Deepbrokenhighways: Road Damage Recognition System Using

Convolutional Neural Networks

Sebastian Peralta-Ireijo, Bill Chavez-Arias and Willy Ugarte

Universidad Peruana de Ciencias Aplicadas, Lima, Peru

Keywords:

Computer Vision, Pothole Detection, Convolutional Neural Network, YOLO, MobileNet.

Abstract:

Road damage, such as potholes and cracks, represent a constant nuisance to drivers as they could potentially

cause accidents and damages. Current pothole detection in Peru, is mostly manually operated and hardly

ever use image processing technology. To combat this we propose a mobile application capable of real-time

road damage detection and spatial mapping across a city. Three models are going to be trained and evaluated

(Yolov5, Yolov8 and MobileNet v2) on a novel dataset which contains images from Lima, Peru. Meanwhile,

the viability of crack detection through bounding box method will be put to the test, each model will be trained

once with cracks annotations and without. The YOLOv5 model was the one with the best results, as it showed

the best mAP50 across all of out experiments. It got 99.0% and 98.3% mAP50 with the dataset without crack

and with crack annotations, correspondingly.

1 INTRODUCTION

Currently, many third world countries struggle to

maintain their roads pothole and crack free. Stud-

ies in Peru like (D

avila Estrada, 2022) showed that

around eight to ten potholes are present on average,

over a one kilometer road range. Furthermore the

lifespan of these roads should be close to twenty ﬁve

years, nevertheless, the real span is about four to six

years. As written in (Ministerio de Transportes y Co-

municaciones, 2015), current government road explo-

ration procedures don’t include a detailed detection of

road damage. This turns the damage mapping into a

lengthy and manual process, if needed. That being

said, nowadays few companies devoted to roadwork

like (Tenorio Construcciones y Soluciones, 2022) use

software to map the damage present in a road before

any work is done. That motivates us to tackle the

problem of slow detection and mapping of road dam-

age.

We aim to develop an application, capable of inte-

grating real time detection and mapping of road dam-

age. Aiding in the optimization of road damage de-

tection for organizations which carry out road main-

tenance. This will allow people to beneﬁt from au-

tomation via Deep Learning models, as well as ob-

taining a detection process free of human subjectiv-

https://orcid.org/0000-0002-7510-618X

ity. Thus, making a contribution to the optimization

of road maintenance projects. While there are solu-

tions that deal with pothole and crack detection, their

method or end product doesn’t line up with our ob-

jectives. For example, vibration based models like

(Egaji et al., 2021) are successful in the damage de-

tection, but lack the capabilities of providing a visual

feedback of the damage and require the use of a vehi-

cle and sensors. Furthermore other solutions that in-

volve computer vision stop at the experimental phase,

maximizing the proposed model detection capability

through the use of innovative image processing like in

(Aparna et al., 2022). Yet, their image dataset features

do not match most of the approaches, since it involves

heatmap imaging and specialised equipment.

Some solutions that involve the integration of road

damage detection model and a application like (Dong

et al., 2022) and (Patra et al., 2021). However, in

(Dong et al., 2022) the UI design and usage wasn’t

friendly to the average person, as results were shown

on a console like interface. Also it required the use of

a small vehicle to carry out the detection which lim-

its it’s capabilities to a speciﬁc viewpoint. In (Patra

et al., 2021) the solution limits itself to the exclusive

detection of pothole as a generalized label. Meaning

that their model is binary, either a part of an image is a

pothole or not, simplifying the training of the model.

Also, road crack detection was left out of the training

of the model leaving various improvement points for

Peralta-Ireijo, S., Chavez-Arias, B. and Ugarte, W.

Deepbrokenhighways: Road Damage Recognition System Using Convolutional Neural Networks.

DOI: 10.5220/0012685600003690

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 26th International Conference on Enterprise Information Systems (ICEIS 2024) - Volume 1, pages 739-746

ISBN: 978-989-758-692-7; ISSN: 2184-4992

739

us to explore. Our approach intends to use a convolu-

tional network architecture that is optimized for real

time detection and a user friendly mobile application

to go alongside it.

The key components of our approach are a

YoloV5 model trained with some of the state of the

art road damage datasets and meant for Bounding Box

detection. This data will be expanded upon by the cre-

ation of our own dataset containing images from our

speciﬁc context, Lima and Callao, Peru. That being

said, using the mobile development framework Flut-

ter, we will develop an app that takes advantage of

the model for road damage detection (potholes of dif-

ferent sizes and depth and road cracks) and damage

mapping using one of the available modules.

