A robust machine learning structure for driving events recognition using smartphone motion sensors

Abstract

Driving behavior monitoring by smartphone sensors is one of the most investigated approaches to ameliorate road safety. Various methods are adopted in the literature; however, to the best of our knowledge, their robustness to the prediction of new unseen data from different drivers and road conditions is not explored. In this paper, a two-phase Machine Learning (ML) method with taking advantage of high-pass, low-pass, and wavelet filters is developed to detect driving brakes and turns. In the first phase, accelerometer and gyroscope filtered time series are fed into Random Forest and Artificial Neural Network classifiers, and the suspicious intervals are extracted by a high recall. Following that, in the next phase, statistical features calculated based on the obtained intervals are used to determine the false and true positive events. To compare the predicted and real labels of the recorded events and calculate the accuracy, a method that covers the limitations of previous sliding windows is also employed. Real-world experimental result shows that the proposed method can predict new unseen datasets with average F1-scores of 71% in brake detection and 82% in turn detection which is comparable with previous works. Moreover, by sensitivity analysis of our proposed model, it is proven that implementing high-pass and low-pass filters can affect the accuracy for turn detection up to 30%.

Full text

Full Terms & Conditions of access and use can be found at
https://www.tandfonline.com/action/journalInformation?journalCode=gits20
Journal of Intelligent Transportation Systems
Technology, Planning, and Operations
ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/gits20
A robust machine learning structure for driving
events recognition using smartphone motion
sensors
Mahdi Zarei Yazd, Iman Taheri Sarteshnizi, Amir Samimi & Majid Sarvi
To cite this article: Mahdi Zarei Yazd, Iman Taheri Sarteshnizi, Amir Samimi & Majid Sarvi (2022):
A robust machine learning structure for driving events recognition using smartphone motion
sensors, Journal of Intelligent Transportation Systems, DOI: 10.1080/15472450.2022.2101109
To link to this article: https://doi.org/10.1080/15472450.2022.2101109
Published online: 24 Jul 2022.
Submit your article to this journal
View related articles
View Crossmark data

