Machine Learning Support for Time-Efﬁcient Processing Dangerous

Driving Detection Using Vehicle Inertial Data

Matheus Jo

ao Silva de Almeida, Julia Kerkoff Ladeira, Caio Gabriel Vicentin,

Andre V. Carvalho Costa, Marcia Pasin

and Vin

ıcius K. Marini

Centro de Tecnologia, Universidade Federal de Santa Maria, Brazil

Keywords:

Efﬁcient Processing, Dangerous Driving, Artiﬁcial Intelligence, Machine Learning Techniques.

Abstract:

Detection of dangerous driving behavior is a key component to improving road safety. It can be successfully

carried out using data collected by sensors widely available in smartphones. Current work focuses on two

groups: either they classify drivers in a binary way, into good and bad drivers, or they provide a scoring

scale, allowing for a larger group of categories. This detection of dangerous driving behavior can be done

with high granularity, evaluating a total distance covered by the driver on a trip, or with minute granularity,

through the evaluation of small sections of driving, also making it possible to identify which maneuvers the

driver is carrying out negligently. However, the process of collecting data for dangerous driving behavior is

complicated because the driver needs to carry out these maneuvers, so that a classiﬁer can later detect them,

adding to situations of insecurity in trafﬁc. Moreover, the solution needs to execute efﬁciently, so that the

detection of dangerous driving behavior can be carried out in real time. Given this problem, we propose a

time efﬁcient dangerous driving detection system using vehicle inertial data. In contrast to other works, we

collected data in a simulation environment with a model car that allows us to perform risky maneuvers, which

would not be possible in a real environment. We identify in our small dataset the dangerous driving behavior

pattern. Thus, given the established pattern, we applied a machine learning method to generate a classiﬁer to

enable the detection of dangerous driving behavior. The resulting system achieved a total average accuracy of

85.61% in our experiments using a small dataset as input towards efﬁcient data processing.

1 INTRODUCTION

The use of motorized vehicles is augmenting more

and more in our country (see Fig. 1). In fact, mo-

tor vehicles are key elements in urban mobility. In

this context, even though there is a few tendency to

improve over the last few years, the number of traf-

ﬁc accidents is still extremely relevant. Due to the

driver behavior, lack of vehicle maintenance, bad road

and bad environment conditions, accidents occur fre-

quently, e and this impacts negatively people’s quality

of life and also on the economy of countries. The

main causes of motor vehicle accidents are human

factors, such as drowsiness while driving or driver

distraction.

Given the human factor, detection of dangerous

driving behavior is mandatory to improve road safety.

It can be successfully carried out using data collected

by sensors widely available in smartphones, such as

accelerometer (ACC) and gyroscope (GYRO). These

https://orcid.org/0000-0001-6649-1488

100

150

200

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

2018

2019

2020

2021

2022

Brazillian Country Fleet (in millions)

Years

cars

trucks

motorcycle

bus

deaths

accidents

Figure 1: History data of the number of vehicles in circu-

lation

, the number of deaths due to trafﬁc accidents

, and

the number of accidents on public highways in Brazil

Figure 1: History data of the number of vehicles in

circulation

, the number of deaths due to trafﬁc accidents

and the number of accidents on public highways in Brazil

sensors can also be easily embedded in vehicles to-

gether with other sensors such as those to detect driver

https://cidades.ibge.gov.br

https://datasus.saude.gov.br

https://cnt.org.br/painel-acidente

Silva de Almeida, M., Ladeira, J., Vicentin, C., Costa, A., Pasin, M. and Marini, V.

Machine Learning Support for Time-Efﬁcient Processing Dangerous Driving Detection Using Vehicle Inertial Data.

DOI: 10.5220/0012686200003690

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 997-1004

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

997

alcohol ingestion (Abu Al-Haija and Krichen, 2022)

(Willis et al., 2019) (Shreshtha et al., 2020), or cam-

eras to capture the driver behaviour inside the vehicle

or even outside the vehicle, which can be used to de-

tect the use of smartphones (Wang et al., 2014) (Yasar,

2017), for example, or any other type of bad driver be-

havior during trips.

Moreover, classifying driver behavior can support

commercial applications such as ride-sharing services

and vehicle insurance services, provide information to

authorities, or serve as support for Advanced Driver

Assistance Systems (ADAS). However, the process of

collecting data for dangerous driving behavior is com-

plicated because the driver needs to carry out these

maneuvers, so that a classiﬁer can later detect them,

adding to situations of insecurity in trafﬁc.

