Content uploaded by Adrian Bosire
Author content
All content in this area was uploaded by Adrian Bosire on Nov 29, 2022
Content may be subject to copyright.
XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
Using Deep Analysis of Driver Behavior for Vehicle
Theft Detection and Recovery
Adrian Bosire. Student Member, IEEE
Computer Science Department
Kiriri Womens University of Science and Technology
Kasarani, Kenya
bosire.adrian@kwust.ac.ke
Damian Maingi
Department of Mathematics
Sultan Qaboos University
Muscat, Oman
dmaingi@squ.edu.om
Abstract—Advancement in technology has resulted to a vast
amount of spatio-temporal data. When this data is properly
mined it translates to knowledge, which is used for auto-theft
detection and recovery. In this study, we analyze vehicle driving
dataset using deep learning algorithms such as the
Convolutional Neural Network, the Recurrent Neural Network
with Long Short-Term Memory, Deep Boltzmann Machines and
Deep Autoencoders, as well as bio-inspired algorithms such as
the Particle Swarm Optimization, Artificial Bee Colony, Ant
Colony Optimization and Bat Algorithm. The performance of
these algorithms is analyzed using benchmark functions such as
the Ackley function, Rastrigin function, Rosenbrock function,
Sphere function, Schaffer function and Himmelblau’s function.
Finally, efficiency of the algorithms was measured using the
Mean Squared Error.
Keywords—Deep neural networks, drivers driving signature,
driver’s behavior, machine learning, vehicular transportation
system
I. INTRODUCTION
The transport industry is essential in the growth of an
economy because it facilitates resource sharing, mainly
through communication and the transfer of goods and
services. The popular means of transport is motor vehicles and
this is marked by the elaborate road network infrastructure in
modern society. Furthermore, as technology advances so does
innovation which necessitates the advancement in the
automation of automobiles with an aim of improving the
transportation process. In order to reduce errors, which may
lead to accidents among other problems, vehicles have been
automated and also led to other inventions like autonomous
vehicles [1], [2]. These next generation technologies are also
associated with vices and crimes such as auto-theft. As such,
a need arises to curb the risks with practical counter-measures.
Among these mitigating options, artificial intelligence
presents itself as a viable option that can be applied in the
monitoring, tracking, detection and recovery of stolen
vehicles [3], [4].
In regard to the suitability of the algorithms present in
various categories, the performance of algorithms based on
deep learning supersedes the statistical-based machine
learning algorithms because of their meta-heuristic properties
[5], [6]. This means that optimality, completeness, precision
and speed are guaranteed via carefully adjusted parameters
and hyperparameters present in the algorithms to assure
confidence in the obtained results [7].
This article is organized in the following order: The next
section covers related work in the use of deep learning for the
automobile industry. The section that follows is the
methodology employed in performing the experiments for this
study. Thereafter, sections four and five present the deep
neural networks and swarm intelligence algorithms used in the
experiments. Then section six discusses the attained results
followed by a conclusion in the final section.
II. RELATED WORK
The approaches used in previous related studies primarily
rely on physical features of the stolen motor vehicle during
their tracking, detection and recovery. Such approaches
mostly rely on the use of license plate recognition, Global
Position System (GPS) tracking devices, sensors and other
telecommunication devices fitted into the vehicle. These
devices are used for geolocation and transmission of the
vehicles position. However, this approach is relatively reliable
until such devices are disabled, tampered with or even
removed [8], [9].
Additionally, novel approaches have been introduced
which cover a broad spectrum from biometrics, driving
patterns as well as driver behavior analysis [10]-[17].
However, the focus has been in areas such as accident
prevention, drivers’ intent prediction and monitoring of the
vehicle, it’s path or the driver’s behavior [18]-[21]. The other
authors who have ventured into auto-theft detection have used
some of the previous methods mentioned earlier which are
reliant on the physical characteristics of the vehicle [22].
Deep analytics, which refers to the use of sophisticated
data processing techniques to yield information, has also been
applied in other studies relating to automobile theft. The
datasets may include both unstructured and semi-structured
data from multiple sources [23], [24]. In this study, deep
analytics is applied on the analysis of the driver’s driving
behavior, consequently yielding the driver’s driving signature.
Monitoring of the driver’s driving signature would help track
a vehicle in transit and also detect a case of its theft.