A robust machine learning structure for driving events recognition using
smartphone motion sensors
Mahdi Zarei Yazd
a
, Iman Taheri Sarteshnizi
b
, Amir Samimi
c
, and Majid Sarvi
b
a
Department of Civil Engineering, Sharif University of Technology, Tehran, Iran;
b
Department of Infrastructure Engineering, University
of Melbourne, Melbourne, Australia;
c
School of Civil Engineering, University of Sydney, Sydney, Australia
ABSTRACT
Driving behavior monitoring by smartphone sensors is one of the most investigated
approaches to ameliorate road safety. Various methods are adopted in the literature; how-
ever, to the best of our knowledge, their robustness to the prediction of new unseen data
from different drivers and road conditions is not explored. In this paper, a two-phase
Machine Learning (ML) method with taking advantage of high-pass, low-pass, and wavelet
filters is developed to detect driving brakes and turns. In the first phase, accelerometer and
gyroscope filtered time series are fed into Random Forest and Artificial Neural Network clas-
sifiers, and the suspicious intervals are extracted by a high recall. Following that, in the next
phase, statistical features calculated based on the obtained intervals are used to determine
the false and true positive events. To compare the predicted and real labels of the recorded
events and calculate the accuracy, a method that covers the limitations of previous sliding
windows is also employed. Real-world experimental result shows that the proposed method
can predict new unseen datasets with average F1-scores of 71% in brake detection and
82% in turn detection which is comparable with previous works. Moreover, by sensitivity
analysis of our proposed model, it is proven that implementing high-pass and low-pass fil-
ters can affect the accuracy for turn detection up to 30%.
ARTICLE HISTORY
Received 4 November 2021
Revised 4 July 2022
Accepted 10 July 2022
KEYWORDS
driving behavior; driving
monitoring; machine
learning; smartphone sensor
Introduction
Road accidents cost most countries approximately 3%
of their GDPs, and 1.35 million people die annually
because of these crashes. Among the main reasons,
speeding, drunk driving, and distracted driving are
evidenced to be the most contributors (World Health
Organization, 2018). Drivers tend to perform some
commonly known maneuvers, like braking and turn-
ing, more frequently in these conditions. If they are
prohibited from driving while being in such situations
that would considerably help avoiding road accidents.
Formerly, traffic policing was the case for the
researchers to improve driving safety (Bates et al.,
2012). However, with the appearance of in-vehicle
sensors (Yuksel & Atmaca, 2021) and the outbreak of
smartphones, developing platforms for driving moni-
toring became a hotspot in this area (Siami et al.,
2021; Toledo et al., 2008). In addition to safety pur-
poses, driving style monitoring also contributes to less
fuel consumption and leads to eco-driving (Jamson
et al., 2015; Tanvir et al., 2021). Providing proper and
on-time feedback to the drivers proved to be effective
and this motivate them to pay more attention to their
performance while driving. Therefore, a precise and
robust method for this aim would lead to fewer daily
accidents and also fuel consumption.
Several methodologies are introduced and tested
formerly for driving behavior monitoring in the litera-
ture and excellent results are achieved using them
employing different datasets. Motion data of individ-
ual vehicles such as acceleration, velocity, angular vel-
ocity, and orientation data is recorded in these
datasets using OBD (On-Board Diagnostic) devices or
smartphones. The behavior of the drivers is also
recorded alongside the data collection using question-
naires or labeling some specific events to investigate
the validity of the methodologies (Chan et al., 2020;
Kazemeini et al., 2022). Despite the existence of such
efforts, the robustness of previous works has remained
an open area of research. A detection method is called
robust if its performance is not susceptible to a sig-
nificant decline when it comes to situations with
CONTACT Iman Taheri Sarteshnizi itaherisarte@student.unimelb.edu.au Department of Infrastructure Engineering, University of Melbourne,
Melbourne, Victoria, Australia
ß2022 Taylor & Francis Group, LLC
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS
https://doi.org/10.1080/15472450.2022.2101109

