Application of Deep Learning Techniques in Negative Road

Anomalies Detection

Jihad Dib

, Konstantinos Sirlantzis

and Gareth Howells

School of Engineering and Digital Arts, Jennison Building, University of Kent, Canterbury CT2 7NT, U.K.

Keywords: Pothole Detection, Road Anomaly, Deep Learning, Deep Neural Network, Convolutional Network,

Image Processing, Object Detection, Object Classification.

Abstract: Negative Road Anomalies (Potholes, cracks, and other road anomalies) have long posed a risk for drivers

driving on the road. In this paper, we apply deep learning techniques to implement a YOLO-based (You Only

Look Once) network in order to detect and identify potholes in real-time providing a fast and accurate

detection and sufficient time for proper safe navigation and avoidance of potholes. This system can be used

in conjunction with any existing system and can be mounted to moving platforms such as autonomous vehicles.

Our results show that the system is able to reach real-time processing (29.34 frames per second) with a high

level of accuracy (mAP of 82.05%) and detection accuracy of 89.75% when mounted onto an Electric-

Powered Wheelchair (EPW).

1 INTRODUCTION

Negative Road Anomalies is the term we have chosen

to describe potholes, cracks, and any anomaly located

at a negative position of the road surface.

Potholes pose the highest risk in all negative road

anomalies as they are a danger to drivers when

driving on roads, and im some cases motorways.

They could cause severe injury to the driver in form

of neck pain, back pain, whiplash and more severe

health risks. Not to mention, the damage which could

be caused to the car’s mechanical system and tires

putting the driver under numerous risks of accidents

and even threatening their life as a result of a torn tire

or other mechanical damage which could be caused

to the vehicle driven when passing over a pothole at

high speed. Potholes were also a significant limitation

to the driverless car advancement projects due to the

stochasticity of their nature and the difference in their

depth and severity making them hard to identify and

detect, and rendering many detection techniques

futile as they sometimes contain some characteristics

which could fall within the limitations of the

detection techniques (for example, water-filled

https://orcid.org/0000-0002-3070-9673

https://orcid.org/0000-0002-0847-8880

https://orcid.org/0000-0001-5590-0880

potholes cannot be detected via ultrasonic sensing

techniques).

In our ongoing project, we apply deep learning

techniques in order to detect and identify potholes.

Our project is mainly focused on object detection

neural networks which can be used in real-time in

order to detect and classify potholes from the video

stream obtained through the use of an RGB Camera,

and to provide a fast and reliable detection method

which allows sufficient time and distance for a safe

avoidance and navigation of manual and autonomous

vehicles and moving platforms.

YOLO (Redmon et al.,2015) (You Only Look

Once) is the candidate network which is used in this

project due to its high accuracy and fast performance,

especially in real-time detection scenarios. The

project was implemented over the Darknet (Redmon

et al.,2020) environment which was developed by the

authors and creators of YOLO and was optimized and

tested in real-time scenarios where it returned

significant promising results.

Many attempts to detect potholes were made, and

different technology was used in order to implement

solutions to the proposed problem. Some solutions

were implemented via the use of laser imaging as

Dib, J., Sirlantzis, K. and Howells, G.

Application of Deep Learning Techniques in Negative Road Anomalies Detection.

DOI: 10.5220/0011336000003332

In Proceedings of the 14th Inter national Joint Conference on Computational Intelligence (IJCCI 2022), pages 475-482

ISBN: 978-989-758-611-8; ISSN: 2184-3236

475

Figure 1: Pothole Detection System Diagram.

input (Yu and Salari,2011), (Vupparaboina et.

al,2015) where different regions of the laser colours

were extracted with the help of image processing,

thermal imaging was also used by fusing thermal

imaging and convolutional neural networks (Aparna

et al.,2019) while others used visible light RGB

cameras with supervised learning techniques by

analysing the road’s surface feature in order to

classify potholes via HOG feature extraction (Azhar

et al.,2016), while (Koch and Brilakis,2011) used

segmentation in order to split mages into two

categories (defective and non-defective) via the use

of the histogram approach with shape-based

thresholds [ (Ryu et. al,2015), (Schiopu et. al,2016),

and (Saluja et. al,2019) on the other hand used video

sequences taken by RGB cameras via their own

thresholding algorithm which considers potholes as

the images with high-intensity values. Other

techniques such as Probabilistic Generative Models

(PGM) fused with Support Vector Machine (SVM)

techniques were used in order to detect the probability

of occurrence of a road crack via the intensity details

was used by (Ai et. al,2018) while (Youquan et.

