Naturalistic Driving Studies Data Analysis Based on a Convolutional

Neural Network

Jamal Raiyn

and Galia Weidl

Computer Science Department, Al Qasemi Academic College, Baqa Al Ghariah, Israel

Connected Urban Mobility, Faculty of Engineering, Technical University of Applied Sciences, Aschaffenburg, Germany

Keywords: Convolutional Neural Network, Deep Learning, Autonomous Vehicle, Anomalous Data.

Abstract: The new generation of autonomous vehicles (AVs) are being designed to act autonomously and collect travel

data based on various smart devices and sensors. The goal is to enable AVs to operate under their own power.

Naturalistic driving studies (NDSs) collect data continuously from real traffic activities, in order not to miss

any safety-critical event. In NDSs of AVs, however, the data they collect is influenced by various sources that

degrade their forecasting accuracy. A convolutional neural network (CNN) is proposed to process a large

amount of traffic data in different formats. A CNN can detect anomalies in traffic data that negatively affect

traffic efficiency and identify the source of data anomalies, which can help reduce traffic congestion and

vehicular queuing.

1 INTRODUCTION

The rapid growth of smart cities, with their severe

traffic congestions, road safety issues and

environmental pollution problems, presents

autonomous vehicles with challenges, including an

increased risk of human operating error. Modern

autonomous vehicles (AVs) are equipped with

multiple sensors, such as cameras, radars and LIDAR,

which help them to better perceive their surroundings

and plan travel routes. These sensors generate a

massive amount of data and one challenge to

autonomous vehicles is how to manage the massive

amount of data that is collected by these various

devices (Arena and Pau, 2019). AVs use the data to

operate independently, communicating, negotiating

and making decisions. These attributes make the new

generation of AVs one of the key areas of applications

for artificial intelligence (Li at et 2018). AVs face

many challenges, such as anomalous data that is

caused by geographical factors and cyber-attackers.

Some road data arrive in an incomplete or falsified

form (Xiang et al, 2019; Qin et al, 2021) and can result

in abnormal conditions on urban roads (Raiyn, 2021;

Zhang eta l, 2020). Currently, many research

enterprises and universities around the world have

taken the initiative to design AVs with artificial

Intelligence, so that they can perform better than the

human brain. A convolutional neural network (a CNN

or ConvNet) (Yamashita et al, 2018) is designed to

mimic the working principle of the human brain and

is trained with large Big Data sets to perform various

tasks. To manage road networks, such a network is

proposed here, with the goal of detecting road sections

that are overloaded because they have been given

inappropriate information. A CNN is a specialized

form of artificial intelligence technology that analyzes

input data which contain some form of spatial

structure. CNNs are used to solve visual

computational problems, such as those associated

with self-driving cars, robotics, drones, security,

medical diagnoses, and agriculture. In this paper, A

CNN (Arena and Pau, 2019) is proposed. CNNs are

considered to be one of the most important forms of

artificial intelligence. They consist of a

computational elements (neurons) heavily connected

to each other, which process perceptual data from the

surrounding environment, such as images of road

traffic from satellites and drones, and information

from sources based on vehicle-to-vehicle (V2V) and

vehicle-to-cloud (V2C) communication and from

digital resources like Google Maps.

1.1 Literature Review

The field of anomaly detection has been widely

researched (Santhosh et al, 2020; Aradi, 2022).

248

Raiyn, J. and Weidl, G.

Naturalistic Driving Studies Data Analysis Based on a Convolutional Neural Network.

DOI: 10.5220/0011839600003479

In Proceedings of the 9th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2023), pages 248-256

ISBN: 978-989-758-652-1; ISSN: 2184-495X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

Anomaly detection approaches are typically divided

into two types: model-based and data analysis-based.