Our main contributions are as follows:

- We elaborate a dataset with the characteristics of

the context in which our application will be utilized.

- We developed a mobile application that is able to

detect in real time types of potholes and cracks using

convolutional networks.

- We test the viability of road cracks detection with

a bouding box method through a quantitative analy-

sis of different models based on metrics (ex. Average

Precision and mAP50).

This paper is distributed into the following sec-

tions: First, we review related work on road damage

detection in Section II. Then, we discuss relevant con-

cepts and theories related to the background of our

research and describe in more detail our main con-

tribution in Section III. Furthermore, we will explain

the procedures performed, and the experiments con-

ducted in this work in Section IV. At the end, we will

show the main conclusions of the project and indicate

some recommendations for future work in Section V.

2 RELATED WORKS

The topic of pothole and damage segmentation has

been around for several years now. In recent years

many methods and techniques have been proposed

with the goal of getting the best results. Those articles

differ greatly in terms of the objective and ﬁnal prod-

uct. The following articles demonstrate studies and

development of Deep Learning models of interest for

our project. Some even mention a model integration

with an application, which is relevant as well. That

being said we will discuss the insights they provide,

the limitations and the differences in comparison to

our intentions.

In (Park et al., 2021) the authors perform a study

comparing the performance of various YOLO mod-

els versions under multiple metrics as precision, recall

and mAP. The ultimate goal being to ﬁnd the optimum

YOLO algorithm for pothole detection in real time.

Results showed that the YOLOv4-tiny was the best

for pothole detection. However, we also will evaluate

other real time models such as SSD and MobileNet.

Moreover, we seek to produce a model which is re-

trained for Peruvian roads. Instead in (Park et al.,

2021) the dataset is mostly web-scraped from Google,

meaning the model isn’t trained for a speciﬁc context

and results could vary greatly depending on the origin

of the image.

In (Rateke and von Wangenheim, 2021) the au-

thors propose a model based on ResNet pre-trained

in ImageNet that segmentates and differentiates the

types of damage between speed-bump, crack and pot-

hole. A transfer learning process was carried out

on two models with different hyper-parameters. The

model with the best conﬁguration was called r34-

DW, has an accuracy of 90% in eleven out of thirteen

classes. However due to the high number of labels,

two classes have a low accuracy (below 72%). In con-

trast, we aim to use more general labels, since having

various highly speciﬁc labels hinders the model per-

formance.

In (Moscoso Thompson et al., 2022) the authors

developed 6 models for evaluation on the SHREC

2022 track, which were compared to the Baseline-

DeepLav3 model. The models were trained to per-

form semantic segmentation on the road surface, de-

tecting potholes and cracks. From the proposed mod-

els the HCMUS-CPS-DLU-Net model had the best

overall results, however these models aren’t able to

differentiate between different types and severity of

the damage. As we seek to provide a more detailed

report of the road damages, it is important to be able

to identify the type and/or severity of the damage.

In (Dong et al., 2021) the authors propose a mo-

bile damage segmentation model, which is developed

to detect multiple damages in the same image, using

an optimizer called DIoU-NMS. The model that per-

forms the segmentation uses a reference object, this

object allows the model to estimate the area, width,

height and ratio of the damage. Even though this

method has an accuracy of 93.4%, for our purposes,

it is not convenient to carry a reference object and de-

pend on having a clear shot with it to use the model.

In our proposal, an accurate detection of the type and

severity of the damage is of higher priority than the

estimate of its dimensions.

In (Patra et al., 2021) a system called PotSpot was

implemented. Through an Android Application, a

CNN model and a Maps API (Application Program-

ming Interface), are integrated to be able to perform

the pothole detection and mapping. The damage de-

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

740

tection is performed exclusively on potholes, more-

over pothole detection is binary, pothole or no pot-

hole, simplifying the training of the model. That be-

ing said despite this project providing an end-to-end

result, it lacks the level of detail we aim to achieve.

The end result fails to show the user the characteris-

tics of the damage (e.g. type, severity, image of the

damage, etc).

3 REAL-TIME POTHOLE

DETECTION WITH MOBILE

DEVICES

3.1 Preliminary Concepts

In this section, the main concepts used in our work are

presented. We aim to train a real-time object detection

model, via Computer Vision methods and a Bounding

Box labeled dataset to be able to identify objects of

interest within our training images.