al,2011), (Zhang et. al,2016) and (Li et. al,2018)

have relied on stereo-vision techniques in different

setups in order to detect potholes via their shapes, or

by estimating the difference between the surface of

the road and the surface of the pothole. (Moazzam et.

al,2013) used a depth camera via the detection of the

area, depth, length, width and volume of the pothole,

while (Avellaneda and López-Parra,2016), (Buttlar

and Islam,2014), and (Forslöf and Jones,2015) have

used the accelerometer, compass and GPS found in

mobile phones in order to achieve post-pothole

detection, when (Chellaswamy et. al,2018) have used

ultrasonic sensors. A more detailed review can be

found in our previous publication A Review on

Negative Road Anomalies (Dib et. al,2020)

The previously mentioned techniques were all

limited due to the fact that post-detection of potholes

cannot be used in order to avoid potholes. Ultrasonic

sensors, laser-imaging techniques, surface difference-

based techniques and depth camera-based techniques

are limited when it comes to water-filled potholes as

water can be reflective, and there will be no or

insignificant surface difference between the surface

of the water and the surface of the pothole.

This paper describes the current progress of the

negative anomaly detection project and proposes the

use of a normal RGB camera where the stream being

fed to a custom-trained YOLO network which will

achieve the real-time detection of the pothole. A

dataset of pothole images was collected, preprocessed

and used to train the neural network in order to fulfil

the task required.

2 PROPOSED APPROACH

We propose the use of deep learning neural networks,

mainly YOLO developed using the Darknet

environment having an input obtained by an RGB

camera mounted on any moving platform. This could

be a self-driven car, driven car, truck, motorcycle, or

even an electric-powered wheelchair or a robot.

A core computing unit will be mounted onto the

moving platform. The RGB Camera’s video feed will

be processed by the computing unit which will be

running the Robot Operating System (ROS) (ROS

Wiki,2020) and will process the video feed and feed

it into the pothole detection system (Figure 1).

First, the captured video feed will be pre-

processed by converting the feed into RGB format (if

it is not already in RGB format), then, the frames are

downscaled to 416x416 pixels. Padding is used in the

event where the downscaled frames have either a

height or a width less than 416 pixels. Then, the

frames will be processed by the YOLO convolutional

neural network in order to detect and localise the

potholes based on the features which the network is

trained to detect. Potential pothole candidates will be

detected, and the probability of the candidate being a

true positive detection will be calculated according to

the formulas discussed in Part 3 of this paper. If the

ROBOVIS 2022 - Workshop on Robotics, Computer Vision and Intelligent Systems

476

network’s confidence is more than 0.7 (70%) the

detected object is considered a pothole, it will be

considered a positive detection, the x1,y1,x2,y2

coordinates of the bounding box are calculated, and

the bounding box is drawn around it marking its

location within the video frame. At this stage, this

approach ensures reliability, scalability, and high

performance.

Our approach’s main goal is to ensure real-time

detection with the least possible computing and

power requirements. This will enable the use of the

system in real-time scenarios without the need to rely

on equipment with high computational power which

could drain the battery used. This will ensure the

ability to use our system in real-life scenarios and

potentially provide a standard platform for both

negative and positive obstacle detection and

avoidance.

3 EXPERIMENT

For this experiment (Figure 2), different versions of

the YOLO network were trained on a self-collected

dataset which includes photos from different

scenarios which could be encountered within an

everyday usage of any vehicle/moving platform.

The images dataset which has been used to train

the neural network was collected during the research

phase using a Samsung Galaxy Note 8 phone camera

taking images at 13 Megapixels. Images were

collected within Kent County in the United Kingdom

from different roads and cities in different weather

conditions (sunny, cloudy, and rainy days and nights).