different characteristics. For instance, in driving
behavior monitoring using smartphones, the driver,
smartphone type, and route condition are some
parameters affecting the collected data of smartphone
motion sensors and changing these parameters may
affect the results of trained models. Analyzing the
robustness of the previous detection methods is over-
looked by the literature and in this paper, we devel-
oped a driving behavior detection method that
performs well even using a test set with different char-
acteristics compared to the training set. In the follow-
ing paragraphs, we deeply discuss the literature on
driving behavior detection and elaborate on our
method contributions.
Broadly, driving behavior recognition models fall
into two categories: unsupervised and supervised
models. The former types deal with a huge mass of
unlabeled data in which the ground truth of the tar-
gets (driving events) is not specified. For instance,
Eftekhari and Ghatee (2018) developed a system that
recognizes driving events by feeding Discrete Wavelet
Transformation (DWT) of the time series data into an
Adaptive Neuro-Fuzzy Inference System (ANFIS). Yao
et al. (2021) also benefited from DTW (Dynamic
Time Warping) and HMM (Hidden Markov Model)
to cluster driver behavior. Although these researchers
fully exploit the massive real-world data collected by
smartphones, the accuracy of their systems remains
open to doubt.
In the supervised structures, which is the case in
this paper, multiple instances of different driving
events are available in the collected data. Data collec-
tion for designing such models is a challenging pro-
cess and also labor-intensive; however, one can
quantify the prediction performance using the labeled
instances and employ it to compare the results in dif-
ferent scenarios. Methods implemented in this area
are divided into three distinct groups, namely thresh-
old-based, pattern matching-based, and learning-based
methods. Detailed information regarding some of the
main studies in this field is represented in Table 1.
Threshold-based methods such as Chhabra et al.
(2018) are simple and easy to implement; but very
case dependent. Pattern matching-based methods such
as Dynamic Time Warping (DTW) (Singh et al.,
2017) measure the similarity between two signals. In
these algorithms, one must define template signals
and the models’performances highly depend on the
selected templates (Chan et al., 2020).
Learning-based structures can learn and construct
predictive models exploiting a large amount of train-
ing data which enables them to find more complex
patterns than the other approaches. For example, Yu
et al. (2017) used Artificial Neural Network (ANN),
and Support Vector Machine (SVM) to classify six
types of driving events using a total of 4029 labeled
driving events. Bejani and Ghatee (2018) developed an
ensemble learning method containing a Decision Tree
(DT), SVM, ANN, and K-Nearest Neighbors (KNN)
to evaluate the different driving styles of 27 drivers.
Nuswantoro et al. (2020) also adopted an ANN algo-
rithm by sliding a five-second time window over the
collected data to specify different driving behaviors.
Furthermore, some other researchers like Zhang et al.
(2019), Wang et al. (2021), and Saleh et al. (2017)
designed deep learning algorithms for this aim. To
elaborate, Saleh et al. (2017) used Long-Short Term
Memory (LSTM) to recognize normal, aggressive, and
drowsy driving applying a 50 percent overlapped slid-
ing window. Zhang et al. (2019) also designed an
attention-based convolutional and recurrent neural
network to classify driving events like brakes and
turns. Deep learning algorithms are unveiled to be
more powerful than simple ML methods since they are
capable of capturing more complex time-related features
of multidimensional data taking advantage of their
complicated nonlinear structure (Shinde & Shah, 2018);
Table 1. Summary of the main previous studies.
Type Reference Method EE
1
method Events Testing approach
Threshold-based (Chhabra et al., 2018) Fixed Threshold –Acc
2
, Brake, Turn Not mentioned
Pattern matching-based (Singh et al., 2017) DTW Using threshold LC
3
, Acc, Brake, Turn Not mentioned
Learning-Based (Bejani & Ghatee, 2018) DT Sliding window LC, Turn, U-Turn Cross-validation
(Carlos et al., 2020) Bag of Words þANN Sliding window Acc, Swerving, Brake Train/Test split
(Yu et al., 2017) ANN, SVM Not mentioned Weaving, Swerving,
Side slipping, U-
Turn, Turn, Brake
Train/Test split
(Xie et al., 2018)RF
4
Sliding window LC, Acc, Brake, Turn Train/Test split
(Ferreira et al., 2017) RF, ANN, SVM, BN
5
Sliding window LC, Acc, Turn, Brake Not mentioned
(Carvalho et al., 2017) RNN
7
Sliding window Acc, LC, Turn Train/Test split
(Saleh et al., 2017) RNN Sliding window Normal, Aggressive,
Drowsy Driving
Train/Test split
(Wang et al., 2021) CNN þRNN Separated events Acc, Brake, Turn Different datasets
(Zhang et al., 2019) CNN þRNN Sliding window Brake, Acc, Turn Cross-validation
(Zhang et al., 2019) CNN þRNN Sliding window Brake, Acc, Turn Cross-validation
2 M. ZAREI YAZD ET AL.