In this paper, we propose a dangerous driving de-

tection system with the support of machine learning

approach which consumes inertial data extracted by

ACC and GYRO devices, towards efﬁcient data pro-

cessing. The analysis of ACC and GYRO data, which

is a feature easily captured through sensors embedded

in the vehicle itself and even in smartphones, is one of

the commonly investigated alternatives to detect dan-

gerous driving behavior. In contrast to other works

(Jeong et al., 2013) (Chen et al., 2015) (Nuswantoro

et al., 2020), to conduct our experiments, we col-

lected data in a simulation environment with a model

car that allows us to perform risky maneuvers which

would not be possible in a real environment. Typi-

cally, the use of data collected in real drive scenarios

allows some form of driver anonymization, while im-

age processing require ﬁlters to drive anonymization.

However, as drivers are subject to complying with the

law when driving on the streets, this represents a lim-

itation on the driver’s opportunities to perform risky

maneuvers. Moreover, our solution seeks to perform

efﬁciently in terms of classiﬁcation time, so that the

detection of dangerous driving can be carried out in

real time.

Controversially, features associated with acceler-

ation events did not play a signiﬁcant role in drivers

classiﬁcation (Van Ly et al., 2013). Braking and turn-

ing events can be more signiﬁcant potential in drivers

classiﬁcation. Time headway in high ﬂow freeways

can also impact the accident risk. Headway is the time

interval between successive vehicles’ head in a lane.

Shorter headway corresponds to higher risk of acci-

dents, and was found for drivers with prior accidents

or violations, young drivers, male drivers, drivers with

no passengers and as well as drivers not wearing seat

belt (Evans and Wasielewski, 1983). With regard to

the vehicles characteristics, shorter headway was as-

sociated with newer vehicles and vehicles of interme-

diate mass. Thus, a classiﬁer for detecting danger-

ous driving needs to consider more parameters, such

as headways, braking (Lattanzi and Freschi, 2021)

and turning events, than just vehicle acceleration. We

have not yet added the headway computing to our so-

lution, but we take data from a complete trip in the

analysis, precisely to consider braking and turning

events in our classiﬁcation approach.

This work is organized as following. In, Section 2,

we describe and categorize previous related work. In

Section 3, we present our methodology to detect dan-

gerous drive behaviour. In Section 4, we discuss the

obtained results. Finally, in Section 5, we present the

paper conclusion.

2 RELATED WORK

Regarding the classiﬁcation of drivers with regard to

dangerous driving behavior, the detector’s output can

be binary (aggressive/non-aggressive behavior), or on

a scale, with three or more distinct groups. In fact,

proposed techniques for detecting dangerous driving

behavior are nothing new. However, the concept of

dangerous driving sometimes involves particular as-

pects of each country. For instance, some country

legislation do not tolerate the drinking of alcoholic

beverages by drivers (e.g. Brazil, Czech Republic,

Romania, Slovakia), while others allows the ingestion

of a low amount of alcohol. Blood Alcohol Content

(BAC) drink driving limits across many Europe coun-

try usually is 0.5 grams per litre

In academia, early works did not apply Machine

Learning (ML) techniques for the detection of dan-

gerous driving behavior. A popular metric for detect-

ing dangerous driving has been the analysis of vehicle

speed. Speed and acceleration data acquired with the

support of sensors and GPS embedded in the vehi-

cle can be used to model and analyze driver behav-

ior with the support of data mining techniques (Con-

stantinescu et al., 2010). Drivers were divided into

5 types: non-aggressive, slightly aggressive, neutral,

moderately aggressive and very aggressive. Also, Dy-

namic Time Warping (DTW) can be used to binary

classify drivers (non-aggressive and aggressive) us-

ing as input data ACC measurements, GYRO, mag-

netometer, GPS, and videos (Johnson and Trivedi,

2011). DTW is a method to calculate the optimal

matching, usually with regard to time, between two

data sequences. However, evaluation was conduced

using a modest dataset. Vehicle speed can also be

used to driver classiﬁcation, such as in (Eboli et al.,

https://etsc.eu/issues/drink-driving/blood-alcohol-

content-bac-drink-driving-limits-across-europe/

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

998

2017). A percentage of the average vehicle speed (50-

80%) can be used as parameter to categorize driving

behaviors as safe, unsafe and safe, but potentially dan-

gerous.