Figure 4 presents a sample from our dataset:

(a) represents a general-looking pothole, i.e. a

hole within the tarmac in an quasi-circular

shape.

(b) represents a shallow circular broken

surface/crack within the tarmac floor which is

usually hard to detect by laser-based systems.

with water and dirt with a completely different

pattern than the tarmac surrounding it. This

cannot be detected by laser-based or sonar-

based systems.

(d) represents a random-shaped pothole filled

with water which is nearly clear and located

on the side of the road where the double lines

are clearly shown making it very difficult to be

detected by normal image-based systems,

laser-based systems, and sonar-based systems.

(e) represents a stochastic-shaped broken side of

the road filled with rabble which is also very

challenging for image-based, laser-based, and

sonar-based systems due to the stochastic

shape, and the reflective randomly-located

rabble.

This dataset will be made available online at a later

stage.

Figure 2: Network Training Process.

The collected images were pre-processed by

downscaling them to 30% in order to obtain a width

of around 415 pixels and then they were labelled

individually using the labelling tool LabelImg

(GitHub,2020) for Python ensuring that all the

surrounding boxes cover the exact corners of the

pothole without adding a lot of background data

which could cause any diversion in the learning

process.

Figure 3: Labeling the dataset images via LabelImg.

Application of Deep Learning Techniques in Negative Road Anomalies Detection

477

The network training phase was done by using

more than one platform and more than one version of

YOLO in order to collect numerous test results and

perform benchmarking.

All training experiments were done using an Intel

Core i7 desktop with an NVIDIA RTX2080 6GB

memory graphical processing unit (GPU) running a

Windows 10 OS with Anaconda (Anaconda,2020) as

a platform to run the Python environment on

Windows.

In order to calculate the accuracy of the detection

denoted by precision, we have considered the

detection a true positive if the maximum overlap

between the detected region (detected box) and the

original annotation (annotated box of the RoI (Region

of interest i.e. pothole) within the validation dataset)

is larger than or equal to the Intersection over Union

(IoU) (Rosebrock,2020) which is the area of overlap

between the detected region and the annotation region

divided by the area of the union which is the union of

both areas combining the detected region and the

annotated region:

After calculation, the precision of every detection,

the mean average precision (mAP) is calculated via

the calculation of the precision envelope, the area

under the curve (points where the recall changes) and

then the summing of those values. This has been

described extensively in (Medium,2020)

The precision and recall formulas are as follows:

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =

𝑇𝑃

𝑇𝑃 + 𝐹𝑃

𝑅𝑒𝑐𝑎𝑙𝑙 =

𝑇𝑃

𝑇𝑃 + 𝐹𝑁

where TP is the number of true positives, FP is the

number of false-positive, and FN is the number of

false negatives.

The mean average precision (mAP) is the mean of

the average precisions calculated, i.e. the sum of all

average precisions divided by the number of

detections.

𝑚𝐴𝑃 =

∑

𝐴𝑃





𝑁

where N is the number of detections, and AP is the

Average Precision (described in (Anaconda,2020))

In addition to the previous values, we have

recorded the frame rates achieved by the algorithm

(a) (b) (c) (d) (e)

Figure 4: Samples images taken from our manually-

collected dataset.

when tested in order to assess whether the detection

is in real-time or not.

In our training experiments, we have set the IoU

to 70 so detection is considered a true positive when

there is an overlap that is more than 70% (YOLO uses

a default IoU threshold of 0.3 (30%) which we have

raised to 0.7).

We have attempted more than one training

experiment and split them into three main sets. In

every set, we have tried several different number of

training and validation datasets as follows:

• 80% of the total number of images in the

dataset is used as training data, the rest is used

as validation images.

• 70% of the total number of images in the

dataset is used as training data, the rest is used

as validation images.

• 60% of the total number of images in the

dataset is used as training data, the rest is used

as validation images.