III. METHODOLOGY
The objective of this study is to evaluate several deep learning
algorithms and swarm intelligence algorithms, described in
the next section, on the driver’s driving behavior pattern.
Generally, the steps followed in the implementation of the
algorithms and subsequent analysis of the Vehicular trace
dataset [25] is shown below.
i. Load the dataset
ii. Pre-process the data
iii. Define the algorithm
iv. Configure the hyperparameters and parameters
v. Define the activation function
vi. Define the Loss/Cost function
vii. Train the algorithm
viii. Optimize the network obtained
ix. Test the results obtained
2021 22nd International Arab Conference on Information Technology (ACIT) | 978-1-6654-1995-6/21/$31.00 ©2021 IEEE | DOI: 10.1109/ACIT53391.2021.9677433
A well-prepared dataset will enhance the likelihood of
obtaining better accuracy. The other configurations of the
algorithm may be done stepwise as the overall reaction is
observed so as to correctly analyze the dataset and obtain
optimal results.
This research proposes the analysis of the driver’s driving
style in conjunction with the GPS data as a signature to
monitor and track a vehicle and also detect its loss and
recovery. The analysis of the driver’s behavior will be done
by the deep neural networks and swarm intelligence
algorithms whose performance will be assessed based on the
Mean Squared Error obtained from the execution of
benchmark functions: Ackley function, Rastrigin function,
Rosenbrock function, Sphere function, Schaffer function and
Himmelblau’s function. The results obtained from these
algorithms will be compared in order to instill the confidence
needed to guarantee the quality in terms of accuracy.
IV. DEEP NEURAL NETWORK ALGORITHMS
The deep learning algorithms under consideration are the
Convolutional Neural Network (CNN), the Recurrent Neural
Network with Long Short-Term Memory (RNN-LSTM),
Deep Boltzmann Machines (DBM) and Deep Autoencoders
(AE).
The Convolutional Neural Network (CNN) is actually
popular in the analysis of images and by extension computer
vision. In this case, this algorithm is used to systematically
evaluate time series data. The algorithm follows the procedure
in which the data’s features are extracted using feature
detectors to create feature maps which are then reduced
through sum pooling. This is followed by flattening of the
pooled feature maps which acts as the input layer for the
Artificial Neural Network with several fully connected hidden
layers. The algorithm for the Convolutional Neural Network
is as follows [6], [26]:
Algorithm 1. Convolutional Neural Network
Input:
Initialize weights to a small randomly generated value, set
learning rate to a small positive value, training period and
the number of layers.
Output:
Begin with iteration
For n<MaxIteration, do
Forward propagate through convolution, pooling and
then fully-connected layers
Derive Loss Function value for the data
Calculate the error term with respect to the weights for
each type of layer
Backpropagate the error generated and calculate the
change in gradient for both the weights wk
(layer) and
bias bk
(layer)respectively
Update the weights and bias respectively
End For
The RNN-LSTM algorithm can process the current input
with respect to previously memorized input. This information
flow through the network is controlled using the input gate ,
forget gate and output gate . These three gates constitute
the memory cell which determines how much information to
propagate through the network and which data should be
discarded. The algorithm of the RNN-LSTM is shown in
algorithm 2 below [27].
Algorithm 2. Recurrent Neural Network with LSTM
Input:
Input sequence xi, training period T, learning rate ϵ, hidden
layer h and output sequence y.
Output:
Calculate the hidden vector sequence
For t=1: T do,
At the input gate, calculate the effective data using
it=σ (wxixt+ whi ht-1+ wcict-1 + bi) where σ is the
logistic sigmoid function
At the forget gate, determine the data to retain using
ft=σ (wxfxt+ whfht-1 + wcfct-1 + bf)
Calculate the memory cell state using ct=σct-1 +
ictan h (wxcxt+ whc ht-1 + bc)
Determine the output to be passed out the memory cell
using Ot=ft (wxoxt+ whoht-1+ wcoct-1 + bo)
Then calculate the effective output of the hidden layer
using ht=Ottan h (ct)
End for
Calculate the predicted output using yt=ft (whyht+by)
The Deep Boltzmann Machine algorithm utilizes the
global optimization capability of simulated annealing. The
visible input neurons are clamped onto specific states while
the visible output neurons and the hidden neurons operate
freely. Boltzmann machines learn their weights using
simulated annealing. The correlation-based learning
procedure of the DBM is presented in algorithm 4 below [28].