Nowadays, in the academia, Machine Learning

(ML) techniques have been frequently applied to

detect dangerous driving using ACC and GYRO

achieved data. The most popular classiﬁers, with re-

gard to dangerous drive detection, are Support Vec-

tor Machine (SVM) (Jeong et al., 2013) (Chen et al.,

2015) (Nuswantoro et al., 2020) (Lattanzi and Fres-

chi, 2021) e CNN, which has gain more attention is

recent years (Baheti et al., 2018) (Shahverdy et al.,

2020) (Masood et al., 2020) (Zhang et al., 2020).

SVM can be applied to detect lateral hazardous

driving movements, such as lane changes and zigzag

driving (Jeong et al., 2013). Data from 172 hazardous

driving movements was collected using the vehicle’s

onboard GYRO. An average total accuracy of ≈85%

was achieved in the experiments. A set of six types

of abnormal driving behaviors was identiﬁed: weav-

ing, swerving, sideslipping, fast U-turn, turning with

a wide radius and sudden braking. Experiments were

conducted in real driving environments with 20 vol-

unteers driving during 4 months, and the proposed ap-

proach achieved an average total accuracy of ≈ 95%.

Another work (Chen et al., 2015) also applied SVM

classiﬁer to detect the dangerous drive behaviour but

using data from ACC and GYRO devices present in a

smartphone.

Recent works also compared SVM with other

classiﬁers, such as in (Nuswantoro et al., 2020) and

(Lattanzi and Freschi, 2021). Artiﬁcial Neural Net-

work (ANN) and SVM can be used for classifying

a broad set of driver movements (normal behaviour,

zig-zag, sleepy, turn right, turn left, U-turn, sudden

braking, sudden acceleration, and speed bumps) per-

formed by motorcycle drivers, such as in (Nuswan-

toro et al., 2020). Experiments were conducted in real

driving environments with 5 volunteers driving all the

movements. The proposed approach with ANN clas-

siﬁer outperformed the SVM classiﬁer with an aver-

age total accuracy of ≈96% and ≈87%, respectively.

A combination SVM and Feed-Forward Neural

Network (FFNN) classiﬁers were used to recognize

safe and unsafe driving behaviors, such as in (Lattanzi

and Freschi, 2021). Acceleration, braking and vehicle

speed values were collected in the experiments with

more than 26 hours of driving by 10 drivers. The con-

ducted experiments shown an average accuracy above

90% for both classiﬁers with a slight advantage of the

FFNN classiﬁer.

Deep learning can also be applied to the detection

of dangerous driving using inertial data from the vehi-

cle, GPS and images captured by cameras, such as in

(Khodairy and Abosamra, 2021). The classiﬁer out-

put was proposed in order to meet different demands

such as driver protection, needs of the automotive in-

surance industry and ADAS requirements. Two clas-

siﬁcation outputs were used, one with three classes

(normal, drowsy and aggressive behaviors) and a bi-

nary classiﬁcation model (safe/unsafe).

A method for dangerous driving behavior predic-

tion using a combination of cloud data and Elman

neural network (CM-ENN), based on vehicle mo-

tion state estimation and passenger’s scores about the

drivers, was proposed by (Xiang et al., 2021). Ex-

periments were conducted in real trafﬁc scenarios and

demonstrate that the proposed method is more ac-

curate and robust than classical neural network ap-

proaches.

Convolutional Neural Networks (CNN) have been

also widely applied to classify driver behavior us-

ing vehicle data (such as acceleration, gravity, RPM,

speed, and throttle (the amount of accelerator pedal

is pushed)) (Shahverdy et al., 2020), and/or images

obtained obtained through in-vehicle cameras (Zhang

et al., 2020) (Masood et al., 2020) (Baheti et al.,

2018). In fact, the most recent academic works show

a trend towards using the CNN classiﬁer. A research

question that is still under wide discussion is the need

to divide drivers in a binary way or in a more detailed

classiﬁcation. In a recent work (Shahverdy et al.,

2020), the CNN classiﬁer is able to distinguish driver

behaviour into ﬁve classes: safe or normal, aggres-

sive, distracted, drowsy, and drunk driving. (Baheti

et al., 2018) (Masood et al., 2020) (Zhang et al., 2020)

divided driver behaviour in a much larger number of

classes.