Model-based approaches mostly use algorithms that

are very accurate, such as machine learning schemes

(Xiang et al, 2019), whereas the approaches based on

data analysis usually use statistical measurements. On

urban roads, anomalies cause discomfort to drivers

and have a negative impact on traffic efficiency (Silva

et al, 2018). When traffic accidents occur and there is

a congestion, the resulting traffic flow becomes

abnormal. In an AV network, anomalies are caused by

traffic accidents, bad weather, road work, and

repeated lane changing attempts. In addition, there

are other challenges (Mishra et al. 2021) faced by

AVs, such as noise and interference, which are further

sources of anomalies in traffic flow. AVs collect

various types of data via onboard devices and

communication with devices on the Internet of Things

(Aradi, 2022). Both the onboard devices and the AV

communications protocols are affected by

interferences and delays (Niknam et al, 2018). AV

positioning data is collected by the Global Navigation

Satellite System (GNSS) (Raiyn, 2017), whose

satellites are mainly located in medium earth orbits

(MEO). The signals transmitted by a satellite

propagates through the atmosphere, where they are

subject to delays caused by ionospheric and

tropospheric media. At ground level, multipath

effects, namely the reception of signals that are

reflected from obstacles such as buildings

surrounding the receiver, can occur, causing one of

the largest types of errors, one that is also difficult to

model, as it depends strongly on the receiver

environment. Increased delays may affect the

performance metrics of a positioning terminal, which

are characterized in terms of availability, accuracy,

and integrity. Delays may also be caused by the weak

performance of network equipment and differences in

equipment attributes. Many untargeted transmitted

signals can even interfere with transmitted signals

(Raiyn, 2021). Now, there are new challenges that

have yet to be considered, such as communication

between AVs in heterogeneous wireless networks.

Heterogeneity can cause delays in communication

between AVs. For AV communication in

differentiated wireless networks that provide differing

quality of service (QoS), middleware is needed to

make adaptations. Another challenge faced by AVs is

identifying road sections with a high degree of noise.

Furthermore, the devices in an AV that has been

involved in an accident may stop working and render

it unable to communicate with surrounding vehicles.

These and many other factors, cause anomalies in data

that can increase daily traffic congestion. To detect

road traffic anomalies, various forecasting schemes

have been proposed (Yan et al, 2018). Recently, some

deep learning approaches have been introduced to

predict urban traffic flow. The most established

algorithm among various deep learning models is

using CNNs, a class of artificial neural networks that

has been the dominant technology for computer vision

tasks since it first started producing astonishing

results. Deep neural networks were proposed to

predict traffic condition on highways by considering

spatio-temporal correlations in traffic data attributes.

To make more accurate forecasts, an advanced CNN

can also incorporate data sources, such as weather

forecasts, and police reports. In (Yamashita et al,

2081), a CNN was designed to automatically and

adaptively learn a spatial hierarchy of features

through backpropagation by using multiple building

blocks, such as convolutional layers, pooling layers,

and fully connected layers. A CNN can exploit the

non-linear regularities of network traffic, providing

significant improvements with respect to the mean

absolute and standard deviation of data (Mozo et al,

2018). The performance of several models was

compared using different accuracy measurement

methods (e.g., Root Mean Square Error [RMSE] and

Mean Absolute Percentage Error [MAPE]). The

results indicate the good performance of the Prophet

and CNN models (Yan et al, 2018).

1.2 Problem Description

Various classical forecast schemes have been

proposed to manage road traffic flow. These schemes

perform well in historical data management, however,

they are lacking in real-time forecasting. Traffic flow

is becoming more complex, and information about it

has evolved from a single format to a conglomeration

of formats providing very large datasets, known as

Big Data. To process Big Data, a new scheme is

proposed which is based on computational

intelligence. Furthermore, the forecasting schemes are

designed to process dynamic traffic flow instead of

fixed traffic flow. Classical computing schemes,

performed short-term forecasting for a given time slot

and considered traffic flow as a fixed entity,

represented by 0 for busy and 1 for free. However,

real-time traffic flow dynamic can change every

millisecond. The proposed CNN considers dynamic

traffic flow, represented by 00 for busy, 11 for free

and 01 for busy changed to free and 10 for free

changed to busy.

Naturalistic Driving Studies Data Analysis Based on a Convolutional Neural Network

249

1.3 Research Contribution

The main contributions of the model described in this

article are as follows: Firstly, it can detect anomalies

in traffic data. Second, it can extract the features of

traffic data and identify the sources of anomalous data

(Raiyn, 2022). Third, it can perform a short-term

travel forecasts by applying CNNs, and fourth, it

makes it possible to compare the accuracy of results

through measurement methods such as RMSE and

MAE. The results are analyzed and the best method

for prediction is selected. This paper is organized as

follows: Section 2 gives an overview of deep

learning. Section 3 describes the methodology and

discusses the development of the CNN. Section 4

concludes the discussion and point out directions for

future research.