Algorithm 3. Deep Boltzmann Machine
Input:
Initialize as uniform random values in -a0 , a0, where
a0=0.5 or 1.
Set the initial temperature T0 and the final temperatures Tf
Output:
At the clamping phase, present the pattern and for each
example , perform simulated annealing until Tf is reached.
i. At each T, relax the network by the Boltzmann
distribution for a length of time through updating the
states of the unclamped (hidden) units
xi= +1, with probability Pi
-1, with probability 1-Pi
Where Pi is calculated using;
Pi=1
1+ e-neti
T
and neti=wi
J
j=1, j≠ix
ii. Then update T by the annealing schedule and at Tf
estimate the correlation in the clamped condition using;
Pij
+=Exi xj where i,j=1,2,…,J;i≠j
iii. At the free-running phase, only the input neurons are
clamped and the output neurons are free-running hence
at Tf estimate the correlation in the free-running
condition using;
Pij
-=Exi xj where i,j=1,2,…,J;i≠j
iv. Perform the weight update using;
∆wij=ηPij
+- Pij
- where i,j=1,2,…,J;i≠j
Where η= ε
T , with ε being a small positive constant
Repeat the steps above for next epoch until there is no change
in wij, i, j
The Deep Autoencoder makes use of several stacked
hidden layers that force the data to converge hence retain only
the optimal data that is suitable for predictions. Furthermore,
the non-linear hidden layers allow the network to learn more
complex encoding functions and enhance higher precision.
The following is the deep autoencoder algorithm [29].
Algorithm 4. Deep Autoencoder Neural Network
Input:
Input matrix A ϵ 0, 1m x n , where rows and columns
correspond to vector values for input features
Output:
Fix a number of hidden units (h ϵ N, h<m), and
a number of hidden layers (d ϵ {1, …, maxhl}
Training: for each driver profile ai of A, where i ϵ [1,m]:
i. for each hidden layer compute hidden activation
from the input vectors
ii. compute reconstructed output from hidden
activation
iii. Perform the stochastic gradient descent and compute
the MSE between ai and ai
iv. back-propagate error gradient to update weight
parameters
Testing: for each driver profile ai of , where i ϵ [1,m]:
i. autoencode ai and produce ai
ii. set ai as ithrow of the output matrix A
.
V. SWARM INTELLIGENCE ALGORITHMS
This study uses the following swarm intelligence
algorithms; Particle Swarm Optimization (PSO), Artificial
Bee Colony (ABC), Ant Colony Optimization (ACO) and Bat
Algorithm (BA).
The PSO algorithm mimics bird flocking and predation.
Thus, the particles represent birds in a search space. The
particles will fly in restricted directions of the bounded search
space which the group perceives to be ideal. Their velocities
are dynamically adjusted based on both their individual
experience and that of the other particles in the population.
Adjustment of the inertia weight facilitates both the global and
local search capability of the algorithm during its
implementation [30].
Algorithm 5. Particle Swarm Optimization
Initialization:
For each particle i=1, …, Np , do
i. Initialize the particle’s position with a uniformly
distribution as Pi0 ~ U LB, UB, where LB and UB
represent the lower
ii. and upper bounds of the search space
iii. Initializeto its initial position: pbest i, 0=
Pi0.
iv. Initialize to the minimal value of the swarm:
gbest 0= argminf[Pi(0)].
v. Initialize velocity: Vi ~ U -UB-LB, |UB-LB|.
Termination:
Repeat until a termination criterion is met
For each particle i=1, …, Np, do
i. Pick random numbers: ri, r2 ~ U(0, 1)
ii. Update particle’s velocity using Vi (t + 1) = ωVi (t) +
c1r1 pbest i, t- pi t+ c2r2 (gbest (t) - Pi (t))
iii. Update particle’s position using Pi (t + 1) = Pi (t) +
Vi (t + 1).
iv. If f[Pit] < f[pbest(i, t)], do
a. Update the best-known position of particle
i: pbesti, t= pi(t).
b. If f[Pit] < f[gbest(i, t), update the swarm’s best-
known position:gbest(t) = pi(t).
v. t←(t+1);
Output gbest(t) that holds the optimal solution.