We have also attempted more than one different

learning rate, as follows:

• Learning Rate = 1e-4

• Learning Rate = 1e-5

• Learning Rate = 9e-5

The first training set of experiments was made

using YOLOv2 (Redmon et al., 2015) trained using

Keras (Team K.,2020) (Keras Website) open-source

neural network library with the following training

parameters:

• Training Images: 574 (80% of the dataset)

• Validation Images: 143 (20% of the dataset)

• Learning Rate = 1e-4

• IoU = 0.7

The training resulted in a best mAP of 71.5%, 102

true positives, 41 false positives, and upon testing the

neural network performance on a “challenging” input

with more than one pothole present, the following

results were obtained:

ROBOVIS 2022 - Workshop on Robotics, Computer Vision and Intelligent Systems

478

Figure 5: YOLOv2 Test Results.

It is evident from the first test results (Figure 5)

that this network is not really fit for purpose as its

result is a larger region of interest (ROI) detected

versus the optimal ROI which should have been

detected. A larger ROI can be acceptable as it means

that the network has detected the object of interest

which means that avoidance can be achieved,

however, our aim is to obtain the most optimal

detection possible. In addition, the network has

detected only 6 potholes within the first example and

only two in the second example along with a larger

ROI where two potholes were considered as only one.

Not to mention that the detection accuracy is only

0.7302 for the bottom pothole in the second example

which is 73.02% for the most evident pothole. As for

the frame rate achieved, it was 44 frames per second

as reported by the algorithm.

The second training set of experiments was made

using YOLOv3 (Redmon and Farhadi, 2018) trained

using Keras with the same parameters as the previous

training.

The network’s best mAP was 75.48%, the number

of true positives was 108 along with 35 false positives

and the test input returned the following results:

Figure 6: YOLOv3 Test Results.

The results obtained via YOLOv3 look slightly

more promising than YOLOv2 as the network has

detected 7 potholes in the first example and only two

in the second but with tighter ROIs covering the exact

borders of the potholes. In the second example, the

evident pothole’s detection accuracy was 0.7719

which is 77.19% which shows a slight improvement

from the first network used. The frame rate was

almost the same as the previous test (47 frames per

second)

The training set of experiments was made using

YOLOv4 (Bochkovskiy et. al, 2020) trained using

Darknet environment with the same parameters as the

previous training. The network’s best mAP was

82.05%, the number of true positives was 117, and the

false positives were 26. The test input returned the

following results:

Figure 7: YOLOv4 Test Results.

The results obtained via YOLOv4 were the most

promising results as the network has detected 12

potholes within the first example along with three

potholes in the second example as the network has

identified the two small potholes as an additional

partial pothole. The accuracy of the detection for the

evident pothole is 0.8975 which is 89.75% which was

the highest obtained accuracy within the tests

aforementioned. This experiment was achieved with

a frame rate of 49 frames per second which is the

highest frame rate achieved in all of our experiments.

Knowing that the speed limit for vehicles in

residential areas in England is 30 mph (48.28 km/h or

13.411 m/s), and for other moving platforms such as

electric-powered wheelchairs (EPWs) is 4 mph (6

km/h or 1.67 m/s) offroad (on a footpath, on a

pavement, etc.) and 8mph (12 km/h or 3.3333 m/s) on

the road (on tarmac) we can easily calculate our

detection-rate via the formula:

𝐷𝑒𝑡𝑒𝑐𝑡𝑖𝑜𝑛 𝑅𝑎𝑡𝑒 =

𝐹𝑟𝑎𝑚𝑒 𝑅𝑎𝑡𝑒

𝑀𝑎𝑥 𝑆𝑝𝑒𝑒𝑑

Application of Deep Learning Techniques in Negative Road Anomalies Detection

479

Figure 8: Real-time testing of the YOLO Network applied on an online video [29] where three consecutive frames have been

extracted. In the test above, the precision rate of the detection is respectively: pothole 1.a: 85%, pothole 1.b: 73%, and pothole

1.c: 95% pothole 2.a: 56%, pothole 2.b: 82%, and pothole 2.c: 97%. pothole 3.a: 53%, pothole 3.b: 70%, and pothole 3.c:

96%. The Average Frame Rate achieved is 52 fps which is considered real-time.