2 DEEP LEARNING

TECHNOLOGY

Deep-learning (DL) methods are some of the most

useful tools in the area of machine learning (Li et al,

2018; Xiang et al, 2019) (Berman et al, 2019). A DL

system learns the features of a dataset and then

combines them to achieve a specific goal. In this

paper, a DL method is proposed for the early

detection of traffic anomalies (Nguyen et al, 2018). It

is composed of two main phases: training and testing.

The system proposed here is composed of four

phases: (1) preparation of the dataset, (2) a training

phase, (3) a testing phase, and (4) a performance

metrics phase (Wang et al, 2019; Raiyn, 2017). The

deep learning (DL) system, in general, is structured in

three main parts: an input layer, a hidden layer, and

an output layer. The input layer for DL consists of a

large amount of data that is received from different

sources. The big dataset for traffic modeling is diverse

and comes from variety of devices and systems, such

as cameras, LIDAR, sensors, and GNSSs (Raiyn,

2017). The neurons have an important influence on

the learning ability of the algorithm; too few can lead

to insufficient learning, and too many can lead to

overfitting. The output layer is responsible for

exporting the values, or the vectors of the values, that

correspond to the format required for the problem,

and it presents the results visually based on

measurements of statistical error.

2.1 Convolutional Neural Network

CNNs are a type of deep learning model for

processing data, one- dimensional, two- dimensional

and three- dimensional (Ensafi et al, 2022; Cordeiro

et al, 2021). They are designed to automatically and

adaptively learn the spatial hierarchies of features,

from low- to high-level patterns. They are typically

composed of three types of layers: convolutional,

pooling, and fully connected layers. The first two,

convolutional and pooling layers, perform feature

extraction, whereas the third, the fully connected

layer, maps the extracted features into final output,

such as classification. The architecture of a CNN

consists of three distinct types of layers: a

convolutional layer, pooling layer, and a

classification layer. The convolutional layers are the

core of the CNN. The weights define a convolutional

kernel applied to the original input, a small window

at a time, called a receptive field. The results from the

application of these filters across the entirety of the

input are then passed through a non-linear activation

function, typically a ReLU, to produce a feature map.

The three layers in the architecture of a CNN are as

illustrated in Figure 1.

Figure 1: CNN architecture.

3 METHODOLOGY

The aim of this study is to make time-series forecasts

with the help of different models, such as, exponential

smoothing and CNNs. In this investigation, the model

with the most accurate results qualifies as the best

model (Alzubaidi et al, 2021). The first step in

anomaly detection is to define normal traffic on a

section of urban road and then to flag as anomalies

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

250

any observations that do not fit this normal pattern.

Finding these patterns is the main challenge in

detecting anomalies in urban road traffic. This section

describes model based on a CNN that can be applied

to all types of normally trafficked roads. Through the

analysis, the CNN extracts anomalous data. An

important step is cleaning and formatting the data.

Travel data are collected by different devices and are

influenced by human and environmental factors.

Data cleaning is the process of modifying the data to

ensure that the datasets are free of irrelevancies and

incorrect or incomplete travel observations. The best

computational data formats have several useful

features; for instance, they are easy for computers to

parse, easy for people to read, and widely usable by

other tools and systems. The travel data input for this

study came from various smart devices in all kinds of

formats. This dataset was cleaned to improve the

efficiency of the data analysis and the quality of the

results. Due to noise and environmental factors that

influence data communication, some of the datasets

collected were not complete; in this case, data

cleaning involved identifying the fields from which

data were missing and then properly compensating

for them. An interactive exploratory travel dataset

was useful for representing errors in rea-time. The

deep learning scheme extracted the attributes from

the input data. The attributes were classified and

assigned scores.

errorwxy

nnAD



(1)

Here, x

is the input signal, w

is the weight

corresponding to each input signal, and y

is the

output signal. F

is the activation function, the means

of calculating the sum of the data coming from the

input.

3.1 Convolutional Neural Network 2D

AV systems use CNNs to detect the behavior of urban

road traffic. CNN can handle the urban road network

in old cities. In this case, the input data consisted of

the geographical locations of all the data sources in an

urban area that was powered by online map services,

such as Google Maps, as illustrated Figure 2a. The

images in the figure are raw snapshots from a traffic

congestion map for an urban area in Kaboul town,

whose roads are constructed according to Palestinian

building standards, which means they are narrow and

crooked. The raw snapshots for this area mainly cover

urban arterial roads; each is 400 pixels wide and 400

pixels high. The snapshots were retrieved during

morning rush hours between 8:00 and 10:00 in June

2, 2022, through a free API provided by an online map

service provider. These snapshots were the initial

source of traffic congestion data. The procedure

included cleaning the data as illustrated in Figure 2b,

c and d.