however,theyaremoreusefulwhenmassivedriving
events are labeled.
The event extraction method, data filtering, and
testing approach are three key factors that should be
more discussed among previous learning-based mod-
els. In terms of event extraction, some previous
works like Yu et al. (2017), Ma et al. (2019), and
Wang et al. (2021) collected separated driving events,
and their aims were just classifying, not recognizing,
different driving behaviors. They supposed that the
driving events are previously extracted from the raw
time series data of smartphones and their focus was
only on determining the type of these extracted driv-
ing events. Other studies by Eftekhari and Ghatee
(2019) and Carlos et al. (2020) took advantage of
sliding windows to split the whole data time series
and implement learning-based algorithms. Using a
sliding window is the case for most of the recent
papers, nevertheless, there are some drawbacks with
this method mentioned in the literature (Ouyang
et al., 2018). For instance, some driving events may
not be entirely captured by rolling a window over a
time series, and label assignment to the slices of a
time series becomes a controversial task while using
a supervised dataset.
Collected data by smartphone sensors comprise
some noise that may affect the result provided by
developed detection methods (Wu et al., 2018).
Different data filtering approaches are implemented in
the literature for smartphone sensor data denoising
namely simple moving average filter (Johnson &
Trivedi, 2011), band-pass filter (Singh et al., 2017),
low and high-pass filters (Chhabra et al., 2018), and
wavelet filter (Eftekhari & Ghatee, 2018). Previous
papers exploited these approaches in their analysis,
however, the impact of these denoising filters on
improving the accuracy is not reported.
From the perspective of the testing approach, most
of the conducted studies (Carvalho et al., 2017; Xie
et al., 2018) randomly divided their datasets into train
and test set to show the performance of the designed
models. Besides, some other studies (Nguyen et al.,
2020;S

anchez et al., 2018) benefited from cross-valid-
ation method to address the problem of overfitting. In
these studies, the number of recorded labeled events
per event type hardly hits 100 (Carvalho et al., 2017;
Daptardar et al., 2015; Eftekhari & Ghatee, 2019;
Ferreira et al., 2017; Nuswantoro et al., 2020; Xie
et al., 2018). The point here is that the predictive
models in the literature are not robust and examined
by rich train and test datasets. In other words, it is
not clear whether the former models still perform well
when it comes to other unseen drivers, mobile
phones, or routes. Some works are developing person-
alized driving behavior monitoring systems (Yi et al.,
2019) but a transferrable and robust model which can
be used for different drivers is not yet explored.
In brief, the literature on driving behavior detec-
tion using smartphone motion data can be improved
in three distinct ways namely: (1) providing a proper
event extraction method for supervised collected
datasets rather than using sliding windows, (2) exam-
ining the effect of data filtering approaches on detec-
tion accuracy, and (3) testing the reliability (or
robustness) of the results using different testing scen-
arios. To address the above-mentioned gaps, we
developed a two-phase ML structure for driving
behavior recognition and demonstrated the profi-
ciency and robustness of our model with different
testing scenarios. Furthermore, a supervised dataset
including 4193 braking and 1434 turning samples is
collected with more than 40 drivers and 12 different
smartphones to design multiple train/test scenarios.
The findings of this paper suggest that our proposed
model is practical especially in situations where a
new driver with a new smartphone joins a monitor-
ing system and it is necessary to detect his/her
behavior in the early stage without using his/her his-
torical data. The main contributions of this research
are highlighted below:
An event extraction phase utilizing Random Forest
and Multilayer Perceptron (ANN) classifiers along
with an algorithm for quantifying the performance
of this phase are proposed in this paper to cover
the limitations of the traditional sliding windows.
We trained and tested our proposed method with
multiple datasets completely different from each
other in terms of environmental properties to
explain the robustness of our method.
Despite other studies, the accuracy of our method
in recognizing braking and turning events is meas-
ured and discussed with a rich supervised dataset,
and the marginal contribution of different input
components in our method, specifically denoising
filters, are also quantified.
The rest of this paper is organized as follows. In
the next section, the data collection method is
described. In Method section, our structure for brake
and turn detection is demonstrated. In Result section,
results achieved by this research are shown and dis-
cussed. Finally, a summary of findings is highlighted
in Summary and conclusion section.
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 3