The ABC algorithm is dependent on the foraging patterns
of the honey bees in nature. The bees are split into groups
based on their responsibilities. The unemployed onlooker bees
will wait for the employed bees who will use the waggle dance
to communicate to them about the viability of the food source
based on the quality and quantity of nectar. These bees will
then be recruited to that food source or alternatively start
scouting for potential food sources. The food sources in this
case are the plausible solutions available in the search space.
The quantity and quality of these solutions is measured using
fitness and probabilistic functions respectively as shown in the
algorithm steps below [31].
Algorithm 6. Artificial Bee Colony
Initialization:
For each bee i=1, …, n, do
i. Initialize the positions of bees (xi=1, …, SN)
ii. Randomly initialize the food sources within the search
space using xij=xj
min+rand(0,1)(xj
max-xj
min) where
xij represents the parameter for ith employed bee on jth
dimension
iii. Evaluate fitness fiti of the population
Termination:
Repeat until (iter ≤MaxCycles)
i. Generate new positions which represent new solutions
vij by the employed bees and calculate the fitness
value fiti on the new population
ii. Apply greedy selection process between the initial
solutions xij and the resultant solutions vij.
iii. Calculate the probability values (Pi) for the solutions
xi
iv. Generate other new solutions vi for the onlookers
from the solutions xi selected depending on their
(Pi) values and evaluate the nectar quality of new
positions using fiti.
v. Apply greedy selection process to solutions found by
onlooker bees.
vi. If there is an abandoned solution for the scout then
replace it with a new random solution xi
vii. Memorize the best solution so far xbest←xi||vi.
viii. Increment iteration iter+1
Output xbest the best memorized solution.
The ACO algorithm is based on the foraging behavior of
ants which start by randomly exploring the area surrounding
their nest. The ant that finds a worthwhile food source will
evaluate the quality and quantity of the food which determines
the rate of pheromone it will deposit on its trail back to the
nest. The rest of the ants evaluate the viability and shortest
route to the source of the food using the pheromones deposited
on the trail as a means of communication. Therefore, the
essence of the ACO approach is the pheromone model which
uses a probability function to evaluate the probable solutions
in each iteration measured by pheromones deposited and
further quantified using the evaporation rate. The algorithm is
presented below [32].
Algorithm 7. Ant Colony Optimization Algorithm
Initialization:
For each ant j=1, …, na do
i. Initialize pheromones trail (T0)
ii. Assign best solution Sbsat any time in each
iteration t
iii. Assign the minimum Tmin and maximum Tmax value
for the pheromone trail
iv. Store the best solution in each iteration Sib
Termination:
Repeat until a termination criterion is met
For j=1, …, na do
i. Construct a solution Sib using a probabilistic
function
ii. If Sib is a valid solution then optionally perform a local
search to find the best solution Sbs
iii. If fSis, t< fSbs , t or Sbs=NULL then Sbs ← Sib
iv. Update the pheromone bounds Tmin, Tmax using
Tij← 1- ρ Tij, i, j where ρ is the evaporation
rate of the pheromones which satisfies 0< ρ≤1
v. If (stagnation behavior detected) then initialize
pheromones trail (Tmax)
vi. Update the iteration t←(t+1);
Output Sbs which is considered the optimal solution
The Bat algorithm is another metaheuristic optimization
algorithm based on the echolocation technique of microbats
which they use for detecting prey, avoiding obstacles and
locating their roosting crevices. These bats emit short bursts
of sound pulses whose echo is detected by the bats and
together with the time delay, detection difference between the
ears and pitch of the echo to build a three dimensional view of
their surroundings. Therefore, the bat algorithm relies on the
way bats fly randomly from their crevices at a given velocity
and emit sound pulses at an adjustable frequency to locate
their food. The bat algorithm is presented below [33].
Algorithm 8. Bat Algorithm
Initialization:
For each bat i=1, n, do
i. Initialize the bat positions xi (i = 1, …, n) and velocity
vi
ii. Define the pulse frequency fi at xi
iii. Initialize pulse rates r and the loudness A
Termination:
Repeat until (t < tmax)
i. Generate new solutions by;
a. adjusting frequency fi=fmin+(fmax-fmin)β, where
β is a random vector in the range of [0,1]
b. updating velocities vi
t=vi
t-1+(xi
t-1-xbest )fi
c. update locations xi
t=xi
t-1 + vi
t
ii. If (rand>r), do
a. Select a solution among the best solutions
b. Generate a local solution around the selected best
solution
iii. Generate a new solution by flying randomly
iv. If (rand < A & f(xi) < f(xbest )) then accept the new
solutions
v. Rank the bats and find the current best xbest
Output the optimal solution (xbest ).