In this work, the classiﬁer delivers a binary classi-

ﬁcation but the methodology can be easily adapted to

deliver a more detailed classiﬁcation. The classiﬁer

process data achieved from ACC and GYRO avail-

able in a GoPro 10 Black camera. In contrast to other

works, which use data from real drivers, we use data

achieved from a remote-controlled scale vehicle plat-

form, which allows us to perform riskier maneuvers.

3 METHODOLOGY

In the following, we detail the methodology for clas-

sifying drivers’ behavior into safe and unsafe, using

the neural network as classiﬁer.

Machine Learning Support for Time-Efﬁcient Processing Dangerous Driving Detection Using Vehicle Inertial Data

999

3.1 Data Acquisition

In this project, a remote-controlled scale vehicle plat-

form was used to obtain inertial and ranging sensors

data. For this purpose, a GoPro 10 Black camera

was attached to the rear face of the vehicle shell (see

Fig. 2).

Figure 2: Assembly of the prototype for the remote-

controlled vehicle used in the data acquisition process.

The GoPro camera has a 5.3K-maximum resolu-

tion imaging sensor, which is also capable of captur-

ing video at 2.7K resolution with maximum frame

rate at 240 Hz (Kirschenbaum, 2021). At the same

time, it carries a Bosch BMI 260 inertial unit (Sen-

sortec, 2019) alongside the image sensor, which pro-

vides the acceleration and turning rate data onto the

video mp4-ﬁle metadata.

The GoPro 10 captures images in several resolu-

tions and frame rates, conﬁgured with a 4K image at

60 Hz frame rate for generating videos and their meta-

data. The mp4 video ﬁles from the camera were in-

terpreted through a bin2csv converter tool applied to

binary data from the ffmpeg codec for extracting the

embedded inertial data at a sampling rate of 400 Hz,

with location data sampled at 10 Hz (Irache, 2020).

In order to successfully detect dangerous driving

behaviour through the use of inertial data, it is es-

sential to complement the hardware capability with

software that can make the data useful. Therefore,

a scheme of our data acquisition system setup is de-

picted in Fig. 3.

Raw data achieved from the system is post-

processed to generate a set of features that are in-

dicative of the driving behavior. The data collec-

tion process is iterative, involving initial data gath-

ering, preliminary model training, and evaluation of

the model’s performance. Based on the model’s per-

formance, further data may be gathered to cover iden-

tiﬁed gaps or to improve performance in areas where

the model currently underperforms.

Location

≈10 Hz

≈400 Hz

≈60 Hz

Acceleration

Roll/Pitch/Yaw

Image

GoPro (image/inertial unit)

Raw Data (.csv)

Safe

Driving

Dangerous

Driving

Signal

Filters

Figure 3: Data acquisition/ﬁltering approach.

3.2 Driving Scenarios

The recognition of dangerous driving behaviour re-

quires a dataset containing data about driving move-

ments such as accelerating, braking, driving straight

and cornering, with differentiating between safe sce-

narios with moderate driving and more complex and

potentially dangerous maneuvers such as sharp turns,

sudden braking, aggressive acceleration, and erratic

lane changes. Varying road surface conditions can

lead to changes in traction, leading to different de-

grees of tire slip and subsequently, changes in accel-

eration and angular velocity.

In the same manner, the track layout, whether it is

straight or curved, and if the road has bumps and de-

pressions, this signiﬁcantly impacts the forces experi-

enced by the vehicle and hence the data captured by

the onboard sensors. The experimentation with driv-

ing scenarios was carried out in a ﬂat track with rough

tarmac with a total length of ≈30 m as represented in

Fig. 4.

Each driver started riding around 15 m straight

line to proceed performing curves over six poles

spaced 3 m from each other, and then running back

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

1000

3 5 6

Figure 4: Schematic of the track on which driving scenarios

were performed.

to the starting line after the sixth pole.

For the purpose of gathering data, 8 drivers us-

ing the remote-controlled vehicle platform performed

laps following the track layout above with speeds up

to 10 m/s. The driving sessions were organized in

safe driving runs, and then in dangerous driving runs.

Node 1 indicates the start/end of the route and each

of the nodes 2-6 indicates a pole the driver should

swerve around. The process of acquiring the dynamic

driving data accounted for wet and dry tarmac con-

ditions - there were two driving sessions - alongside

the safe and dangerous driving scenarios, aiming to

improve the output accuracy of the detection model.