Data
To record the drivers’behavior, accelerometer and
gyroscope sensors embedded in different smartphones
are utilized. These two sensors record motion data of
smartphones in three different dimensions (x, y, and z
indicated in Figure 1(a)). The physical and gravita-
tional acceleration (m/s
2
) of the devices is collected by
the accelerometer and the angular velocity (rad/s)
relative to the three axes of the smartphones is stored
by the gyroscope. Since the coordination of the smart-
phone should be aligned with the vehicle direction,
every device must be affixed to the car during the
data collection. In our experiment, smartphones are
connected to the vehicles in a way that the z-axis is
toward the sky and the y axis is aligned with the dir-
ection of the car movement (Figure 1(b)). To make
sure that the smartphones are not affected by
unwanted movements inside the vehicles, we used
some equipment to strictly connect the smartphones
to the vehicles such as mobile holders or adhesive
tapes. Moreover, smartphones in our experiment were
placed mostly on the vehicle’s center console, over the
glove box (using a holder), or on the small ledge pro-
vided on car doors as these locations were mostly
used in the literature (Carlos et al., 2020; Carvalho
et al., 2017; Ferreira et al., 2017; Yu et al., 2017;
Zhang et al., 2019).
For conducting the data collection phase of our
research, 11 people were trained to label driving
events (brakes and turns) while seating next to the
driver during the trip. We asked the drivers to also
announce every brake and turn maneuver before per-
forming that to enhance the correctness of the real
recorded labels. An android-based application was
developed to simultaneously record the accelerometer
and gyroscope data along with the assigned labels
(Figure 2). Data collectors were supposed to tap a
brake or turn button provided in the application
when a driving event is about to take place and tap it
off when the maneuver is finished. Before recording
the data, passengers were able to select the desired
sensors (accelerometer and gyroscope in our case) for
data collection (Figure 2(a)).
All the experiments were executed in an urban
environment with 42 different drivers and 12 types of
smartphones (with 9 different brands). Nine of the par-
ticipated drivers were between 21 and 35, eight of
them were between 36 and 49, and the rest of them
were experienced taxi drivers working within the city
with more than 50 years of age. The reason behind
using these different characteristics is to test the robust-
ness of our proposed model and address the generaliz-
ability problem of the results which is overlooked by
the previous models in the literature. We have parti-
tioned our dataset according to the smartphone brands
into six groups. A description of the gathered data in
each group is demonstrated in Table 2.245tripsare
recorded with an average duration of 10 minutes
including 5627 driving events (brakes and turns).
Method
A two-phase ML-based structure is developed in this
paper to recognize brakes and turns during a trip
when smartphone sensory data is available. A general
picture of this method is illustrated in Figure 3. The
Event Extraction Phase (Phase 1) is designed to find
the potential driving event intervals. Since some out-
puts of the first phase may contain false-positive
events, the second phase is then created to better clas-
sify false and true positive intervals from the previous
phase. The filters and classifiers used in our method
as well as our proposed structure will be further dis-
cussed in this section.
Filters
Fourier-based Low-pass, high-pass, as well as wavelet
filters, are applied in our method to eliminate the
undesirable recorded noise by smartphones. Although
just a few studies in the area of driving behavior
Figure 1. (a) Smartphone coordination system, (b) Smartphone placement in the vehicle during the experiment.
4 M. ZAREI YAZD ET AL.