By applying this formula to vehicles in rural areas,

we can conclude that our detection rate is

13.411

=3.65

i.e. 3.65 frames per meter, as for the off-road

wheelchair, it is 29.34 frames per meter and the on-

road wheelchair detection rate is 14.70 frames per

meter. These detection rates achieved are more than

sufficient for safe navigation and avoidance of the

pothole as the lowest detection rate (3.65 achieved

when the system is mounted to cars in rural areas)

allows at least 3 frames to be detected within every

meter and as is known, 49 frames per second are

larger than the commonly-used real-time threshold

for frame-rate which is 30 frames per second. Figure

8 represents the real-time results obtained when using

an mp4 video (ITV News YouTube Channel, 2018)as

input to the network. In order to demonstrate the

result, we have extracted three consecutive frames

from the resulting video showcasing the detection rate

achieved with the real-time frame rate of an average

of 52 frames per second.

In addition to the previous training attempts, a

separate attempt was made with the same training and

validation ratios, along with different learning rates in

order to attempt to find additional methods of

improving the detection and studying the effect of the

learning rate and the training/validation ratio on the

mean average precision of the network’s detection

performance. The results were as follows:

Table 1: YOLO with Learning Rate = 1e-4.

Ratio

Trainin

/Validation

0.8 0.7 0.6

mAP 0.765 0.725 0.702

True Positives 109 156 302

False Positives 34 59 128

Table 2: YOLO with Learning Rate = 2e-4.

Ratio

Trainin

/Validation

0.8 0.7 0.6

mAP 0.5939 0.474 0.455

True Positives 85 102 196

False Positives 58 113 234

Table 3: YOLO with Learning Rate = 1e-5.

Ratio

Trainin

/Validation

0.8 0.7 0.6

mAP 0.521 0.497 0.434

True Positives 75 107 187

False Positives 68 108 243

Table 4: YOLO with Learning Rate = 9e-5.

Ratio

Trainin

/Validation

0.8 0.7 0.6

mAP 0.616 0.605 0.603

True Positives 88 130 259

False Positives 55 85 171

ROBOVIS 2022 - Workshop on Robotics, Computer Vision and Intelligent Systems

480

Table 5: YOLO with Learning Rate = 8e-5.

Ratio

Trainin

/Validation

0.8 0.7 0.6

mAP 0.681 0.671 0.532

True Positives 97 144 229

False Positives 46 71 201

Table 6: YOLO with Learning Rate = 7e-5.

Ratio

Trainin

/Validation

0.8 0.7 0.6

mAP 0.617 0.582 0.532

True Positives 88 125 229

False Positives 55 90 201

Table 7: YOLO with Learning Rate = 6e-5.

Ratio

Trainin

/Validation

0.8 0.7 0.6

mAP 0.615 0.647 0.543

True Positives 88 139 233

False Positives 55 76 197

As we can observe from this experiment, the

detection’s mAP is generally larger when the number

of images used for training is larger which can be

observed when comparing the mAP values at the

different training/validation ratios. In addition, we

can observe that setting the learning rate to 1e-4

returned the highest mAP value. However, 8e-5 and

9e-5 did return acceptable results at a

training/validation ratio of 0.8 which leads us to

conclude that a variable learning rate would be more

ideal in respect to the validation loss, this could

improve the results obtained and will be assessed in

future work.

4 CONCLUSIONS

In this paper, we have implemented a deep learning-

based system which detects and localises different

types of potholes regardless of the stochasticity in

their shapes, textures, patterns, and colours in real-

time (i.e. high frame rates achieved within the

experiments undertaken), and with high accuracy.

The results obtained show that the accuracy of

the detection was very high even in the case of water-

filled potholes which is usually considered the main

limitation of many sensing techniques. Not to

mention that the detection rate and the frame rates

achieved were more than sufficient for our detection

rate to be considered real-time providing sufficient

detection speed and distance for a safe navigation and

avoidance of potholes.

We can also conclude that the training results

could be improved by varying the learning rate

throughout the learning process, and by increasing the

size of the training dataset used.