3.3 Data Pre-Processing

The data pre-processing phase involves the identiﬁ-

cation and treatment of noise and situations that the

data may manifest. Therefore, it is necessary to have

the appropriate tools that help in the identiﬁcation and

processing of data.

Therefore, our dataset was reformatted and pro-

cessed before the neural network processing. In this

work, GoPro telemetry data was extracted using a free

online tool, GoPro Telemetry Extractor

When extracting data relating to ACC and GYRO

by the tool, it is accompanied by data relating to the

camera temperature, date and recording time in ms in

a tabular .csv ﬁle. The columns are renamed to more

accessible names and unnecessary data needs to be re-

moved, leaving only the recording time and ACC and

GYRO data in the 3-axes x, y and z. Additionally, the

data on the vertical axis must have the gravity value

(9.8) subtracted and the recording time is converted to

seconds for convenience. Fig. 5 depicts a sample of

the dataset after the conclusion of the pre-processing

step.

In the data, there is still a lot of high-frequency

noise, so it is necessary to ﬁlter it using a LowPass

Filter (LPF). LPFs are a good way to remove noise

(both mechanical and electrical) from the ACC. The

LPF used is the Inﬁnite Impulse Response (IIR) ﬁlter.

The IIR ﬁlter is used so that the computing process

can be carried out as quickly as possible. IIR ﬁlters

https://goprotelemetryextractor.com/free/

Figure 5: Sample of the dataset after the conclusion of the

pre-processing step.

Figure 6: ACC data: in red, ﬁltered data. In back, raw data.

do not have a phase delay, so there is no time delay.

Fig. 6 depicts a sample of our data before and after

the application of the IIR ﬁlters.

Subsequently, downsampling in the data fre-

quency must be carried out so that they can be pro-

cessed quickly by the neural network. The GoPro col-

lects data at a frequency of 200 Hz, which has been re-

duced to 10 Hz. Fig. 7 depicts a sample of our dataset

after frequency downsampling to 10 Hz.

Figure 7: ACC data after downsampling to 10 Hz.

Finally, each 5s stretch within the dataset will be-

come an observation labeled as safe or dangerous. To

extend our dataset, the concept of sliding window was

applied: between 0-5s will be considered a stretch,

between 1-6s also and so on, as depicted in Fig. 8. A

sample of the ﬁnal dataset is depicted in Fig. 9.

3.4 Dataset Calibration

For experimental evaluation, a total of 374 cases were

collected, of which 268 were safe and 106 (size) were

Machine Learning Support for Time-Efﬁcient Processing Dangerous Driving Detection Using Vehicle Inertial Data

1001

...

sliding window

Figure 8: Sliding window approach.

Figure 9: Sample of the ﬁnal dataset.

dangerous. Since, each data in our dataset contains

6 kbytes, the total dataset size is 2244 kbytes or ≈ 2

Mbytes. The imbalance in the dataset occurred due

to the speed at which the dangerous laps were per-

formed, which ended up generating shorter videos.

The treatment of datasets with unbalanced data is

a recent problem. Algorithms used to training neural

networks have difﬁculty to learn in the presence of

unbalanced data, i.e. when there is a large difference

in the number of samples in each class (safe/unsafe).

This difference can lead the classiﬁer to have a biased

output, favoring the outstanding class.

Because of this and to prevent the data from being

trained in an unbalanced way in the neural network,

162 safe cases were removed from the dataset. The

remaining 212 cases were divided between a training

dataset containing 180 data and a testing dataset con-

taining 32 cases.

3.5 Neural Network Model

Our classiﬁer, which the code is depicted in Fig. 11,

follows a neural network schematic model used in this

work consists of three layers, as depicted in Fig. 10.

In each network layer, with a set of artiﬁcial neurons,

the network can learn about the data, by extracting its

characteristics.

The input layer has 300 nodes (ACC and GYRO

data for 5 seconds of the section at a frequency of

10 Hz). The two dense intermediate layers with the

Figure 10: Neural network schematic model.

ReLU activation function, have 256 nodes and with

128 nodes, respectively. The output layer with the

sigmoid activation function, the values close to 1.0

indicate a safe trip and values close to 0.0 indicate a

dangerous trip.

3.6 Training

The training process of the neural network occurs ev-

ery round or every epoch. During each epoch, the

process of updating the weights takes place in order

to improve the results obtained. The weights are ad-

justed in the neural network during the learning pro-

cess.