detection such as Eftekhari and Ghatee (2018) applied
them to smoothen the data of smartphones instead of
simple or exponential moving average (Yu et al.,
2017), these filters are widely used in different areas
related to time series analysis and showed promising
results over the basic filters (Haque et al., 2016;
Malghan & Hota, 2020). Therefore, we will follow the
way of Eftekhari and Ghatee (2018) and also include
Fourier-based high and low-pass filters and show that
their contributions in improving the detection accur-
acy are considerable.
The Fourier transform plays a crucial role in high
and low pass filters which is formulated as in (1):
^
ft
ðÞ¼X
n1
t¼0
ft
ðÞ
exp 2pipt
n
 (1)
where
^
ft
ðÞis the Fourier transform of the time series
ft
ðÞ
, p is the specified frequency (p ¼0, …, n-1) and
n is the number of observations in ft
ðÞ
:In a low-pass
filter, the frequency components of a time series are
first decomposed by the Fourier transform. Then, the
components with lower frequencies than the cutoff
frequency are passed, and the other ones are
neglected. With the usage of this filter, only the low-
frequency patterns of the time series remain and the
noises contributing to the high-frequency components
are smoothed. The performance of the high-pass fil-
ters is similar to the low-pass types except that,
reversely, in the high-pass filters, components with
higher frequencies than the cutoff are passed.
In wavelet filter, instead of Fourier transform, a
Discrete Wavelet Transform (DWT) is applied to sep-
arate different frequency parts of the time series. In
DWT, time resolution is also an important factor.
This means that the decomposed frequencies of a sig-
nal can be turned on or off depending on the time,
which is not the case in Fourier transform. The DWT
is demonstrated by (2):
Wa,b
ðÞ
¼1
ffiffiffi
b
pX
n1
t¼0
ft
ðÞwta
b
 (2)
where a¼k2j,b¼2j, j is scale index, k is wavelet
transform signal index, and wðtÞis the mother wave-
let. Multi-level decomposition can be applied by
DWT. The input time series at the first level is
Figure 2. an overview of the proposed android application for data collection. (a) selecting the desired sensors for data collection,
(b) the application during data collection where no maneuver is taking place, (c) the application during data collection where a
brake is happening and being labeled.
Table 2. Collected data description.
Group Number of trips
Number of
turn events
Number of
brake events
Nokia 95 260 966
Samsung 63 377 1214
Asus 21 91 129
Huawei 12 304 467
Xiaomi 19 182 589
Others 35 220 828
Sum 245 1434 4193
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 5

decomposed to Approximation and Detail coefficients
(A
1
and D
1
). Then, A
1
could be also fed again into a
DWT to generate A
2
and D
2
. This process can be
continued, but in this research, A
1
and A
2
will
be extracted.
Classifiers
Random Forest (RF) and Artificial Neural Network
(ANN) with fully connected layers are two types of
classifiers employed in our method. In RF (Ho, 1995),
many decision trees work together as an ensemble to
predict the output of unseen data. To elaborate, in a
single decision tree (or C4.5), prediction is provided
based on some trained values for different features of
the dataset. In each step of training, a greedy algo-
rithm is utilized to specify the best choice for splitting
the dataset using the Gini impurity loss function,
measuring the class distribution of samples. Gini
impurity of different classes can be calculated by the
following equations:
G node
ðÞ
¼X
n
k¼1
pkð1pkÞ(3)
pk¼#observations with class k
#all observations (4)
where pkis the probability of selecting a sample from
class k and n is the number of observations in a node.
Figure 3. A brief overview of the methodology. (a) Event Extraction Phase (Phase 1), (b) Performance Improvement Phase
(Phase 2).
6 M. ZAREI YAZD ET AL.

