Federated Road Surface Anomaly Detection Using Smartphone

Accelerometer Data

Oussama Mazari Abdessameud

and Walid Cheriﬁ

Dept. Computer Science, Ecole Militaire Polytechnique, Algiers, Algeria

Keywords:

Road Anomaly Detection, Smartphone, GPS Turn-by-Turn Navigation, Federated Learning.

Abstract:

Road surface conditions signiﬁcantly impact trafﬁc ﬂow, vehicle integrity, and driver safety. This importance

is magniﬁed in the context of service vehicles, where speed is often the only recourse for saving lives. Detect-

ing road surface anomalies, such as potholes, cracks, and speed bumps, is crucial for ensuring smooth and safe

driving experiences. Taking advantage of the widespread use of smartphones, this paper introduces a turn-

by-turn navigation system that utilizes machine learning to detect road surface anomalies using accelerometer

data and promptly alerts drivers. The detection model is personalized for individual drivers and continuously

enhanced through federated learning, ensuring both local and global model improvements without compromis-

ing user privacy. Experimental results showcase the detection performance of our model, which continually

improves with cumulative user contributions.

1 INTRODUCTION

Roads are fundamental to transportation networks, fa-

cilitating seamless connectivity and mobility. A ro-

bust road network enables efﬁcient travel for individ-

uals and the transportation of goods. However, vari-

ous natural and human-induced factors can contribute

to road surface damage, resulting in the formation

of potholes, cracks, and other irregularities. While

potholes and cracks typically arise from natural wear

and tear, speed bumps represent a deliberate interven-

tion by human actors to achieve speciﬁc trafﬁc man-

agement objectives. Nonetheless, poorly designed or

maintained speed bumps can also pose safety hazards,

increasing the risk of accidents and injuries, particu-

larly when drivers are unaware of their presence or

fail to perceive them.

Road anomalies, including defects and speed

bumps, pose signiﬁcant risks to both human occu-

pants and vehicular components. Indeed, encounter-

ing road surface anomalies at high speeds can lead

to vehicle damage, malfunctioning of vehicle com-

ponents, or even injury to occupants (Kosakowska,

2022). Proactive measures are essential for mitigating

these risks. Drivers can take precautions by reducing

their speed before traversing a speed bump or encoun-

tering surface irregularities. However, the challenge

https://orcid.org/0000-0001-7198-8974

becomes more pronounced when drivers are unaware

of upcoming anomalies on their route. In such scenar-

ios, the need for an automatic solution becomes im-

perative. By leveraging technology, such as advanced

sensors and real-time monitoring systems, drivers can

be automatically notiﬁed of impending anomalies, en-

abling them to take necessary precautions and adapt

their driving behavior accordingly. An automatic so-

lution enhances driver awareness and safety and con-

tributes to the overall efﬁciency and maintenance of

road infrastructure.

Smartphones with GPS and navigation systems

have transformed driving habits, allowing for efﬁ-

cient route navigation. Integrating automatic road

anomaly notiﬁcations seamlessly complements these

habits, providing timely alerts during journeys. Es-

tablishing a comprehensive database of road surface

anomalies is essential, facilitated by crowd-sourcing

from drivers via automated marking. Smartphones’

accelerometer sensors capture signals reﬂecting the

vehicle’s acceleration, enabling precise identiﬁcation

and location of road surface anomalies when coupled

with GPS data.

Classifying the footprint left in the acceleration

signal by road anomalies necessitates a pre-trained

classiﬁcation model. Given potential variations in

signals generated by different phones and vehicles,

the model must be trained across various scenarios.

While data from different users could be used to re-

376

Mazari Abdessameud, O. and Cheriﬁ, W.

Federated Road Surface Anomaly Detection Using Smartphone Accelerometer Data.