The next steps will focus on further use of deep

learning object detection convolutional neural

networks. In future work, there will focus on

including more functionalities, such as object

localization in real-world 3D coordinates and more

additional functionalities via the use of additional

sensing techniques fused by a data fusion algorithm.

This algorithm will combine the use of more than one

sensing technique in such a way that every technique

used will cover the other techniques’ weaknesses and

limitations via the use of multimodal sensing

techniques combined with deep learning. The

proposed algorithm could be the backbone of a wide

range of systems and it can be used to make decisions

ensuring safe navigation of the moving platform

when needed.

ACKNOWLEDGEMENTS

We would like to acknowledge the help Engineering

and Physical Sciences Research Council (EPSRC)

who funded the project via studentship reference

2112938 and Assistive Devices for empowering dis-

Abled People through robotic Technologies

(ADAPT). ADAPT is selected for funding by the

INTERREG VA France (Channel) England

Programme which is co-financed by the European

Regional Development Fund (ERDF).

REFERENCES

J. Redmon, S. Divvala, R. Girshick. and A. Farhadi, 2015.

IEEE Conference on Computer Vision and Pattern

Recognition (CVPR). You Only Look Once: Unified,

Real-Time Object Detection. [online] IEEE. Available

at: <https://arxiv.org/abs/1506.02640> [Accessed 1

July 2020].

J. Redmon, 2020. Darknet: Open Source Neural Networks

In C. [online] Pjreddie.com. Available at:

<https://pjreddie.com/darknet/> [Accessed 1 July

2020].

X. Yu and E. Salari, 2011. ‘‘Pavement pothole detection

and severity measurement using laser imaging,’’ in

Proc. IEEE Int. Conf. Electro/Inf. Technol., Mankato,

MN, USA, pp. 1–5.

Y. Aparna, R. Bhatia, V. Rai, N. Gupta, Aggarwal, and A.

Akula, 2019. ‘‘Convolutional neural networks based

Application of Deep Learning Techniques in Negative Road Anomalies Detection

481

potholes detection using thermal imaging,’’ J. King

Saud Univ.-Comput. Inf. Sci.,, doi:

10.1016/j.jksuci.2019.02.004.

K. Azhar, F. Murtaza, M. H. Yousaf, and H.A. Habib

,2016. ‘‘Computer vision based detection and

localization of potholes in asphalt pavement images,’’

in Proc. IEEE Can. Conf. Electr. Comput. Eng.

(CCECE), pp. 1–5.

C. Koch and I. Brilakis, 2011. ‘‘Pothole detection in asphalt

pavement images,’’ Adv. Eng. Inform., vol. 25, no. 3,

pp. 507–515.

S.-K. Ryu, T. Kim and Y.-R. Kim, ‘‘Image based pothole

detection system for ITS service and road management

system, 2015.’’ Math. Problems Eng.,vol. 2015, Art.

no. 968361, doi: 10.1155/2015/968361.

I. Schiopu, J. P. Saarinen, L. Kettunen, and I. Tabus, 2016.

‘‘Pothole detection and tracking in car video

sequence,’’ in Proc. 39th Int. Conf. Telecommun.

Signal Process. (TSP), Vienna, Austria, pp. 701–706.

D. Ai, G. Jiang, L. Siew Kei, and C. Li, 2018. ‘‘Automatic

pixel level pavement crack detection using information

of multi-scale neighborhoods,’’ IEEE Access, vol. 6,

pp. 24452–24463.

H. Youquan, W. Jian, Q. Hanxing, W. Zhang, and X.

Jianfang, 2011. ‘‘A research of pavement potholes

detection based on three-dimensional projection

transformation,’’ in Proc. 4th Int. Congr. Image Signal

Process., pp. 1805–1808.

L. Zhang, F. Yang, Y. Daniel Zhang, and Y. J. Zhu, 2016.

‘‘Road crack detection using deep convolutional neural

network,’’ in Proc. IEEE Int. Conf. Image Process.

(ICIP), pp. 3708–3712.

Y. Li, C. Papachristou, and D. Weyer, 2018. ‘‘Road pothole

detection system based on stereo vision,’’ in Proc. IEEE

Nat. Aerosp. Electron. Conf. (NAECON), pp. 292–297.