(a) (b)

(c ) (d)

Figure 2: Preparing the input data.

The second phase included a hidden phase and an

output phase as explained below.

3.2 Forecasting Model

In this section, the framework for representing the

traffic congestion data for a road network is described.

The proposed representation framework consists of

two steps. The first step, which segments the original

traffic congestion is prepared according to the data

science life cycle. The second step reduces all the

values in each grid using CNN one- and two-

dimensional operations to reduce in a manner similar

to image down-sampling. To detect data anomalies,

two approaches are proposed: use of the exponential

moving average and use of a CNN.

3.2.1 Classical Time-Series Forecasting

In (Li, et al, 2018), the exponential smoothing scheme

is used for time- series forecasting. The aim of this

scheme is to detect data anomalies in road flow traffic.

The exponential smoothing scheme lends greater

weight to the most recent travel observations, so that

the older the observation, the less it affects the

forecast.

....

221101

+++=

−−+ tttt

yyyy

θθθ

(2)

Where

)1(

ααθ

−=

is the smoothing

constant, and

10 ≤<

Hence the forecast scheme results in

....)1()1(

+−+−+=

−−+ tttt

yyyy

ααααα

(3)

-1 -1 -1 -1 -1 -1 -1 -1 -1 -1

-1 -1 -1 -1 -1 1 -1 -1 -1 -1

-1 1 -1 -1 -1 1 -1 -1 -1 -1

-1 -1 1 -1 -1 1 -1 -1 -1 -1

-1 -1 -1 1 1 1 1 -1 -1 -1

-1 -1 -1 -1 -1 1 -1 1 -1 -1

-1 -1 -1 -1 1 -1 -1 -1 1 -1

-1 -1 -1 1 -1 1 -1 -1 -1 1

-1 -1 1 -1 -1 -1 -1 -1 -1 -1

Naturalistic Driving Studies Data Analysis Based on a Convolutional Neural Network

251

which can be easily described as follows:

ttt

yyy

)1(

αα

−+=

(4)

Exponential smoothing allows us to update the

prediction when a real time observation is available.

Anomaly detection based exponential smoothing is

performed based on a travel data analysis.

Exponential smoothing uses statistical error

measurements to characterize the integrity and

consistency of the travel data. Data anomalies

associated with road accidents are characterized by

alterations in travel speed which starts to decrease in

the upstream direction and increase in the downstream

direction. Furthermore, significant differences

become evident between travel speeds and standard

deviations.

3.2.2 Convolutional Neural Network

• Input layer

The input for the model is the time- series data for

road traffic. The lengths of input and output time

intervals can be expressed as F and P, respectively.

The model input can be written as:

]1,1[],,...,,[

+−−∈=

−++

FPNimmmx

Piii

(5)

Where I is the sample index, N is the length of the

time intervals, and m

is a column vector representing

the traffic speed of all the road sections in a

transportation network within one time unit. In some

highways in Middle East countries, the average speed

is 90 km. Travel observations that exceed that

threshold are represented by a value between 0 and -

1; otherwise, they are represented by a positive value

between 0 and 1 as illustrated in Figure 3. The CNN

is applied to detect anomalies in the road traffic data.

In this case, only the negative values are considered

and at the same time present the travel data that are

smaller than the threshold.

-0.3

-0.5

-0.9

0.9

0.8

0.7

-0.3

-0.5

-0.9

0.1

0.4

-1

1,1

1,2

1,3

)(

111



jjj

bxw

Figure 3: Features extraction.

The extraction of traffic features involves a

combination of the convolutional and pooling layers.

The output of the first convolution and pooling layers

can be written as:

]

,1[)),

111

((

Wpool

o ∈+=

(6)

and the output of the last convolutional and

pooling layers can be written as

],1[)),((

Wpool

o ∈+=

where

is the activation function. In the

prediction, the features learned and outputted by

traffic feature extraction are concatenated into a dense

vector that contains the final and the highest-level

features of the transportation network input. The

dense vector can be written as

flatten

cjoooflatteno == ]),,...,,([

(7)

where L is the depth of CNN. Finally, the vector

is transformed into output through a fully connected

layer. The output can be written as:

boWy

flatten

(8)

bbxWpoolflatenW

++=



−

)))))(((((

(9)

where W

and b

are the parameters of the fully

connected layer, and

represents the predicted

network-wide data anomalies. The CNN uses

convolutional filters on its input layer and obtains

local connections only where local input neurons are

connected to an output neuron (in the convolutional

layer). Hundreds of filters are sometimes applied to

the input, and the results are merged in each layer.