DOI: 10.5220/0012766200003756

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

In Proceedings of the 13th International Conference on Data Science, Technology and Applications (DATA 2024), pages 376-383

ISBN: 978-989-758-707-8; ISSN: 2184-285X

train the model, concerns about user privacy arise due

to the revealing nature of acceleration data, which

can disclose driving habits. Consequently, training

the classiﬁcation model locally on users’ devices and

aggregating these models into a single global model

helps preserve user privacy.

This paper presents a proactive solution to en-

hance road safety by utilization of smartphone tech-

nology for real-time detection and notiﬁcation of road

surface anomalies. It introduces a turn-by-turn nav-

igation system that exploits the widespread use of

smartphones equipped with GPS, navigation systems,

and accelerometer sensors. The system’s primary ob-

jective is to detect road surface anomalies in real-time

and provide prompt alerts to drivers to enhance road

safety. Through machine learning techniques, the sys-

tem automatically detects and classiﬁes road surface

anomalies based on the footprints observed in the ac-

celerometer data. Upon detecting an anomaly, the

system marks its location to alert approaching drivers.

Furthermore, to enhance detection performance, the

machine learning model undergoes additional train-

ing using user data. To address privacy concerns as-

sociated with user data, the system employs a feder-

ated learning approach. Local models are trained on

users’ devices, and the aggregated global model is uti-

lized for anomaly detection while safeguarding user

privacy.

2 RELATED WORKS

Al-Sabaeei et al. (2024) present a comprehensive

review of smartphone applications for road surface

monitoring. These applications predominantly rely

on two distinct approaches: vision-based data col-

lection and the analysis of acceleration and vibration

data. By exploiting these techniques, they aim to as-

sess the condition of road pavements.

Camera-based methods have shown promise in

detecting road surface defects and anomalies, with

reported accuracy levels ranging from 80% to 98%.

However, their effectiveness is inﬂuenced by fac-

tors like lighting conditions and image quality (Rahi-

man V et al., 2021; Lee et al., 2021; Kim and Kim,

2023). Despite their potential, vision-based methods

have limitations. These methods require stable ori-

entation of the smartphone camera towards the road

surface, which conﬂicts with our system’s objective

of providing real-time notiﬁcations and navigation to

users. As our system prioritizes positioning the smart-

phone camera facing the driver, this approach imposes

constraints on phone placement, rendering it unsuit-

able for our application.

On the other hand, the use of vibrations data also

depicts promising results. Chen et al. (2022) de-

veloped a convolutional neural network that based

on smartphones’ accelerometer data detects road sur-

face transverse cracks with a 97% accuracy. Mar-

tinelli et al. (2022) exploited the same accelerom-

eter data with with several machine learning tech-

niques including decision trees, support vector ma-

chine to detect with an accuracy ranging from 84%

to 97% in determining whether the road pavement

is distressed or not. Several similar works showed

that vibrations data are a reliable source for road sur-

face anomalies detection (Dong and Li, 2021; Shole-

var et al., 2022; Mazari Abdessameud et al., 2022;

Tomiło, 2023; Yuan et al., 2023). Therefore, the fu-

sion of vibration data and advanced machine learning

techniques holds great promise for improving road

safety.

Crowd-sourcing offers a valuable avenue for col-

lecting precise road anomaly locations. In their study,

Xin et al. (2023) propose a novel probabilistic-based

crowdsourcing technique. By aggregating data from

a large number of users, the accuracy of anomaly

detection and localisation can be signiﬁcantly im-

proved. This approach effectively ﬁlters out poten-

tial false positives and false negatives by combining

dynamic events detected from various smartphones

onboard vehicles. Also, by harnessing this crowd-

sourced data, errors associated with GPS for anomaly

localisation is mitigated, leading to more reliable re-

sults.

In their recent work, Jeong and Jo (2024) in-

troduce a machine learning approach that leverages

crowdsourced data to evaluate pavement conditions.