Deﬁnition 1 (Computer Vision(Baek and Chung,

2020)): Computer vision is a technology that ex-

tracts useful information by inputting visual data into

a computer and analyzing it. The goal of computer

vision is to extract interesting information from pat-

tern recognition, statistical learning, and projection

geometry through object detection, segmentation, and

recognition in images.

Deﬁnition 1 (Convolutional Neural Network (Baek

and Chung, 2020)). CNN aims to reduce the complex-

ity of the model and extract signiﬁcant features by ap-

plying a convolution operation. In visual data, object

detection is a technique to ﬁnd a candidate region for

a detection target to recognize a speciﬁc target and to

predict the type and location of the object (bounding

box).

Deﬁnition 2 (Vision-based Pothole Detection (Kim

et al., 2022)). A vision-based method uses images

or videos as input data and determines the presence

of potholes on the road surface by applying image-

processing and deep-learning technology. ...suitable

for determining the number and approximate shape of

potholes.

Example 1 (Vision-based Pothole Detection). Is

shown in Figure 1.

Deﬁnition 3 (Global Positioning System (Wu et al.,

2020)). GPS, which is used to record location. As

mentioned in (Pandey et al., 2022) and (Egaji et al.,

2021), location data means latitude and longitude co-

ordinate information.

Figure 1: Example of vision-based Pothole Detection via

Bounding Box labeling in (Park et al., 2021).

Example 2 (Global Positioning System). Is shown in

Figure 2.

Figure 2: Example of GPS data usage for pothole mapping

in (Patra et al., 2021).

3.2 Method

In this section, the main contributions proposed in this

project will be detailed.

3.2.1 Peruvian Road Damage Dataset

The ﬁrst contribution of this work is the creation of

a dataset containing information of potholes on the

roads of Lima, Peru. This is due to the fact that,

currently, there is no publicly available road damage

dataset of Peru. The dataset without augmentation has

618 images and three labels (Cracks, Pothole and Se-

vere Pothole), the labels for the potholes were chosen

as follows: the most critical potholes were labeled as

Severe while medium sized and small potholes were

labeled as Pothole.

Deepbrokenhighways: Road Damage Recognition System Using Convolutional Neural Networks

741

3.2.2 Real-Time Pothole Severity and Crack

Detection Model

The second contribution of this work is the implemen-

tation of a model for real-time detection of pothole

severity. We will use the pre-trained Yolov4 model,

this model will be trained further using our data aug-

mented Peruvian Road Damage Dataset (Figure 4).

The architecture of the Yolov5 model is composed of

3 fundamental parts; the ﬁrst is the Backbone, where

Darknet or CSPDarknet53 is generally used to ex-

tract features in different dimensions from the input.

The second is the Neck, where FPN or SPP modules

are generally used to fuse the features extracted from

the backbone at different scales (sizes) and ﬁnally the

Head, which is responsible for performing detection

at multiple scales (sizes).

3.2.3 Road Damage Detection and Mapping

System

The third contribution of this work is a mobile appli-

cation Fig. 3, developed with the React Native frame-

work, with road damage detecting and mapping func-

tionalities. Using the model we developed, damage in

roads will be detected in real-time through a smart-

phone’s camera. As damage is detected the user will

be capable of taking a picture of the detection and

choose to register (upload) the damage to a Cloud

Firestore Database. The damage registration data in-

cludes: the picture, latitude and longitude of where

the picture was taken and the detected label. Once

damage is detected and registered, users will be able

to see damage detected throughout the city through a

map implemented into the application.

Figure 3: System Flow Diagram.

4 EXPERIMENTS

In this section, we will discuss the experiments that

were performed for our project. Detailing the con-

ﬁguration and resources needed to replicate said ex-

periments. Moreover, the results produced by these

experiments will also be discussed and interpreted.

4.1 Experimental Protocol

In this section, the details about the conﬁguration, re-

quired resources and applications used to perform the

experiments. The experiments were developed on the

Google Colaboratory Platform with a GPU Runtime.

This includes a Intel(R) Xeon(R) CPU @ 2.30GHz

The GPU we utilised for our experiments was a Tesla

T4, alongside 12 GB of RAM. Our code was devel-

oped on Python 3.9.16. Data augmentation related

work relies mainly on the Albumentations library.