One filter can extract one traffic feature from the input

layer and, therefore, hundreds of filters can extract

hundreds of traffic features as illustrated in Figure 4.

The fully connected layer expresses the negative

values, that represent the anomalies in each road

section.

Figure 4: Fully connected layer.

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

252

3.3 Comparative Methods

To manage road traffic flow, two approaches are

proposed.

• The first scheme is the exponential smoothing

scheme. Used for short-term forecasting, this

scheme is based on both historical and real-time

travel data. It calculates a short-term forecast for

a given period of time (e.g. 15 minutes, a day, a

week) without considering changes in road traffic

flow after that. The output of the exponential

smoothing scheme is a value that represents the

speed on the road. The result is saved in 0, 1

format, which described the road traffic as busy

or free respectively. The second approach

involves CNNs with one and two dimensions. A

1D CNN is used to handle time- series road

traffic; the 2D CNN can handle images such as

maps.

• An exponential smoothing scheme needs a large

amount of input data to make forecasts, however,

the CNN can use an image of a road section to

perform traffic analysis and make forecast.

• The input travel data should be structured for an

exponential smoothing scheme; however, the

advantage of a CNN (1D, 2D, and 3D) is that it

can process any data format.

The CNN can identify sources of abnormalities.

Detecting abnormalities is a complex task due to the

variability of activity from one case to another, and

due to the absence of standardized datasets

representing normal and abnormal activity. For

anomaly detection, the CNN is trained on data

representing the activity and behavior of the sources

of abnormalities. The initial set of data, collected from

time- series representing the usual behavior of road

traffic, serves as a baseline for training the novelty

detection models. Road observations recognized as

anomalies are explored further to determine the

sources of the abnormalities, as depicted, and the

CNN is trained to identify further abnormalities based

on the behavior and activity of other sources.

4 EVALUATION OF

FORECASTING SCHEMES

In general, AVs collect a variety of traffic data. Low-

quality data can lead to traffic congestion and

collisions. Furthermore, data reception may be

incomplete due to the urban noise produced by

network tunnels. The raw traffic data used in this

project was mainly obtained from smart phones in

2.5-minute cycles. When a statistical analysis of the

original data was carried out, it was found that it

involves two main defects; the dataset was

incomplete, and the data contained noise. Urban roads

are divided into sections each 300 meters long. The

data were processes by several operations: statistical

measurements were used to detect incomplete data,

and unnecessary material such as noise was removed

(Alrajhi et al, 2019). The loss function was a

customized mean squared error (MSE), as defined in

Eq. 10. lMSE which is described in Eq. 11, was

applied to calculate the mean squared error at

different congestion levels with different priorities.

pMSE in Eq. 12 was applied to calculate the mean

squared error at different congestion levels with

different positioning.



−=

MSE

)

(

(10)



−=

iii

yyl

lMSE

)

(

(11)



−=

iii

yyp

pMSE

)

(

(12)

The cost function is a special type of functions

that helps to minimize error and to approach as close

as possible the expected output. It uses two

parameters to calculate error: one is an estimated

output of the CNN model (also called the prediction)

and other is the actual output. The mean squared error

is considered one of the most familiar loss functions

as it is much like the least square loss function. It

directly calculates the difference between the

predicted result and the true label, which is denoted as

in Eq.13. A basic working out of error can be formed

with the input data when there is an actual value and

a predicted value. Error can be characterized as the

difference between the predicted value and the actual

value. The loss function in the first convolutional and

pooling layers can be formulated as



−=

yxh

2)()(

))((

),J(

θθ

(13)

The loss function of all the convolutional and

pooling layers can be formulated as



−=

conv

yxh

2)()(

1-n

))((

),J(

θθ

(14)

The loss function of all the convolutional and

pooling layers for all the road section can be

formulated as

Naturalistic Driving Studies Data Analysis Based on a Convolutional Neural Network