Their methodology aims to develop classiﬁers by in-

corporating diverse data sources. However, a no-

table limitation arises: these classiﬁers are still con-

strained by the number of instances used during train-

ing since the classiﬁcation model does not evolve. In-

terestingly, Jan et al. (2023) echo a similar concern

in their research. They emphasize that challenges in

pavement condition detection stem from differences

in sensing platforms and vehicle parameters. These

variations introduce complexities that impact the ac-

curacy and adaptability of detection models.

To address the limitations of static training of

classiﬁcation models, the proposed solution involves

employing federated learning techniques. Federated

learning allows the model to adapt dynamically to dif-

ferent situations while preserving user data privacy.

By combining this approach to detection base on ac-

celerometer data and crowdsourcing for localisation,

we can create more effective and adaptable systems

for road safety.

Federated Road Surface Anomaly Detection Using Smartphone Accelerometer Data

377

3 METHODOLOGY

Driving over road surface anomalies without prior no-

tice can jeopardize driver safety and vehicle integrity.

Therefore, drivers should be aware of road anoma-

lies along their routes. In this section, we elaborate

on how our system utilizes drivers data to detect the

presence of road surface anomalies. These detected

anomalies are stored in a central database and utilized

to alert subsequent drivers.

To streamline the use of our system, we en-

gage with end users through smartphones. These

devices offer users the ability to navigate their pre-

ferred routes while receiving alerts about upcoming

road surface anomalies along the way. Also, the ac-

celerometers embedded in these smartphones enable

automatic detection of new road surface anomalies,

enriching our database with real-time information.

To establish a robust and accurate database of

road surface anomalies, the system relies on user-

generated annotations. Users report encountered

anomalies, such as potholes or cracks, detected by the

integrated model. Each report includes details such

as the anomaly type, precise location using GPS co-

ordinates, and a conﬁdence rate indicating the level

of conﬁdence in the model’s detection, which can

be enhanced through continuous training. The sys-

tem consolidates conﬁrmed anomalies by aggregating

reports for the same location, considering the conﬁ-

dence scores associated with individual reports to de-

termine the anomaly type.

This methodology ensures that anomalies fre-

quently reported with high conﬁdence levels are clas-

siﬁed as conﬁrmed, thereby enhancing the system’s

accuracy and reliability in identifying critical road

surface issues.

3.1 General Architecture

The proposed system comprises eight interconnected

modules designed to collect, process, and ana-

lyze data from various sources, ultimately provid-

ing timely notiﬁcations to drivers regarding upcoming

road anomalies. The general architecture of the pro-

posed system is illustrated in ﬁgure 1. These modules

are distributed over the client side and the server side.

On the client side, we deploy ﬁve modules to manage:

• Navigation: The system retrieves essential infor-

mation from a remote geographic database to cal-

culate routes for the user, facilitating navigation

along these routes with step-by-step directions.

Furthermore, this module provides notiﬁcations

regarding upcoming road surface anomalies to en-

sure user awareness and safety during the journey.

• Upcoming conﬁrmed anomalies: Upon receiving

the user-selected route, the navigation module in-

terfaces with this module to obtain information

regarding conﬁrmed anomalies along the chosen

route, including their locations. These anoma-

lies are subsequently utilized to notify the user

and evaluate the effectiveness of the automatic

anomaly detection module.

• Anomaly detection: To detect whether the vehicle

hit a road surface anomaly, the module receive ac-

celerometer signals. These signals are then sub-

jected to various operations, including reorienta-

tion from relative to absolute reference, and data

ﬁltering to eliminate artifacts and unwanted noise.

The detection module retrieves one signal window

at a time and performs an inference on the data

(acceleration x, acceleration y, acceleration z and

velocity) using the detection model management

module. The result is a prediction vector indicat-

ing the class of the signal present in this window.