For the implementation of the Yolo models, we

used the Ultralytics open-source project that fa-

cilitates the training of Yolo models. It can be

found at https://github.com/ultralytics/yolov5. Some

of the libraries required for these experiments in-

clude opencv-python>=4.1.1, Pillow>=7.1.2 and

torch>=1.7.0 as well as others detailed on the re-

quirements ﬁle of the ultralytics repository.

For the MobileNet v2 model, we’ll use the

TF2Lite Model Detection API. For this experiment

packages required for the Model Detection API as

well as the tensorﬂow=2.8.0 library are required.

The experiments consist of training the two mod-

els just mentioned with our Peruvian Road Dam-

age Dataset alongside images from widely used Road

Damage Dataset 2020 (RDD20). Moreover, to test

the viability of crack detection using a Bounding Box

method each model will be trained twice, with and

without including the crack damage annotations. The

Dataset used for these experiments contains 325 im-

ages from Peru and 293 from the RDD20 (Arya et al.,

2021), which includes images from India and Japan.

After Data Augmentation, which involves 3 different

types of transformations, the ﬁnal dataset for exper-

imenting has 4994 images. A example of the per-

formed transformations is shown in (Figure 4).

Figure 4: Data Augmentation performed on Dataset (Hueco

= Pothole; HuecoGrave = SeverePothole).

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

742

4.2 Results

In this subsection the experiments carried out and the

corresponding results obtained in each one will be de-

tailed and discussed.

4.2.1 YOLOv5

Training with YOLOv5 for the dataset without crack

annotations lasted about 4 hours with 21 minutes,

training a total of 350 epochs. On the other hand,

training with crack annotations lasted 4 hours with 31

minutes, training a total of 350 few. The training was

performed with the pre-trained Yolov5s weights and

with a batch size of 16 for both the datasets, with

crack annotations (see Figure 5a) and without crack

annotations (see Figure 5b).

Although good results were obtained, the train-

ing could not be more extensive due to the limited

resources used, being a maximum of 350 epochs as

previously mentioned. However, the loss obtained in

the last epochs of the training is minimal. This means

that the model is highly optimized and training be-

yond this point would not cause a drastic change in

results.

4.2.2 YOLOv8

Training with YOLOv8 for the dataset without crack

annotations lasted about 4 hours with 45 minutes,

training a total of 220 epochs. On the other hand,

training with crack annotations lasted 2 hours with 25

minutes, training a total of 110 epochs. The training

was performed with the pre-trained Yolov8n weights

and with a batch size of 16 for both the datasets, with

crack annotations and without crack annotations.

Due to the limited GPU resources the Google Co-

lab platform offers, it was not possible to perform a

more exhaustive training; however, when observing

Fig. 6a and Fig. 6b, it is evident that as more training

epochs are added, the loss will continue to decrease.

It is worth noting that the epochs in YOLOv8 train-

ing are fewer, this is due to YOLOv8 having higher

computational resource requirement than YOLOv5.

4.2.3 SSD MobileNet V2

The training for this model with the dataset with

both datasets lasted 35000 epochs for approximately

4 hours. When performing detection on test im-

ages the minimum conﬁdence was set at 0.15%, with

the purpose of ﬁltering very low quality, inaccurate

and redundant overlapping detections. Moreover,

the observed total loss during training had a con-

stant downward tendency, however the MobileNetV2

model trained with no crack annotations still showed

pronounced signs of a decrease in loss as shown

in Fig. 7a. That being said there’s a potential im-

provement to be reached if trained for even longer

(more epochs). Meanwhile the MobileNetV2 trained

with cracks annotations showed signs of stabilizing as

shown in Fig. 7b.

4.3 Discussion

In this subsection, the results obtained in the previous

section are detailed and discussed.

Models trained without crack annotations, mani-

fest excellent pothole detection capabilities as shown

in Table 1. The three models included in the experi-

ments show a pattern, which is a higher Average Pre-

cision for the Severe Pothole class than the Pothole

class. This is due to severe potholes, being a more

critical damage, it possess more notorious character-

istics and therefore is more distinguishable. On the

other hand, our ”regular” pothole class, which con-

tains smaller and less dangerous potholes, being a less

signiﬁcant damage in the road it involves a higher dif-

ﬁculty of detection. That being said, within the batch

of models trained without crack annotations, YoloV5

stands out as the best. As it has the top mean average

precision (mAP50), that being 99.0%.