253



−=

tion conv

yxh

sec 1

2)()(

1-n

))((

),J(

θθ

(15)

where

)(xh

is the prediction that is closest to the

actual value, y. To reduce forecasting error, the CNN

uses a self-management strategy to control the output

of each convolutional layer. The calculation of loss

function is related to road sections. In training the

CNN, weights are selected that capture a desired

input-output relationship. This training objective can

be framed as a minimization of a loss function, which

quantifies the difference between the output of the

network and the ground truth values from the training

set. The forecast of the CNN are given in terms of

traffic speed on different road sections, and the mean

squared errors (MSEs) are employed to measure the

distances between the forecasts and actual traffic

speed. Minimizing MSEs is taken as the training goal

of the CNN. MSE can be written as



−=

MSE

)

(

when the model parameters are set , the optimal

values of

can be determined according to the

standard back propagation algorithm, which is used

in other studies of CNN:

),(min

212

101

nnn

θθ

−

(16)

where,



−=

)

(

argmin

||||

argmin ybW

flatten

−+=

σθ

||)))))(((((||

argmin

ybbXWpoolflattenW

−++=



−

σθ

The quality of a forecast depends on the training of

the neural network

);(

, with the aim of finding a

suitable set of parameters

so that the model can

achieve good performance. The task of training a

neural network is equivalent to optimizing the loss

function by back-propagation iteration. More

precisely, a loss function outputs a scalar value which

is regarded as a criterion for measuring the difference

between the predicted result and the true label for one

sample. During training, our goal was to minimize the

scalar value for m training samples (i.e., the cost

function).

• Detecting of the source of data anomalies

The CNN was capable of extracting a variety of

data from a given 2D image, such as road traffic data,

AV positioning, and anomalous data and their sources.

Anomalies in the data may stem from geographical

factors or experimental or human error. To elicit good

performance from the CNN, the training dataset

should be free from anomalies. An exponential

smoothing scheme employing statistical functions was

used to detect anomalous data. Simple statistical

functions were applied to detect univariate anomalous

feature values in the data sets as illustrated in Figures

5. An anomaly score was computed for all

observations; if the score was greater than a given

threshold value, the data were considered anomalous.

Figure 6 show data with missing values that were

collected with a uBlox device. Figure 7 show the

detected anomalies in road traffic.

Figure 5: Statistical measurment error.

Figure 6: Missing data.

Figure 7: Variation in travel time.

speed(km/h)

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

254

5 CONCLUSION AND FUTURE

WORK

This paper proposes a CNN scheme that can predict

road traffic speed based on extracted features from 2D

images. The main goal of this project was to detect

anomalies in road data and their sources. The

discussion began by introducing the new challenges

that faces modern AVs. This was followed by an

overview of applications of artificial intelligence in

AVs, such as deep learning algorithms and more

specifically CNNs, including 1D, 2D, and 3D

convolution processing alternatives. The development

of a new generation of AVs equipped with various

sensors and Internet of Things devices calls for new

data management schemes. The main drawback of

traditional traffic forecasting schemes is that they

cannot manage data in different formats. Furthermore,

traditional forecasting schemes process a traffic road

network in its entirety, which increases the processing

time. On the other hand, the application of CNNs in

road traffic detection has demonstrated significant

improvements over traditional approaches. A CNN

can process three forms of input data such 1D, 2D and

3D. The advantages of using a CNN are that it can

process data in many independent layers and each

layer can be optimized.

ACKNOWLEDGMENTS

This study has been supported by the Project

101076165 — i4Driving within Horizon Europe

under the call HORIZON-CL5-2022-D6-18 01-03,

which is programmed by the European Partnership on

‘Connected, Cooperative and Automated Mobility’

(CCAM).

CONFLICT OF INTEREST

The authors declare that they have no conflict of

interest.

DATA AVAILABILITY

The majority of the datasets used in this paper are

publicly available. Private datasets can be furnished

upon request.

REFERENCES

Arena, F., Pau, G.: An Overview of Vehicular

Communications, future internet,11,27. pp.112.

(2019).

Li, J., Cheng, H., Guo H., Qiu, S.: Survey on Artificial

Intelligence for Vehicles”,

Automotive Innovation, Springer, (2018). https://doi.org/

10.1007/s42154-018-0009-9

Xiang, W., Huang, T., Wan, W.: Machine learning based

optimization for vehicle-to-infrastructure Communica-

tions, Future Generation Computer Systems 94, 488–

495. (2019).