VI. EXPERIMENTS AND RESULTS
The deep neural network and swarm intelligence
algorithms are implemented in python programming
language. The parameters and hyperparameters of the
algorithms are configured and consistently adjusted on several
runs based on performance. The performance of the
algorithms is evaluated by using the mean squared error
(MSE) defined in (1) below.
MSE= yi- yi2
N
i = 1
N(1)
where is the predicted value and is the actual value.
The MSE is the average squared difference between the
predicted values and the actual values.
The dataset is split into three divisions; 70% for training,
20% for validation and the other 10% for testing. The idea
behind splitting the dataset is to compare the results obtained
with the original values. This makes it easy to verify the
accuracy of the respective algorithms as shown in Table I and
Table II.
TABLE I THE MEAN SQUARED ERROR (MSE) OBTAINED WHEN PERFORMING DRIVER BEHAVIOR ANALYSIS USING DEEP NEURAL NETWORKS.
Loss/Cost Function
Convolutional
Neural Network
(CNN)
Recurrent Neural
Network with Long
Short-Term Memory
(RNN-LSTM)
Deep Boltzmann
Machines (DBM)
Deep Autoencoders
(AE)
f1
Ackley Function
1.7329
0.4578
0.4101
1.7305
f2
Rastrigin Function
1.6521
1.5673
1.02947
0.6958
f3
Rosenbrock Function
1.7952
0.1203
1.0592
1.3942
f4
Sphere Function
2.0178
1.2995
2.3061
3.9148
f5
Schaffer Function
1.1262
1.4189
2.2979
2.7427
f6
Himmelblau’s Function
2.4911
2.2194
3.9647
3.4031
TABLE II THE MEAN SQUARED ERROR (MSE) OBTAINED WHEN PERFORMING DRIVER BEHAVIOR ANALYSIS USING BIO-INSPIRED ALGORITHMS.
Cost Function
Particle Swarm
Optimization (PSO)
Artificial Bee Colony
(ABC)
Ant Colony
Optimization (ACO)
Bat Algorithm (BA)
f1
Ackley Function
0.7741
1.5038
1.2596
2.1733
f2
Rastrigin Function
2.4205
1.7396
2.6681
2.6908
f3
Rosenbrock Function
1.7103
1.8529
3.7305
2.5842
f4
Sphere Function
2.3059
0.4099
0.9426
3.2227
f5
Schaffer Function
1.9904.
1.0733
0.4471
2.6932
f6
Himmelblau’s Function
1.5390
2.5412
4.7841
4.6117
During the parameter tuning, results also show that the
depth of the deep neural networks in terms of the number of
hidden layers had a relatively minimal influence on the
predictability accuracy. This is because increasing the number
of hidden layers past 50 slowed down the network and this
had a minimal increase in relation to the accurate predicted
values.
Furthermore, the breadth of the network in terms of the
number of neurons per layer had a correlation with the number
of features that were being observed in the dataset. This is
because increasing the number of neurons past 20 did not
make a significant change in the MSE. The RNN-LSTM
performed relatively better than the other algorithms.
Generally, increasing the breadth and depth of the network
had a relatively small impact on the performance in terms of
classification and matching of the drivers with their driving
pattern. However, there was some dependence on the features
being observed. This means that observing fewer features for
a larger dataset increased the confidence levels for the
predictability compared to many features for the same large
dataset.
The ABC algorithm and the PSO algorithm also made
good predictions although the bio-inspired algorithms made
significant changes when the parameters were slightly
changed. This is unlike the deep neural networks which were
dependent on the hyperparameter changes in order to lower
the errors in predictions.
Overall, there was a high dependence of hyperparameter
and parameter tuning to feature selection to achieve accurate
prediction by the algorithms.
VII. CONCLUSION
The ultimate objective of this study was to use the results
as a means of specifically identifying a driver based on their
driving pattern and using this information to detect and
possibly recover a stolen vehicle. The results obtained show
that this is possible and presents possibilities for further
research.
The analysis of driving behavior provides insight to the
vehicles design, safety of the occupants, quality and
maintenance. Therefore, there are several aspects of any
driver’s behavior that can be studied in order to help us
understand how we can improve the quality of driving and
assure safety.