On the side of the models trained with crack an-

notations, present an average precision over 90% in

all classes as shown in Tab. 1. The only exception

to this is the MobileNetV2, which shows a consid-

erable amount of drop off in average precision for

the pothole class in comparison to its counterpart

trained with no crack annotations. This is caused due

to having more classes to predict makes the model

more likely to confuse one for another. Especially

when the class objects are similar in appearance as

it is the case here. Another thing to consider is that

road cracks come in various shapes and types, which

also has an impact on the pothole detection. YoloV5

is again identiﬁed as the best model as it obtained

the best metrics in comparison to Yolov8 and Mo-

bileNetV2. Also, the deﬁciencies observed from the

MobileNetV2 model don’t apply. YoloV5 provides a

excellent average precision that is 99.3% for potholes

and severe potholes. While also having a 96.2% av-

erage precision for cracks. An acceptable drop off in

average precision, in comparison to the other classes,

considering how irregular and varied the appearance

of road cracks can be, making it the toughest class to

label and detect.

Lastly, through these experiments we have found

that crack detection via a Bounding Box method to be

viable. This being due to the competent average pre-

Deepbrokenhighways: Road Damage Recognition System Using Convolutional Neural Networks

743

(a) YOLOv5 training loss without crack annots.

(b) YOLOv5 training loss with crack annots.

Figure 5: Comparison of YOLOv5 training (a) without crack annots and (b) with crack annots.

Table 1: Av. Precision at 0.5 IoU.

without crack damage annots with crack damage annots

Model Pothole SeverePothole mAP50 Pothole SeverePothole Crack mAP50

SSD MobileNetV2 89.6% 95.3% 92.5% 83.2% 92.2% 90.8% 88.7%

YoloV5 98.6% 99.5% 99.0% 99.3% 99.3% 96.2% 98.3%

YoloV8 98.0% 99.2% 98.6% 94.2% 98.9% 94.4% 95.8%

cision obtained of 96.2% for cracks with the YoloV5

model, alongside 99.3% average precision for pothole

classes. Taking that into consideration, the model se-

lected for implementing into the mobile application

will be a YoloV5 model trained with crack annota-

tions. It presents the best performance from our set of

experiments, having a minimal decrement in mAP50

in comparison to the ”crack-less” YoloV5 model as it

involves a wider range of labels to detect.

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

744

(a) YOLOv8 training loss without crack annots.

(b) YOLOv8 training loss with crack annots.

Figure 6: Comparison of YOLOv8 training loss (a) without crack annots and (b) with crack annots.

(a) Without crack annots. (b) With crack annots.

Figure 7: Comparison of MobileNetV2 training total loss (a) without crack annots and (b) with crack annots.

5 CONCLUSIONS

In this paper, a system for detection and monitoring

of road damages is devised. The scope for this pa-

per includes the experimentation and selection of the

model which is most capable for road damage detec-

tion. Moreover, road crack detection has been proven

to be viable as the tested models got an average pre-

Deepbrokenhighways: Road Damage Recognition System Using Convolutional Neural Networks

745

cision above 90% for that speciﬁc label. That being

said we have found YOLOv5 to be the best model for

road damage detection in Peru. As it reports the best

mAP50 across both experiments 99.0% and 98.3%,

with and without cracks correspondingly. As we have

found crack detection to be viable a YoloV5 model

trained with crack annotations will be the one chosen

for our mobile application.

Due to the lack of data availability corresponding

to our context which is Lima, Peru; pushed us into

forming our own novel dataset. That being said, time

was also a constraint allowing us to elaborate our ex-

periments with 618 photos (325 from Peru and 293

from RDD20(Arya et al., 2021)), thus a data aug-

mentation phase was required in order to be able to

train a robust model with plenty of data. This pro-

cess allowed us to train our models with 2497 images,

in contrast to our original 618 images. Our results

demonstrated the effectiveness and usefulness of data

augmentation for road damage detection.

As future work, we would like to be able to use

drone or satellite imaging in order to further opti-

mize and speed up the road damage detection process.

While also testing its practicality as Peru’s roads are

highly transited, it poses the question if cars would be

a major obstacle to the visibility of potholes similarly

to other topics (Rodr

ıguez et al., 2021; Fernandez-

Ramos et al., 2021; Alfaro-Paredes et al., 2021).

REFERENCES

Alfaro-Paredes, E., Alfaro-Carrasco, L., and Ugarte, W.

