintenance: Prompt identification of x Traffic load: The constant weight and pressure of
potholes is essential for timely road maintenance. Traditional vehicles passing over the road cause fatigue and wear
methods of pothole detection, such as manual inspection or on the pavement. Heavy vehicles and high traffic
surveying, are time-consuming and labour intensive.

979-8-3503-0884-6/23/$31.00 ©2023 IEEE 1

Authorized licensed use limited to: East West University. Downloaded on July 18,2026 at 15:50:55 UTC from IEEE Xplore. Restrictions apply.
 volumes contribute to the accelerated degradation of al. On pothole detection using yolov5 and convolutional
road surfaces, increasing the likelihood of potholes. neural networks has resulted in a good accuracy to detect
potholes [1]. In another work by Dharaneeshkar et al. they
x Age and deterioration: As roads age, their materials have worked on pothole detection with the earlier versions of
degrade and become more susceptible to damage. YOLO such as YOLOv3, YOLOv2, YOLOv3-tiny and they
Over time, the combination of traffic, weather, and have compared the results of these model to find the best one
environmental factors weakens the road’s structure, [2]. Similarly in a work done by Priyanka, Manish dixit to
making it more prone to pothole formation. detect the potholes they have trained model from various
x Poor construction and maintenance: If the road is not datasets images and finally achieved an accuracy of about
con- structed properly, with insufficient compaction 98percentage [3]. In other work conducted by Sathya et al.
or poor- quality materials, it can develop weaknesses have worked on pothole detection using YOLOv3 and
that contribute to potholes. Additionally, inadequate compared these results with other previous existing models.
maintenance, such as delayed repairs or patching, can They have used mAP, precision and recall as parameters for
allow small cracks to grow into larger potholes. comparing with other models [4].
x Ground movement: Geographical features like earth- On the other hand, Tharun and Rajaram had done a work
quakes, landslides, or natural shifts in the ground can using CNN and Yolov7 network where they have also used
put stress on roads, leading to cracks and potholes. the camera and GPS location to count the number of potholes
in a particular area and update them in Google [5]. Similarly,
x Chemicals and salt: Road salts and de-icing chemicals Sathvik et al. have worked on pothole detection using
used during winter to combat ice and snow can YOLOv7 and CNN where they used YOLOv7 to annotate
corrode road surfaces, accelerating the deterioration and train the model. Finally they have achieved an F1 score
of pavement materials and contributing to potholes. of 0.51 and they observed the increase in F1 score with
increase in epochs [6]. Similarly Madhuri et al. have worked
x Tree roots: Trees planted near the road can send out
roots that disturb the road’s base and pavement, on a model which can detect potholes for this they have used
creating cracks and spaces where water can infiltrate, Convolutional Neural Network (CNN) and Deep Neural
causing potholes. Network (DNN). Finally they have obtained accuracy of
87percent and probability of 0.534 for pothole detection [7].
x Heavy rainfall or flooding: Excessive rainfall or Similarly in other work done by Ming-An Chung et al. they
flooding can erode the roadbed and wash away have used YOLOv7 to implement a visual impairment
materials, leaving voids beneath the surface, which guidance system using pothole analysis so that the visually
can eventually collapse and form potholes. impaired people can get an alert to avoid obstacles on road
[8].
To prevent and minimize pothole formation, regular road
maintenance and repair are essential. This includes filling Similarly in other work done by Rohit Ranjan et al. they
cracks, resurfacing damaged areas, and addressing used yolo algorithm to detect pothole and alert the drivers by
underlying drainage issues to minimize water infiltration. giving a beep sound. To accomplish this the vehicles must be
equipped with a camera to capture real-time images of road
B. Objective [9]. In other work performed by yang zou et al. they used
YOLOv5 model for the detection of potholes on road and has
Getting better and accurate bounding boxes around the
given a considerable amount of improvement in the result
pothole. To detect pothole faster so that it can be used in real
compared to previous works at that time [10]. Similarly on
time applications. Preprocess data for training: Perform data
the other hand there was research going on pothole detection
preprocessing techniques to ensure optimal training of the
using techniques such as Artificial neural networks and deep
YOLOv8 model. This includes image augmentation,
neural networks to detect the potholes accurately and in less
resizing, and data normalization to enhance model
amount of time [11], [14], [15], [16].
generalization. Address false positives and false negatives:
Analyse and mitigate false positive and false negative Omer and Pradeep have performed a work on pothole
detections to enhance the overall reliability and precision of detection using YOLOv4 where they have used annotated
the pothole detection system. images to train the model and finally results were evaluated
based on recall, precision and mAP, this model can also be
The paper is organized into six sections. Section II
used with images and videos [12]. Whereas in the other
presents the literature survey of latest technologies for the
works done based on YOLO for pothole detection, the model
detection of potholes. In section III, Requirements and
is able to detect the potholes but taking a long time for
specifications for the model has been discussed such as Data
detection which is not suitable for real time applications
base used and YOLOv8 model specifications. In section IV,
[13],[17].
the design and implementation of the proposed model is
discussed. In section V, Results from the proposed model are
presented and we have also made a comparative analysis of III. REQUIREMENTS AND SPECIFICATION
various variants of YOLOv8 models used such as YOLOv8
nano, small and medium. Finally, section VI presents the A. Database
conclusion and future scope of the proposed model. We used the Kaggle Pothole dataset and included some
real-world and online image examples. We used 631 photos
II. LITERATURE SURVEY in total for this operation. 562 of those photos were utilized
to train the model. 11 photos were used to test the model,
Literature has witnessed several advancements in pothole while 58 images were utilized for validation. photos with a
detection methodologies. In a work done by Au Yang her et single pothole and photos with many potholes are both