Qin, H., Yan, M., Ji, H.: Application of Controller Area

Network (CAN) bus anomaly detection based on time

series prediction, Vehicular Communications, 27.

(2021).

Raiyn, J.: Classification of Road Traffic Anomaly Based

on Travel Data Analysis, International Review of Civil

Engineering (IRECE), 12, 4. (2021).

Zhang, S., Li, S., Li, X., Yao, Y.: Representation of Traffic

Congestion Data for Urban Road Traffic Networks

Based on pooling Operations, Algorithms, 13(84), 1-18.

(2020).

Yamashita, R., Nishio, M., Do, R.K.G., Togashi, K.:

Convolutional neural networks: an overview and

application in radiology, Insights into Imaging,9, pp.

611-629. (2018).

Alzubaidi, L., Zhang, J., Humaidi, A.J., Al‑Dujaili, A.,

Duan, Y., Al‑Shamma, O., Santamaria, J., Fadhel,

M.A., Al‑Amidie, M., Farhan, L.: Review of deep

learning: concepts, CNN architectures, challenges,

applications, future Directions, Journal of Big Data,

8(23), 1-74. (2021).

Silva, N., Shah, V., Soares, J. Rodrigues, H.: Road

Anomalies Detection System Evaluation, Sensors, 18,

1-20. (2018).

Santhosh, K.K., Dogra, D.P., Roy, P.P.: Anomaly Detection

in Road Traffic Using Visual Surveillance: A Survey,

ACM Computing Surveys, 6(53), 1-26. (2020).

Mishra, A., Cohen, A., Reichherzer, T., Wilde, T.:

Detection of data anomalies at the edge of pervasive IoT

Systems, Computing, Springer, (2021).

Aradi, S.: Survey of Deep Reinforcement Learning for

Motion Planning of Autonomous Vehicles, IEEE

Transactions On Intelligent Transportation Systems,

Vol. 23, No. 2, 740-759. (2022).

Berman, D.S., Buczak, A.L., Chavis, G.S. Corbett, C.L. A

Survey of Deep Learning Methods for Cyber Security,

Information, 10, 122. 1-35. (2019). doi:10.3390/info10

040122

Nguyen, H., Kieu, L.-M., Wen, T., Cai, C.: Deep learning

methods in transportation domain: a review, IET

Intelligent Transport System, 12, 9.998-1004. (2018).

Alrajhi M., Kamel, M.: A Deep-Learning Model for

Predicting and Visualizing the Risk of Road Traffic

Accidents in Saudi Arabia: A Tutorial Approach,

International Journal of Advanced Computer Science

and Applications, Vol. 10, No. 11, 475-483. (2019).

Wang, J., Cehn, R., He, Z.: Traffic speed prediction for

urban transportation network: A path based deep

Naturalistic Driving Studies Data Analysis Based on a Convolutional Neural Network

255

learning approach, Transportation Research Part C, 100.

372-385. (2019).

Raiyn, J.: Developing Vehicle Locations Strategy on Urban

Road, Transport and Telecommunication, 18. (4). 253–

262. (2017).

NIknam, S., Natarajan, B., Barazidh, R.: Interference

Analysis for Finit-Area 5G mm Networks Considering

Blockage Effects, IEEE Access, Vol. 6. 23470-23479.

(2018).

Mozo, A., Ordozgoiti, B., GoÂmez-Canaval, S.:

Forecasting short-term data center network traffic load

with convolutional neural networks, PLoS ONE 13(2):

1-31. (2018). e0191939. https://doi.org/10.1371/

journal.pone.0191939.

Ensafi, Y., Amin, S.H., Zhang, G., Shah, B.: Time-series

forecasting of seasonal items sales using machine

learning –A comparative analysis, International Journal

of Information Management Data Insights,2, 1-16.

(2022).

Cordeiro, J.R., Raimundo, A., Postolache, O., Sebastião, P.:

Neural Architecture Search for 1D CNNs—Different

Approaches Tests and Measurements, Sensors, 21. 1-

23. (2021). https://doi.org/10.3390/s21237990.

Raiyn, J. Using Dynamic Market- Based Control for Real-

Time Intelligent Speed Adaptation Road Networks,

Advances in Science, Technology and Engineering

Systems Journal, Vol. 7. No. 4. 24-27.

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

256