• Detection results veriﬁcation: This involves the

interface of this module with the Upcoming

Conﬁrmed Anomalies module, utilizing smart-

phone location data to compare the outcomes of

the anomaly detection module with conﬁrmed

anomalies in the same location. Four scenarios

may arise: i. When the same type of anomaly is

detected, if the detection rate is low, accelerome-

ter data is utilized to augment the training of the

local model, thereby enhancing accuracy. ii. If

the model fails to detect a conﬁrmed anomaly or

iii. identiﬁes an anomaly with an incorrect type,

accelerometer data is employed to reﬁne the lo-

cal model due to the model’s likely inaccuracy. iv.

In the event the model detects an anomaly not yet

conﬁrmed, a ﬂag is dispatched to the distant server

to report the anomaly’s existence. In this case,

accelerometer data awaits conﬁrmation of the de-

tected anomaly’s type.

• Local detection model management: This module

manages the local detection module. It conducts

local model training using data collected from the

device. Additionally, the module facilitates com-

munication with the server to either contribute the

local detection model to the global model or up-

date the local model based on the received global

model.

These client modules interact with a server that man-

ages centralised actions and data. The server relies on

three modules to deal with:

• Road Anomalies Flags Storage: This module

maintains a record of the various ﬂags received,

as well as the conﬁrmed road anomalies. Each re-

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

378

Server Side

Driver (user) Side

3D accelerations

data

Location

Registered

road anomalies

Routes for

navigation

Route and anomalies

notification

Detection results

Accelerations data

Detection

results

verification

anomaly

detection

Confirmed

anomaly not

detected

Anomalies

data fusion

module

Local detection

model management

Local model

training

Local model

update

Updated detection

model

Global

detection

model

management

Clients

models

management

Models

aggregation /

evaluation

Up coming confirmed

anomalies

Navigation module

Map and

anomalies

visualisation

Navigation

and

anomalies

warning

Routes

calculation

Anomalies detection module

classification

Signal

acquisition

and filtering

Figure 1: General architecture of the proposed system.

ceived ﬂag includes details such as anomaly type,

location, and ﬂag reliability or conﬁdence level.

Conﬁrmed road anomalies are those ﬂags with

high reliability or conﬁdence levels.

• Anomalies data fusion: The aim of this module

is to achieve a consistent state of the anomaly

database. This data fusion process enables the

ﬂags received from users to be combined, pro-

viding a more complete and accurate informa-

tion. Due to potential inaccuracies, ﬂags from

various sources might describe the same anomaly

with different types and slightly varying locations.

To address this, the Anomalies data fusion mod-

ule employs a Weighted Majority Voting (WMV)

technique (Tao et al., 2018) to consolidate ﬂags

for each anomaly and select the most reliable type.

This process involves three key steps: data clus-

tering, data fusion, and decision making. Sub-

section 3.3 describes the data fusion process ﬂow.

This process effectively combines ﬂags from vari-

ous contributors, resulting in a comprehensive and

reliable record of anomalies.

• Global detection model management: This mod-

ule ensures the update of the global detection

model, and the update of the different user’s lo-

cal models. The update of the global detection

model involves the aggregation of several local

users models. This process involves retrieving

weights from saved different users and combining

them using a predeﬁned algorithm (e.g., FedAvg).

Finally, the resulting aggregated model undergoes

evaluation to assess its performance.

3.2 Anomaly Detection

The proposed system provides an efﬁcient automatic

road anomaly detection service, leveraging smart-

phone capabilities to identify irregularities in road

surfaces. It is a well-known phenomenon for vehicles

to experience vibrations when encountering anoma-

lies such as potholes. These vibrations are captured

as a three-axis acceleration signal generated by the

smartphone’s integrated sensor. However, this sig-

nal is collected in the smartphone’s local coordinate

system, which depends on how the phone is mounted

in the vehicle. To overcome this challenge, the pro-

posed system employs an automatic correction ser-

vice that utilizes Euler angles to reorient the accel-

eration data. This approach eliminates the depen-

dence on the smartphone’s physical orientation, en-

suring consistent data interpretation regardless of how

the user mounts the phone. The reoriented signal can

still contain noise from non-informative vibrations,

such as constant engine hum. To address this, we

employ a high-pass ﬁlter. This ﬁlter removes low-

frequency data, effectively resulting in a ﬁltered vi-

bration signal that’s more suitable for further process-

Federated Road Surface Anomaly Detection Using Smartphone Accelerometer Data

379

ing and anomaly detection. After preprocessing the

vibration signal, it is divided into equal-length sliding

windows to produce a series of consecutive samples.