2
Authorized licensed use limited to: East West University. Downloaded on July 18,2026 at 15:50:55 UTC from IEEE Xplore. Restrictions apply.
included in the dataset. Additionally, we have made sure that convolutional neural network that can be divided into two
the training photos include potholes of various forms and main parts: the backbone and the head. A modified version
sizes. We used the Kaggle Pothole dataset and included some of the CSP- Darknet53 architecture forms the backbone of
real-world and online image examples. We used 631 photos YOLOv8. This architecture consists of 53 convolutional
in total for this operation. 562 of those photos were utilized layers and employs cross-stage partial connections to
to train the model. 11 photos were used to test the model, improve information flow between the different layers. The
while 58 images were utilized for validation. photos with a head of YOLOv8 consists of multiple convolutional layers
single pothole and photos with many potholes are both followed by a series of fully connected layers. These layers
included in the dataset. Additionally, we have made sure that are responsible for predicting bounding boxes, object scores,
the training photos include potholes of various forms and and class probabilities for the objects detected in an image.
sizes. One of the key features of YOLOv8 is the use of a self-
attention mechanism in the head of the network. This
B. YOLOv8 mechanism allows the model to focus on different parts of
YOLO (You Only Look Once) is an object detection the image and adjust the importance of different features
algorithm that utilizes a single neural network to based on their relevance to the task. Key Features of
simultaneously predict bounding boxes along with class YOLOv8
probabilities for each object in an image. YOLOv8 is the Multiple features are to be focused on in YOLOv8. Here
latest version of the YOLO algorithm, which outperforms are some key features of YOLOv8:
previous versions by introducing various modifications such
as spatial attention, feature fusion, and context aggregation x Improved Accuracy: YOLOv8 improves object
modules. YOLO (You Only Look Once) is a popular real- detection accuracy
time object detection algorithm in computer vision. It was x compared to its predecessors by incorporating new
first introduced in 2015 by Joseph Redmon et al. YOLO is techniques and optimizations.
known for its speed and accuracy, making it suitable for
various applications like autonomous vehicles, surveillance x Enhanced Speed: YOLOv8 achieves faster inference
systems, and image recognition tasks. YOLOv8 was speeds than other object detection models while
developed by Ultralytics an organization which is working main- training high accuracy.
on AI enabling modules.
The architecture of YOLOv8 builds upon the previous
versions of YOLO algorithms. YOLOv8 utilizes a