(2021). Query by humming for song identiﬁcation us-

ing voice isolation. In IEA/AIE (2), volume 12799 of

Lecture Notes in Computer Science, pages 323–334.

Springer.

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

Akula, A. (2022). Convolutional neural networks

based potholes detection using thermal imaging. J.

King Saud Univ. Comput. Inf. Sci., 34(3):578–588.

Arya, D., Maeda, H., Ghosh, S. K., Toshniwal, D., and

Sekimoto, Y. (2021). Rdd2020: An annotated im-

age dataset for automatic road damage detection using

deep learning. Data in Brief, 36:107133.

Baek, J.-W. and Chung, K. (2020). Pothole classiﬁcation

model using edge detection in road image. Applied

Sciences, 10:6662.

Dong, J., Li, Z., Wang, Z., Wang, N., Guo, W., Ma, D.,

Hu, H., and Zhong, S. (2021). Pixel-level intelli-

gent segmentation and measurement method for pave-

ment multiple damages based on mobile deep learn-

ing. IEEE Access, 9:143860–143876.

Dong, J., Wang, N., Fang, H., Wu, R., Zheng, C., Ma, D.,

and Hu, H. (2022). Automatic damage segmentation

in pavement videos by fusing similar feature extrac-

tion siamese network (sfe-snet) and pavement damage

segmentation capsule network (pds-capsnet). Automa-

tion in Construction, 143:104537.

avila Estrada, H. A. (2022). Propuesta de un concreto para

pavimentos r

ıgidos con adici

on de polvo de vidrio

en reemplazo parcial del cemento y agregado ﬁno,

ın de reducir la contaminaci

on producida por la

construcci

on de la capa de rodadura en la carretera

mayocc-huanta, tramo allccomachay-huanta departa-

mento de ayacucho.

Egaji, O. A., Evans, G., Grifﬁths, M. G., and Islas, G.

(2021). Real-time machine learning-based approach

for pothole detection. Expert Syst. Appl., 184:115562.

Fernandez-Ramos, O., Johnson-Ya

nez, D., and Ugarte, W.

(2021). Reproducing arm movements based on pose

estimation with robot programming by demonstration.

In ICTAI, pages 294–298. IEEE.

Kim, Y.-M., Kim, Y.-G., Son, S.-Y., Lim, S.-Y., Choi, B.-Y.,

and Choi, D.-H. (2022). Review of recent automated

pothole-detection methods. Applied Sciences, 12(11).

Ministerio de Transportes y Comunicaciones (2015). Re-

sumen ejecutivo del inventario basico de la red vial

departamental o regional.

Moscoso Thompson, E., Ranieri, A., Biasotti, S., Chic-

chon, M., Sipiran, I., Pham, M.-K., Nguyen-Ho, T.-L.,

Nguyen, H.-D., and Tran, M.-T. (2022). Shrec 2022:

Pothole and crack detection in the road pavement us-

ing images and rgb-d data. Computers & Graphics,

107:161–171.

Pandey, A. K., Iqbal, R., Maniak, T., Karyotis, C., Akuma,

S., and Palade, V. (2022). Convolution neural net-

works for pothole detection of critical road infras-

tructure. Computers and Electrical Engineering,

99:107725.

Park, S.-S., Tran, V.-T., and Lee, D.-E. (2021). Applica-

tion of various yolo models for computer vision-based

real-time pothole detection. Applied Sciences, 11(23).

Patra, S., Middya, A. I., and Roy, S. (2021). Potspot: Partic-

ipatory sensing based monitoring system for pothole

detection using deep learning. Multim. Tools Appl.,

80(16):25171–25195.

Rateke, T. and von Wangenheim, A. (2021). Road surface

detection and differentiation considering surface dam-

ages. Auton. Robots, 45(2):299–312.

Rodr

ıguez, M., Pastor, F., and Ugarte, W. (2021). Classi-

ﬁcation of fruit ripeness grades using a convolutional

neural network and data augmentation. In FRUCT,

pages 374–380. IEEE.

Tenorio Construcciones y Soluciones (2022). Manten-

imiento en v

ıas.

Wu, C., Wang, Z., Hu, S., Lepine, J., Na, X., Ainalis,

D., and Stettler, M. (2020). An automated machine-

learning approach for road pothole detection using

smartphone sensor data. Sensors, 20(19).

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

746