Furthermore, the driver’s driving signature is dependent
upon their experience and skill as exhibited by their reactions
over time. This is dynamic owing to the fact that their
reactions may change, more so, when other extenuating
factors are involved such as the weather, road condition or the
kind of occupants present in the vehicle.
ACKNOWLEDGMENT
This work was supported by Kiriri Women’s University of
Science and Technology.
REFERENCES
[1]
Z. Dong, W. Shi, G. Tong and K. Yang, "Collaborative
Autonomous Driving: Vision and Challenges," in 2020
International Conference on Connected and Autonomous Driving
(MetroCAD), Detroit, MI, USA, 2020.
[2]
J. Rodge and S. Jaiswal, ""Comprehensive Overview of Neural
Networks and its Applications in Autonomous Vehicles."," in
Computational Intelligence in the Internet of Things, IGI Global,
2019, pp. 159-173.
[3]
N. O. Alsrehin, A. F. Klaib and A. Magableh, "Intelligent
Transportation and Control Systems Using Data Mining and
Machine Learning Techniques: A Comprehensive Study," IEEE
Access, vol. 7, pp. 49830-49857, 2019.
[4]
J. Li, H. Cheng, H. Guo and S. Qiu, "Survey on Artificial
Intelligence for Vehicles," Automotive Innovation, vol. 1, no. 1,
pp. 2-14, 19 March 2018.
[5]
R. Chalapathy and S. Chawla, "Deep Learning for Anomaly
Detection: A Survey," arXiv:1901.03407v2 [cs.LG], 2019.
[6]
A. Shrestha and A. Mahmood, "Review of Deep Learning
Algorithms and Architectures," IEEE Access, vol. 7, pp. 53040-
53065, 2019.
[7]
A. Koutsoukas, K. . J. Monaghan, . X. Li and J. Huan, "Deep-
learning: investigating deep neural networks hyper-parameters
and comparison of performance to shallow methods for modeling
bioactivity data," Journal of Cheminformatics, vol. 9, no. 42,
2017.
[8]
M. Alhussein, K. Aurangzeb and S. I. Haider, "Vehicle License
Plate Detection and Perspective Rectification," Elektronika Ir
Elektrotechnika, vol. 25, no. 5, pp. 47-56, 2019.
[9]
S. Singh and P. Kumari, "Automatic Car Theft Detection System
Based on GPS and GSM Technology," International Journal of
Trend in Scientific Research and Development, pp. 689-692, 2019.
[10]
A. A. Elngar and M. Kayed, "Vehicle Security Systems using Face
Recognition based on Internet of Things," Open COmputer
Science, vol. 10, no. 1, pp. 17-29, March 2020.
[11]
M. Ramesh, S. Akruthi, K. Nandhini and S. Meena,
"Implementation of Vehicle Security System using GPS,GSM and
Biometric," in 2019 Women Institute of Technology Conference
on Electrical and Computer Engineering (WITCON ECE),
Dehradun Uttarakhand, India, 2019.
[12]
K. Aishwarya and R. Manjesh , "A Novel Technique for Vehicle
Theft Detection System Using MQTT on IoT," in Lecture Notes
in Electrical Engineering, 2020.
[13]
Y. G. Kang, K. H. Park and H. K. Kim, "Automobile Theft
Detection by Clustering Owner Driver Data," in 17th escar Europe
: embedded security in cars, 2019.
[14]
A. Girma, X. Yan and A. Homaifar, "Driver Identification Based
on Vehicle Telematics Data using LSTM-Recurrent Neural
Network," in 2019 IEEE 31st International Conference on Tools
with Artificial Intelligence (ICTAI), Portland, OR, USA, 2019.
[15]
W. Wang, J. Xi and D. Zhao, "Driving Style Analysis Using
Primitive Driving Patterns With Bayesian Nonparametric
Approaches," IEEE Transactions on Intelligent Transportation
Systems, vol. 20, no. 8, pp. 2986-2998, 2019.
[16]
F. Martinelli, F. Mercaldo, A. Orlando, V. Nardone, A. Santone
and A. K. Sangaiah, "Human Behavior Characterization for
Driving Style Recognition in Vehicle System," Computers and
Electrical Engineering Journal, vol. 83, 29 January 2018.