Fig. 2. Architecture of YOLOv8

3
Authorized licensed use limited to: East West University. Downloaded on July 18,2026 at 15:50:55 UTC from IEEE Xplore. Restrictions apply.
 x Multiple Backbones: YOLOv8 supports various Dataset Collection: Collect a dataset of images that
backbones, such as Efficient Net, Res Net, and CSP contain potholes. These images should be labeled with
Darknet, giving users the flexibility to choose the best bounding boxes around each pothole to indicate their
model for their specific use case. location.
x Adaptive Training: YOLOv8 uses adaptive training to Data preprocessing: In this phase, we attempt to delete
optimize the learning rate and balance the loss any photographs that do not clearly show a pothole, are
function during training, leading to better model redundant, or show a pothole that is directly or indirectly
performance. hidden by another object. Since the entire model is based on
the data that has been analyzed, this stage is extremely
x Advanced-Data Augmentation: YOLOv8 employs important. Therefore, if any errors are made during
advanced data augmentation techniques such as Mix- preprocessing, the model picks them up and as a result,
up and Cut Mix to improve the robustness and performs worse.
generalization of the model.
Split Dataset: Divide the dataset into training, validation,
x Customizable Architecture: YOLOv8’s architecture is testing sets. The training set will be used to train the
highly customizable, allowing users to easily modify YOLOv8 model, and the validation set will be used to
the model’s structure and parameters to suit their monitor the model’s performance during training. Finally,
needs. we will test the trained model on real-time data or any new
x Pre-Trained Models: YOLOv8 provides pre-trained data which the model has not seen before. in this project we
models for easy use and transfer learning on various have divided the dataset as follows: Total number of images
datasets used = 631 Number of images used for training = 562
Number of images used for Validation = 58 Number of
images used for testing = 11.
IV. DESIGN AND IMPLEMENTATION
Configure YOLOv8 Architecture: YOLOv8 is a deep
The block diagram of the proposed pothole detection neural network architecture. You can use pre-trained weights
model is shown in Figure 3. from previous YOLO versions or start from scratch. In this
project we have trained the weights from scratch, Configure
the architecture to output bounding box coordinates and class
predictions for potholes. We have also used variations of
model such as nano, small and medium variations of
YOLOv8 model. So that we can compare the performance of
the three variations of model in detecting potholes.
Training: Train the YOLOv8 model on the training
dataset, adjust hyperparameters like learning rate, batch size,
and number of epochs. In this step the model learns how to
detect the potholes in a real time based on the training data
given to it. Setting appropriate values for the
hyperparameters is also important in the process of learning.
Validation: Evaluate the model’s performance on the
validation dataset to monitor its accuracy and detect
overfitting or underfitting issues and based on the scores of
Fig. 3. Block diagram of proposed methodology accuracies and other evaluation metrics we try to change the
hyperparameters values or we may sometime change the
dataset used to improve the performance of the model.
A. Implementation
Post processing: After the model predicts bounding
Implementing pothole detection using YOLOv8 involves boxes for potholes, apply non-maximum suppression to
several steps. YOLO is an object detection algorithm that can remove redundant and overlapping detections. You can also
be adapted for pothole detection by training it on a dataset set a confidence threshold to filter out low confidence
containing images with labelled potholes. Here’s a high-level predictions.
overview of the steps:

Fig. 4. Results of training process

4
Authorized licensed use limited to: East West University. Downloaded on July 18,2026 at 15:50:55 UTC from IEEE Xplore. Restrictions apply.
 Inference: Use the trained YOLOv8 model to detect The result corresponding to YOLOv8 nano model is
potholes in new images. The model will output bounding shown in Figure 7. The small model is able to detect two
boxes and corresponding confidence scores for each detected potholes in the image. The values 0.34,0.29 in the Figure 7
pothole. denotes the confidence scores of the pothole detected.
The result corresponding to YOLOv8 nano model is
V. RESULTS shown in Figurvalue 0.28 in the Figure 8 denotes the
The training process of the pothole detection model using confidence score of the pothole detected.
YOLOv8 nano model is shown in Figure 4. Similarly we
have performed the same process for two other versions of
YOLOv8 which are YOLOv8 small and YOLOv8 medium.
The result of testing an image with no pothole is shown in
Figure 5. The model has been successfully working without
any errors.

Fig. 8. Result of YOLOv8 Medium model

The comparative analysis of various YOLOv8 versions is

Fig. 5. Image with no pothole shown in Table 1.

The result of an image with pothole is shown in Figure 6. TABLE I. COMPARATIVE ANALYSIS
The result corresponding to YOLOv8 nano model is shown
in Figure 6. The nano model performed well with multiple Model Precision Recall
pothole detection. It has given better accurate results in less YOLOv8 nano 0.55 0.6
amount of time. In the Figure 6 0.40,0.27,0.33 denotes the YOLOv8 small 0.45 0.35
confidence scores of the potholes. YOLOv8 medium 0.5 0.5

VI. CONCLUSION AND FUTURE SCOPE

This approach allows us to identify potholes in any image
or video. This model can identify potholes in a short period
of time, enabling real-time applications. The drawbacks of
earlier models, such as mistaking the shadows of other
roadside items for potholes, have been overcome in this
model. Similar to how the previous model occasionally
mistook manhole covers for potholes, the current model is
resistant to this and has a high degree of confidence in its
ability to identify potholes. The model can help elderly
drivers who have trouble spotting potholes on the road. The
YOLOv8 nano model gave good results in detecting multiple
potholes in a single image but in some cases the medium
Fig. 6. Result of YOLOv8 Nano model model gave best results. So, based on the input image one of
the three models are working better. For majority of input
images the nano model gave better results but in some cases
small and medium models also given better results.
Further this model can be improved by training it with
a larger dataset containing relevant images. The accuracy and
mAP (mean average precision) can be further improved.
Detecting of potholes in night times is a challenge due to
inadequate lighting. Therefore, we can further extend the
model to detect potholes in night times.

REFERENCES
[1] Au Yang Her; Weng Kean Yew; Pang Jia Yew; Melissa Chong Jia
Ying. “Real-time pothole detection system on vehicle using improved
YOLOv5 in Malaysia”. In: IEEE Transactions on Pattern Analysis
Fig. 7. Result of YOLOv8 small model and Machine Intelligence 22.4 (2022), pp. 348–357.