After running all epochs, the trained model is ob-

tained, which has learned features about the dataset

used as input. Therefore, when displaying new infor-

mation to this network, which is within the class in

which the model was trained, in this case the detec-

tion of dangerous driving behavior, it is expected that

the model will have the ability to correctly classify the

information presented (in safe/unsafe).

All our experiments were conducted with a con-

ventional entry-level laptop from 2016: Intel(R)

Core(TM) i7-5500U CPU @ 2.40GHz, 8GB DDR3

RAM and a Nvidia GeForce 940M 2GB DDR3 GPU

running Linux Mint 21.1. We divided our training

dataset with 180 examples into batches of 16 exam-

ples. Training was carried out using 25 epochs (as

depicts Fig. 12).

4 RESULTS

The proposed model achieved an accuracy rate of

85.61% for test cases that were not seen by the neu-

ral network during training (again see Fig. 12). The

achieved accuracy rate indicates that there is a pat-

tern to be found in the trips, however, it is not a high

enough rate for the purpose of classifying and report-

ing dangerous driving, since it presents many false

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

1002

model = tf.keras.Sequential([

preprocessing_layer,

tf.keras.layers.Dense(256, activation = ’relu’),

tf.keras.layers.BatchNormalization(),

tf.keras.layers.Dropout(0.1),

tf.keras.layers.Dense(128, activation = ’relu’),

tf.keras.layers.BatchNormalization(),

tf.keras.layers.Dropout(0.1),

tf.keras.layers.Dense(1, activation=’sigmoid’),

])

model.compile(

loss=’binary_crossentropy’

optimizer = tf.keras.optimizers.Adam(learning_rate=0.01),

metrics =[’binary_accuracy’]

)

Figure 11: Classiﬁer implementation.

Figure 12: Neural Network Learning Curve with regard to

the accuracy.

negatives. This happens because the dataset needs to

be better calibrated. However, we want to keep the

amount of data small, so that the neural network can

readily process it. To achieve better accuracy results,

possible solutions are (i) expanding the database, (ii)

improving pre-processing phase and (iii) testing dif-

ferent hyper-parameters.

5 CONCLUSION

Data quality is one of the main concerns for ML al-

gorithm’s output. Most of the available methods in-

duces knowledge strictly from the data, without using

any external information. Thus, the quality of the al-

gorithm’s output largely determined by the quality of

the input data.

In this way, future work includes to augment the

database data and to improve data pre-processing

phase, seeking precisely to improve the accuracy of

the algorithm here applied. Furthermore, we aim

to improve the neural network accuracy, using other

techniques such as recurrent neural networks, in addi-

tion to evaluating the use of different neural network

models to solve the given problem, such as multi-

class classiﬁers for identifying maneuvers performed

by the vehicle or for recognizing drivers. Finally, we

aim to apply the classiﬁcation of dangerous detection

behavior in a real time environment to evaluate the

performance of the solution given the computational

power and resources available in the vehicle.

ACKNOWLEDGEMENTS

We would like to thank the Brazilian Government and

the FUNDEP Foundation for supporting this project

under the Rota 2030 program, and we appreciate the

support provided by the following people at the Poly-

technic College at the University of S

ao Paulo work-

ing with us in the project: Prof. Marcelo K. Zuffo,

Prof. Agenor T. Fleury, Prof. Fl

avio Trigo.

REFERENCES

Abu Al-Haija, Q. and Krichen, M. (2022). A lightweight

in-vehicle alcohol detection using smart sensing and

supervised learning. Computers, 11(8).

Baheti, B., Gajre, S., and Talbar, S. (2018). Detection of dis-

tracted driver using convolutional neural network. In

Proceedings of the IEEE conference on computer vi-

sion and pattern recognition workshops, pages 1032–

1038.

Chen, Z., Yu, J., Zhu, Y., Cheny, Y., and Li, M. (2015).

D3: Abnormal driving behaviors detection and identi-

ﬁcation using smartphone sensors. 2015 12th Annual

IEEE International Conference on Sensing, Commu-

nication, and Networking (SECON), pages 524–532.

Constantinescu, Z., Marinoiu, C., and Vladoiu, M. (2010).

Driving style analysis using data mining techniques.

International Journal of Computers, Communications

& Control (IJCCC), V:654–663.