Given the Long Short-Term Memory (LSTM)

neural networks’ demonstrated capability to automat-

ically extract pertinent features from raw time series

data and handle temporal dependencies, we advocate

for a stacked LSTM architecture comprising two hid-

den LSTM layers, each containing multiple memory

cells. This architecture is used to classify whether the

window of the captured signal represents a smooth

road or a road surface anomaly.

3.3 Anomalies Fusion

As the system relies on user input to detect road sur-

face anomalies, multiple ﬂags may be received for

the same anomaly, each varying in reliability and po-

tentially indicating different anomaly types. Conse-

quently, when multiple ﬂags are accumulated for the

same approximate location, a fusion process is ini-

tiated. As shown in ﬁgure 2, this fusion process

incorporates three pieces of information: location,

anomaly type, and conﬁdence rating.

Flag n

Location n Anomaly type n Confidence rate n

Confirmed anomaly flag

Location Anomaly type Confidence rate 3

Flag 1

Location 1 Anomaly type 1 Confidence rate 1

Flag 2

Location 2 Anomaly type 2 Confidence rate 2

...

Average if ...

Weighted Majority

Voting

Sum if ...

YES

>= Thershold

Figure 2: Schema of the fusion process of Road surface

anomaly ﬂags.

To determine the retained anomaly type, a WMV

approach is employed, with the conﬁdence rating of

each ﬂag used to weigh the different types. Once the

elected type is identiﬁed, if the sum of conﬁdences of

the ﬂags holding this type exceeds the deﬁned thresh-

old, a conﬁrmed road surface anomaly is generated.

However, if the threshold is not met, no conclusion is

drawn. In terms of location, the ﬁnal location is deter-

mined as the average of all locations associated with

ﬂags indicating the elected type.

3.4 Federated Detection Enhancement

The federated enhancement of the detection model re-

lies on the local and global detection model manage-

ment modules. These two modules interact with each

other to exchange model parameters. The local de-

tection model management module manages the lo-

cal model training using captured signals that were

wrongly classiﬁed. Global model management per-

forms the aggregation of the different models received

from users.

Local training is a crucial functionality designed

to reﬁne the local model by utilizing data stored in

the local database. This feature is activated when the

system accumulates a substantial number of misclas-

siﬁed signals. Upon retrieving the signal instances,

the local training module proceeds to acquire label

updates, as anomalies may have been removed or

changed type during the data fusion process. Subse-

quently, the training process is initiated, adhering to

pre-deﬁned parameters such as the number of epochs.

Upon completion, the new weights of the resulting

model are applied locally for subsequent detection,

along with essential metadata such as the user iden-

tiﬁer, dataset size used in training, and loss func-

tion. This iterative approach ensures the local model

is continuously updated and enhanced, leveraging re-

cent data to improve accuracy and anomaly detection

performance.

On the server side, a scheduled aggregation pro-

cess is implemented. The aggregation process com-

prises several steps. Firstly, the weights of models

trained locally by customers are retrieved. Next, an

aggregation operation is performed to combine these

weights using a FedAvg algorithm as described in al-

gorithm 1. The aggregation algorithm begins by cal-

culating the aggregate count of instances utilized in

training all local models for the current aggregation

cycle. Subsequently, each local model contributes

to the new model in proportion to the number of in-

stances employed for its individual training.