I. Moazzam, K. Kamal, S. Mathavan, S. Usman, and M.

Rahman, 2013. ‘‘Metrology and visualization of

potholes using the microsoft kinect sensor,’’ in Proc.

16th Int. IEEE Conf. Intell. Transp. Syst. (ITSC),pp.

1284–1291

D. A. Casas Avellaneda and J. F. López-Parra, 2016.

‘‘Detection and localization of potholes in roadways

using smartphones,’’ DYNA, vol. 83, no. 195, pp. 156–

162.

W. G. Buttlar and M. S. Islam, 2014. Integration of Smart-

PhoneBased Pavement Roughness Data Collection

Tool With Asset Management System—National

Transportation Library. Accessed: May 20, 2019.

[Online]. Available: https://rosap.ntl.bts.gov/view/

dot/38287

Forslöf, L. and Jones, H., 2015. Roadroid: Continuous Road

Condition Monitoring with Smart Phones. Journal of

Civil Engineering and Architecture, 9(4).

C. Chellaswamy, H. Famitha, T. Anusuya, and S. B.

Amirthavarshini, 2018.‘‘IoT based humps and pothole

detection on roads and information sharing,’’ in Proc.

Int. Conf. Comput. Power, Energy, Inf. Commun.

(ICCPEIC), pp. 084–090.

Wiki.ros.org, 2020. ROS/Introduction - ROS Wiki.

[online] Available at: <http://wiki.ros.org/ROS/

Introduction> [Accessed 1 July 2020].

GitHub, 2020. Tzutalin/Labelimg. [online] Available at:

<https://github.com/tzutalin/labelImg> [Accessed 1

July 2020].

Anaconda, 2020. Anaconda | The World's Most Popular

Data Science Platform. [online] Available at:

<https://www.anaconda.com/> [Accessed 1 July 2020].

Medium, 2020. Breaking Down Mean Average Precision

(Map). [online] Available at: <https://towardsdata

science.com/breaking-down-mean-average-precision-

map-ae462f623a52> [Accessed 1 July 2020].

A. Rosebrock, 2020. Intersection Over Union (Iou)

For Object Detection - Pyimagesearch. [online]

PyImageSearch. Available at: <https://www.pyimage

search.com/2016/11/07/intersection-over-union-iou-

for-object-detection/> [Accessed 1 July 2020].

Team, K., 2020. Keras: The Python Deep Learning API.

[online] Keras.io. Available at: <https://keras.io/>

[Accessed 1 July 2020].

J. Redmon, and A. Farhadi, 2018. Yolov3: An Incremental

Improvement. [online] arXiv.org. Available at:

<https://arxiv.org/abs/1804.02767> [Accessed 1 July

2020].

A. Bochkovskiy, C. Wang, and H. Liao, 2020. Yolov4:

Optimal Speed And Accuracy Of Object Detection.

[online] arXiv.org. Available at: <https://arxiv.org/abs/

2004.10934> [Accessed 1 July 2020].

J. Dib, K. Sirlantzis and G. Howells, 2020. "A Review on

Negative Road Anomaly Detection Methods," in IEEE

Access, vol. 8, pp. 57298-57316, 2020, doi:

10.1109/ACCESS.2020.2982220.

Vupparaboina, K.K., Tamboli, R.R., Shenu, P.M., & Jana,

S., 2015. Laser-based detection and depth estimation of

dry and water-filled potholes: A geometric approach.

2015 Twenty First National Conference on

Communications (NCC), 1-6.

I. Saluja, R. Karwa, A. Bharambe, M Sabnis,

2019."Recognition and Depth Approximation of Dry &

Filled Potholes", International Journal of Innovative

Research in Science, Engineering and Technology, Vol.

8, Issue 8.

ITV News YouTube Channel,11 May 2018. Potholes

costing drivers and insurers ‘£1m a month’ | ITV News.

Available at: <https://www.youtube.com/watch?v=

MlDI-zlcc8I> [Accessed 15 September 2022].

ROBOVIS 2022 - Workshop on Robotics, Computer Vision and Intelligent Systems

482