Eboli, L., Guido, G., Mazzulla, G., Pungillo, G., and

Pungillo, R. (2017). Investigating car users’ driv-

ing behaviour through speed analysis. Promet - Traf-

ﬁc&Transportation, 29(2):193–202.

Evans, L. and Wasielewski, P. (1983). Risky driving related

to driver and vehicle characteristics. Accident Analysis

& Prevention, 15(2):121–136.

Irache, J. (2020). Gopro metadata format parser. Technical

report, GoPro.

Jeong, E., Oh, C., and Kim, I. (2013). Detection of lat-

eral hazardous driving events using in-vehicle gyro

sensor data. KSCE Journal of Civil Engineering,

17(6):1471–1479.

Johnson, D. A. and Trivedi, M. M. (2011). Driving style

recognition using a smartphone as a sensor platform.

In 14th International IEEE Conference on Intelligent

Transportation Systems (ITSC), pages 1609–1615.

Khodairy, M. A. and Abosamra, G. (2021). Driving

behavior classiﬁcation based on oversampled sig-

nals of smartphone embedded sensors using an opti-

mized stacked-LSTM neural networks. IEEE Access,

9:4957–4972.

Machine Learning Support for Time-Efﬁcient Processing Dangerous Driving Detection Using Vehicle Inertial Data

1003

Kirschenbaum, M. (2021). Gopro hero 10 teardown. Tech-

nical report, GoPro.

Lattanzi, E. and Freschi, V. (2021). Machine learning tech-

niques to identify unsafe driving behavior by means

of in-vehicle sensor data. Expert Systems with Appli-

cations, 176:114818.

Masood, S., Rai, A., Aggarwal, A., Doja, M. N., and Ah-

mad, M. (2020). Detecting distraction of drivers us-

ing convolutional neural network. Pattern Recognition

Letters, 139:79–85.

Nuswantoro, F. M., Sudarsono, A., and Santoso, T. B.

(2020). Abnormal driving detection based on ac-

celerometer and gyroscope sensor on smartphone us-

ing artiﬁcial neural network (ann) algorithm. In 2020

International Electronics Symposium (IES), pages

356–363.

Sensortec, B. (2019). BMI260 : Accurate, low-power In-

ertial Measurement Unit (IMU). Bosch Sensortec

GmbH.

Shahverdy, M., Fathy, M., Berangi, R., and Sabokrou, M.

(2020). Driver behavior detection and classiﬁcation

using deep convolutional neural networks. Expert Sys-

tems with Applications, 149:113240.

Shreshtha, S., Singh, P., Singh, R., Arif, S., and Sinha, D.

(2020). Non-invasive alcohol detection for drunk driv-

ing prevention. In 2020 2nd International Conference

on Advances in Computing, Communication Control

and Networking (ICACCCN), pages 332–337.

Van Ly, M., Martin, S., and Trivedi, M. M. (2013). Driver

classiﬁcation and driving style recognition using iner-

tial sensors. In 2013 IEEE Intelligent Vehicles Sympo-

sium (IV), pages 1040–1045. IEEE.

Wang, D., Pei, M., and Zhu, L. (2014). Detecting driver

use of mobile phone based on in-car camera. In 2014

Tenth International Conference on Computational In-

telligence and Security, pages 148–151.

Willis, M., Zaouk, A., Bowers, K., Chaggaris, C., Shannon-

Spicer, R., Bahouth, G., and Strassburger, R. (2019).

Driver alcohol detection system for safety (DADSS)

- pilot ﬁeld operational tests (PFOT) vehicle instru-

mentation and integration of DADSS technology. In

NHTSA – 26th International Technical Conference on

The Enhanced Safety of Vehicles (ESV).

Xiang, H., Zhu, J., Liang, G., and Shen, Y. (2021). Predic-

tion of dangerous driving behavior based on vehicle

motion state and passenger feeling using cloud model

and elman neural network. Frontiers in Neurorobotics,

15.

Yasar, H. (2017). Detection of driver’s mobile phone us-

age. In 2017 IEEE 9th International Conference on

Humanoid, Nanotechnology, Information Technology,

Communication and Control, Environment and Man-

agement (HNICEM), pages 1–4.

Zhang, C., Li, R., Kim, W., Yoon, D., and Patras, P. (2020).

Driver behavior recognition via interwoven deep con-

volutional neural nets with multi-stream inputs. IEEE

Access, 8:191138–191151.

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

1004