Once aggregation is complete, the resulting model

is evaluated to measure its performance. This involves

the use of a test and validation dataset to assess the

accuracy and quality of the aggregated model. De-

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

380

pending on the results of the evaluation, a decision is

made regarding the deployment of the model. If the

aggregated model is deemed satisfactory, it is retained

as the new global detection model and made available

to users. This enables users to beneﬁt from the im-

proved detection performance offered by the aggre-

gated model.

Data: Current global detection model, List of

users detection models with their

metadata

Result: Updated global detection model

Total Data Size ←

∑

Train Data Size

foreach local Model ∈ local Models do

size ← Train Dataset of local Model

Rate ← size / Total Data Size

Names, Weights ← Get local Model

parameters names and weights

foreach name ∈ Names do

New Global Model[name] ←

New Global Model[name] + (Rate

× Weights[name])

end

Save (New Global Model)

Algorithm 1: Local models aggregation algorithm.

4 EXPERIMENTAL SETUP AND

PERFORMANCE EVALUATION

The performance evaluation of the proposed system

comprises two main components. Firstly, the evalu-

ation of the initial detection model that was trained

using a public dataset. The second part consists of

performance evaluation of the same model after a fed-

erated learning phase.

4.1 Training Dataset and Evaluation

Metrics

We utilized the ”Pothole lab” dataset (Lab, 2016),

which comprises over 2500 instances of various road

surface anomalies, including speed bumps, metal

bumps, and potholes, alongside over 1500 instances

of smooth roads. This dataset offers a diverse range of

anomalies commonly encountered on roadways, en-

abling comprehensive evaluation of our detection sys-

tem’s performance. Each instance in the dataset pro-

vides acceleration data along three axes, facilitating

a detailed analysis of vehicle motion and response to

road surface conditions. To assess the effectiveness of

our system, we employed standard evaluation metrics,

including precision, recall, and F1-score. These met-

rics provide valuable insights into the system’s ability

to accurately detect and classify road surface anoma-

lies while minimizing false positives and negatives.

4.2 Initial Classiﬁcation Model

Evaluation

To evaluate the effectiveness of the proposed clas-

siﬁcation model, we explored various combinations

of training hyperparameters. These parameters en-

compassed the number of iterations on the training

dataset, the window size representing an anomaly, the

overlap between windows, and the number of mem-

ory cells in each LSTM layer. The evolution of accu-

racy and loss function during training for the combi-

nation yielding the best results is depicted in Figure

Figure 3: Evolution of the accuracy and the loss function

during training.

The evolution of accuracy and loss show that the

model did not over or under learn. The obtained re-

sults for the evaluation parameters are shown in table

Table 1: Achieved parameters values for the evaluation of

the classiﬁcation model based on Pothole Lab dataset.

Precision Recall F1-score

0.96 0.95 0.95

4.3 Federated Classiﬁcation Model

Evaluation

To evaluate the performance of the federated learning

process, we conducted testing of the detection model

in two distinct phases. In the ﬁrst phase, we utilized

the pre-trained model to assess the initial detection

capabilities across multiple devices. This allowed

us to gauge the baseline performance of the model

before any federated training. Following this initial

phase, we proceeded to the second phase, where the

detection model underwent federated training. Dur-

ing this process, data from multiple devices were ag-

Federated Road Surface Anomaly Detection Using Smartphone Accelerometer Data

381

gregated and used to reﬁne the model’s parameters,

aiming to enhance its ability to generalize across di-

verse datasets while preserving user privacy. Subse-

quently, we conducted testing again to evaluate the

performance of the model after federated training. By

comparing the results of these two phases, we could

assess the effectiveness of the federated learning ap-

proach in improving the detection model’s accuracy

and robustness across different devices and datasets.

Figure 4 shows the followed process to achive the two

phases evaluation.

Classification result

Initial Model

Yes

Correct classification

Registered road

anomalies

Federated training

Before Federated training

evaluation (first phase)

Classification result

Yes

Correct classification

Registered road

anomalies

After Federated training

evaluation (Second phase)

Figure 4: Schema of the fusion process of Road surface

anomaly ﬂags.