[17]
M. Villa, M. Gofman and S. Mitra, "Survey of Biometric
Techniques for Automotive Applications," in Information
Technology - New Generations. Advances in Intelligent Systems
and Computing, vol. 738, S. Latifi, Ed., Springer, Cham, 2018, pp.
475-481.
[18]
J. Bhalerao, A. Kadam, A. Shinde and V. Mugalikar, "Proposed
Design on Driver Behavioral Analysis," 2020.
[19]
J. Zhang, Z. C. Wu, F. Li, C. Xie, T. Ren, J. Chen and L. Liu, "A
Deep Learning Framework for Driving Behavior Identification on
In-Vehicle CAN-BUS Sensor Data," Sensors, vol. 19, no. 6, pp. 1-
17, March 2019.
[20]
M. L. Bernardi, M. Cimitile, F. Martinelli and F. Mercaldo,
"Driver and Path Detection through Time-Series Classification,"
Journal of Advanced Transportation, vol. 2018, pp. 1-20, March
2018.
[21]
K. Zinebi, N. Souissi and K. Tikito, "Driver Behavior Analysis
Methods: Applications oriented study," in The 3rd International
Conference on Big Data, Cloud and Applications - BDCA'18,
2018.
[22]
R. Kommaraju, R. Kommanduri, S. R. Lingeswararao, B.
Sravanthi and C. Srivalli, "IoT Based Vehicle (Car) Theft
Detection," in International Conference on Image Processing and
Capsule Networks. ICIPCN 2020, 2020.
[23]
H. K. Bhuyan and S. K. Pani , "Crime Predictive Model Using Big
Data Analytics," in Intelligent Data Analytics for Terror Threat
Prediction: Architectures, Methodologies, Techniques and
Applications, S. K. Pani, S. K. Singh, L. Garg, R. B. Pachori and
X. Zhang, Eds., Scrivener Publishing LLC , 2021, pp. 57-78.
[24]
M. Feng, J. Zhen, J. Ren, A. Hussein, X. Li, Y. Xi and Q. Liu, "Big
Data Analytics and Mining for Effective Visualization and Trends
Forecasting of Crime Data," IEEE Access, vol. 7, no. 2019, pp.
106111-106123, August 2019.
[25]
P. H. Rettore, "Vehicular Trace Data," 22 January 2018. [Online].
Available: http://www.rettore.com.br/prof/vehicular-trace/.
[26]
A. Khan, A. Sohail, U. Zahoora and A. S. Qureshi , "A survey of
the recent architectures of deep convolutional neural networks,"
Artificial Intelligence Review, April 2020.
[27]
A. Sherstinsky, "Fundamentals of Recurrent Neural Network
(RNN) and Long Short-Term Memory (LSTM) Network,"
Physica D: Nonlinear Phenomena, vol. 404, pp. 1-43, March
2020.
[28]
K. Du and M. Swamy, "Probabilistic and Bayesian Networks," in
Neural Networks and Statistical Learning, 1 ed., London,
Springer, 2019, pp. 589-649.
[29]
D. Chicco, P. Sadowski and P. Baldi, "Deep Autoencoder Neural
Networks for Gene Ontology Annotation Predictions," in The 5th
ACM Conference on Bioinformatics, Computational Biology, and
Health Informatics, New York, 2014.
[30]
D. Gao, X. Li and H. Chen, "Application of Improved Particle
Swarm Optimization in," Mathematical Problems in Engineering,
vol. 2019, pp. 1-10, March 2019.
[31]
H. Hakli and M. S. Kiran, "An improved artificial bee colony
algorithm for balancing local and global search behaviors in
continuous optimization," International Journal of Machine
Learning and Cybernetics, vol. 11, p. 2051–2076, 26 February
2020.
[32]
M. Mavrovouniotis, F. M. Müller and S. Yang, "Ant Colony
Optimization With Local Search for Dynamic Traveling Salesman
Problems," IEEE Transactions on Cybernetics, vol. 47, no. 7, pp.
1743-1756, July 2016.
[33]
W. H. Bangyal, J. Ahmad, H. T. Rauf and S. Pervaiz, "An
Improved Bat Algorithm based on Novel Initialization Technique
for Global Optimization Problem," International Journal of
Advanced Computer Science and Applications (IJACSA), vol. 9,
no. 7, pp. 158-166, 2018.