5
Authorized licensed use limited to: East West University. Downloaded on July 18,2026 at 15:50:55 UTC from IEEE Xplore. Restrictions apply.
[2] Dharneeshkar J Soban Dhakshana V Aniruthan S A Karthika R Latha International Conference on Information Technology, Big Data and
Parameswaran. ”Deep Learning based Detection of Potholes in Indian Artificial Intelli- gence (2023)
Roads using YOLO” In: 2020 International Conference on Inventive [11] Narayana Darapaneni,Naresh Suresh Reddy,Anitha Urkude,Anwesh
Computation Technologies (2020), pp. Reddy Paduri,Arati Alok Satpute,Aakash Yogi,Dilip Krishna Nate-
[3] Priyanka Gupta, Manish Dixit. ”Image-based Road Pothole Detection san,Sarang Surve,Utkarsh Srivastava. ”Pothole Detection Using Ad-
using Deep Learning Model”. In: 2022 14th International Conference vanced Neural Networks” In: 2021 IEEE 12th Annual Information
on Computation Intelligence and Communication Network (2022) Tech- nology, Electronics and Mobile Communication Conference
[4] R. Sathya, B. Saleena, B. Prakash. ”Pothole Detection Using (2021)
YOLOv3 Model” In: 2023 IEEE International Students’ Conference [12] Mohd Omar,Pradeep Kumar. ”Detection of Roads Potholes using
on Electrical, Electronics and Computer Science (2023) YOLOv4” In: 2020 International Conference on Information Science
[5] E Sai Tarun Kumar Reddy, Rajaram V. ”Pothole Detection using and Communications Technologies (2020)
CNN and YOLO v7 Algorithm” In: 2022 6th International [13] K. Pushpa Rani,Mohammad Arshad,A. Sangeetha. ”Pothole
Conference on Electronics, Communication and Aerospace Detection Using YOLO (You Only Look Once) Algorithm” In: 2022
Technology (2022) International Conference on Advancements in Smart, Secure and
[6] Madarapu Sathvik,G. Saranya,S. Karpagaselvi. ”An Intelligent Con- Intelligent Comput- ing (2022)
volutional Neural Network based Potholes Detection using Yolo-V7” [14] Rohitaa R,Sangoju Shreya,Amutha R. ”Intelligent Deep Learning
In: 2022 International Conference on Automation, Computing and based Pothole Detection and Reporting System” In: 2021 Fourth
Renewable Systems (2022) International Conference on Electrical, Computer and
[7] Madhuri Pagale,Sunanda Mulik,Richa Purohit,Anuradha Communication Technologies (2021)
Thakare,Sulaxan Jadhav. ”Design and Implementation of Pothole [15] Anup Kumar Pandey,Rahat Iqbal,Saad Amin,Tomasz Maniak,Vasile
Detection Instance Detection Based on A Unified Approach For Palade, Charalampos Karyotis. ”Deep Neural Networks Based
Globally Roads” In: 2023 1st International Conference on Innovations Approach for Pothole Detection” In: 2021 4th International
in High Speed Communication and Signal Processing (2023) Conference on Signal Processing and Information Security (2021)
[8] Ming-An Chung,Sung-Yun Chai,Ming-Chun Hsieh,Chia-Wei [16] Bhavan Kumar S B,Guhan S,Manyam Kishore,Santhosh R,Alfred
Lin,Chia- Chun Hsu,Shang-Jui Huang,Kai-Xiang Chen. ”Road Daniel J. ”Deep Learning Approach for Pothole Detection - A
Pothole Detection Algorithm and Guide Belt Designed for Visually Systematic Review” In: 2023 Second International Conference on
Impaired” In: 2023 IEEE 3rd International Conference on Electronic Electronics and Renewable Systems, pp.1-4, 2023.
Communications, Internet of Things and Big Data (2023) [17] Kavitha R,Nivetha S. ”Pothole and Object Detection for an
[9] D Rohit Rajan,Mohammad Khaja Faizan,Rajinikanth Kundelu,Neha Autonomous Vehicle Using YOLO” In: 2021 5th International
Nandal, Vasu Sena Gunda. ”Deep Learning Based Pothole Detection” Conference on Intel- ligent Computing and Control Systems
In: 2023 International Conference on Emerging Smart Computing and [18] [Link]
Informatics (2023) Architecture, visualisation made by GitHub user RangeKing.
[10] Yang Zou,Jiangcheng Wan,Binjie Wang. ”Lightweight pothole
detection method based on improved YOLOv5” In: 2023 IEEE 3rd

6
Authorized licensed use limited to: East West University. Downloaded on July 18,2026 at 15:50:55 UTC from IEEE Xplore. Restrictions apply.