We deployed the detection model across four dif-

ferent devices to evaluate its performance in a real-

world setting. Each device was equipped with the

trained model and tasked with detecting road sur-

face anomalies during typical driving scenarios. The

anomaly database of the system was populated with

existing anomalies known to occur on the test road.

As the vehicles traversed various road conditions,

including smooth surfaces and areas with known

anomalies such as potholes and speed bumps, the

accelerometer sensors in the smartphones captured

data reﬂecting the vehicle’s acceleration along three

axes. This data was then processed by the detection

model to identify and classify road surface anomalies.

During detection, the system compared the detected

anomalies with those stored in the database. Any

discrepancies or inaccuracies in classiﬁcation were

noted, and anomalies that were wrongly classiﬁed

were earmarked for further training of the model. Fol-

lowing the test, the results were analyzed to assess the

precision and recall of the detection model under real-

world conditions. The obtained results for the two

phases are detailed in table 2

Table 2: Achieved parameters values for the evaluation the

classiﬁcation model before and after the federated training

(FT) process.

Phase Precision Recall F1-score

Before FT 0.77 0.73 0.73

After FT 0.84 0.79 0.79

5 DISCUSSION

The evaluation of our system comprised two dis-

tinct phases. Initially, we conducted an assessment

of the classiﬁcation model’s performance before im-

plementing federated learning. This evaluation, con-

ducted solely on the trained dataset, yielded promis-

ing results, with a precision of 96% and recall of 95%.

These high scores indicate the reliability and effec-

tiveness of the model in accurately detecting and clas-

sifying road surface anomalies.

Subsequently, we proceeded with the evaluation

of the federated learning component. Upon deploy-

ment of the model across four different devices, an

initial test revealed a slight decline in performance,

resulting in a precision of 77% and recall of 73%.

This drop in performance can be attributed to varia-

tions in the data used for training and testing, high-

lighting the challenge of ensuring consistency across

diverse datasets.

To address this issue, we initiated a federated

training process, aggregating results from multiple

devices and relaunching the test. The subsequent

evaluation demonstrated a notable improvement in

performance, with a precision of 85% and recall of

79%. These enhanced results underscore the effec-

tiveness of the federated learning approach in reﬁning

and optimizing the detection model.

Overall, the ﬁndings indicate that while the ini-

tial performance of the model was commendable, the

incorporation of federated learning signiﬁcantly en-

hanced its performance, underscoring the importance

of collaborative and distributed learning methods in

improving the accuracy and robustness of detection

systems.

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

382

6 CONCLUSION AND

PERSPECTIVES

This paper presents a proactive solution leveraging

smartphone technology for real-time detection and

notiﬁcation of road surface anomalies. Through the

integration of machine learning techniques and ac-

celerometer data, our turn-by-turn navigation sys-

tem effectively identiﬁes and alerts drivers to po-

tential road surface anomalies, thereby enhancing

overall road safety. The evaluation of our system

demonstrated promising results, with the classiﬁca-

tion model exhibiting high precision and recall rates

in detecting anomalies. Furthermore, the implemen-

tation of federated learning proved instrumental in re-

ﬁning the detection model’s performance across di-

verse situations, highlighting the efﬁcacy of collabo-

rative learning approaches in improving detection ac-

curacy while preserving user privacy. Overall, our

system offers a practical and effective approach to ad-

dressing road safety concerns, with the potential to

signiﬁcantly reduce the incidence of accidents and

improve the overall driving experience. As future

work, further optimization and reﬁnement of the de-

tection model could be explored, along with the in-

tegration of additional sensors or data sources to en-

hance anomaly detection capabilities in various road

conditions. The most promising additional sensor is

the camera technology that can augment anomaly de-

tection capabilities. Cameras can capture visual in-

formation about road surface conditions, allowing for

the detection of anomalies before vehicles encounter

them.

DISCLOSURE OF AI TOOLS

USAGE

The preparation of this manuscript involved the use

of Copilote and ChatGPT to correct and improve the

language through the manuscript. Subsequently, the

authors reviewed and edited the content as necessary,

and take full responsibility for the paper’s content.

REFERENCES

Al-Sabaeei, A. M., Souliman, M. I., and Jagadeesh, A.

(2024). Smartphone applications for pavement condi-

tion monitoring: A review. Construction and Building

Materials, 410:134207.

Chen, C., Seo, H., and Zhao, Y. (2022). A novel pavement

transverse cracks detection model using wt-cnn and

stft-cnn for smartphone data analysis. International

Journal of Pavement Engineering, 23(12):4372–4384.

Dong, D. and Li, Z. (2021). Smartphone sensing of

road surface condition and defect detection. Sensors,

21(16).

Jan, M., Khattak, K. S., Khan, Z. H., Gulliver, T. A., and

Altamimi, A. B. (2023). Crowdsensing for road pave-

ment condition monitoring: Trends, limitations, and

opportunities. IEEE Access, 11:133143–133159.

Jeong, J.-H. and Jo, H. (2024). Toward real-world

implementation of deep learning for smartphone-

crowdsourced pavement condition assessment. IEEE

Internet of Things Journal, 11(4):6328–6337.

Kim, M. J. and Kim, Y. M. (2023). Rss model improvement

considering road conditions for the application of a

variable focus function camera. Sensors, 23(2).

Kosakowska, K. (2022). Evaluation of the impact of speed

bumps on the safety of residents-selected aspects.

Transportation research procedia, 60:418–423.

Lab, P. (2016). Roadway surface dis-

ruptions dataset, [online] available:

https://www.accelerometer.xyz/pothole lab/.

Lee, T., Chun, C., and Ryu, S.-K. (2021). Detection of road-

surface anomalies using a smartphone camera and ac-

celerometer. Sensors, 21(2).

Martinelli, A., Meocci, M., Dolﬁ, M., Branzi, V., Morosi,

S., Argenti, F., Berzi, L., and Consumi, T. (2022).

Road surface anomaly assessment using low-cost ac-

celerometers: A machine learning approach. Sensors,

22(10).

Mazari Abdessameud, O., Cheriﬁ, W., Kribi, M. A. E. I.,

and Dahmani, A. (2022). Navisaf: A safe navigation

system for road anomalies detection. In IECON 2022–

48th Annual Conference of the IEEE Industrial Elec-

tronics Society, pages 1–6. IEEE.

Rahiman V, A. et al. (2021). Pothole detection and volume

estimation based on disparity transformation with his-

togram thresholding. Proceedings of the Yukthi 2021-

The International Conference on Emerging Trends in

Engineering.

Sholevar, N., Golroo, A., and Esfahani, S. R. (2022).

Machine learning techniques for pavement condition

evaluation. Automation in Construction, 136:104190.

Tao, D., Cheng, J., Yu, Z., Yue, K., and Wang, L. (2018).

Domain-weighted majority voting for crowdsourcing.

IEEE transactions on neural networks and learning

systems, 30(1):163–174.

Tomiło, P. (2023). Classiﬁcation of the condition of

pavement with the use of machine learning meth-

ods. Transport and Telecommunication Journal,

24(2):158–166.

Xin, H., Ye, Y., Na, X., Hu, H., Wang, G., Wu, C., and

Hu, S. (2023). Sustainable road pothole detection: A

crowdsourcing based multi-sensors fusion approach.

Sustainability, 15(8).

Yuan, W., Yang, Q., and Zhang, R. (2023). Asphalt pave-

ment surface repair areas detection based on smart-

phone sensors. International Journal of Transporta-

tion Science and Technology.

Federated Road Surface Anomaly Detection Using Smartphone Accelerometer Data

383