Using a greedy approach may cause some error in the
result since the greedy models only determine local
optimums. By using multiple decision trees in the RF
classifier, different combinations of the training data-
set are produced and used as the input of different
decision trees. Finally, when the trees are learned by
the various bags of the training dataset, the prediction
result would be the mode of the decision trees’out-
puts. Taking advantage of this method in RF leads to
more accurate results as more randomness
is considered.
ANN (Haykin, 1994) is another popular type of
classifier consisting of several connected nodes trans-
mitting some information. An activation function is
embedded into the nodes of an ANN, adding nonli-
nearity to the results. In other words, the activation
function decides to activate a neuron or not based on
the signal received from the other nodes. A mathem-
atical representation of the activation function (ReLU)
is shown below:
Rz
ðÞ¼maxð0, zÞ(5)
where zis the output of each neuron. While training
ANNs, different features of the training dataset are
fed to the first layer of the nodes, and then using the
weights between the nodes, signals are sent to the
other nodes as below:
fx,w
ðÞ
¼X
n
i¼1
xiwiþb(6)
where xiand wiare the i-th feature and weight from
xand wn-dimensional vectors, and bis the bias. This
continues until the information is received by the out-
put layer’s nodes which is the final step to determine
the prediction of the model (feedforward function). In
this step, a decision is simply made with the reformat-
ted data by the previous nodes. Then an error func-
tion (binary cross-entropy in our case) is defined to
measure how far the output is from reality:
L¼ylog p
ðÞ
1y
ðÞ
logð1pÞ(7)
where yis the real label and pis the predicted prob-
ability by the model. After that, by backpropagation
function, different weights between the nodes are
modified exploiting the gradient descent technique.
These two functions (feedforward and
backpropagation) iteratively continue working until a
low and reasonable error is achieved. The selected
hyperparameters of NN and RF are provided in Table
3. These hyperparameters are set experimentally by
comparing the achieved accuracy.
Event Extraction Phase (Phase 1)
The main aim of the first phase is to determine the
potential intervals from a time series data recorded in
a whole trip. The raw data recorded by smartphone
sensors according to the data section of this paper can
be described by (8) and (9):
RS
A¼frS,A
ðÞ
t1,rS,A
ðÞ
t2,rS,A
ðÞ
t3,:::,rS,A
ðÞ
ti ,:::,rS,A
ðÞ
tN g(8)
LE¼flE
t1,lE
t2,lE
t3,:::,lE
ti,:::,lE
tN g(9)
where RS
Adenotes data time series recorded by the A
axis (x, y, or z) of the sensor S (acc or gyr) in the
smartphone and the rS,A
ðÞ
ti is a single data recorded at
the ti-th time. Moreover, LEis the label time series
recorded for the event type E (brake or turn) and lE
ti is
always 0 or 1, denoting whether rS,A
ðÞ
ti is part of a spe-
cific event or not. Additionally, the total number of
the recorded data points in a whole trip is N. Based
on Figure 1, the raw input for the phase one in the
brake detection part is RAcc
yand in the turn detection
part are RGyr
zand RAcc
x:The raw inputs, in this phase,
are first modified by different filters as shown. In fact,
by noise elimination, RS
Abecomes FRS
A, where F dem-
onstrate the type of filter (L for low-pass, H for high-
pass, W
1
for level one and W
2
for level two of wavelet
filter). After data filtration, the modified inputs are
fed into the proposed classifiers in each part.
Therefore, every recorded single point in the relevant
FRS
Ais the input of the classifier, and the classifier
predicts lE
ti ð
^
lE
tiÞ:The shape of a sample input for the
brake and turn detection classifiers is demonstrated in
(10) and (11) respectively:
brakeinput ¼fW1rAcc,y
ðÞ
t1,Lr Acc,y
ðÞ
t1g(10)
turninput ¼fLr Gyr,z
ðÞ
t1,Hr Gyr,z
ðÞ
t1,Lr Acc,x
ðÞ
t1g(11)
All the notations are defined previously. The above
equations show that, for example, the neural network
classifier in the brake detection part is trained to
Table 3. Classifiers’hyperparameters.
Random forest Artificial neural network
Criterion Gini Number of hidden layers 2
Maximum depth 5 Number of neurons per layer 80 and 40
Number of estimators 40 Learning rate 0.001 (Constant)
Min_samples_split 2 Maximum iteration without meeting improvement 10
Min_samples_leaf 1 Solver Adam
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 7

determine whether a single data point is part of a
brake or not by using two features, Wr Acc,y
ðÞ
t1and
Lr Acc,y
ðÞ
t1:Similarly, this is done in the turn section by
three features namely Lr Gyr,z
ðÞ
t1,Hr Gyr,z
ðÞ
t1, and Lr Acc,x
ðÞ
t1:
The predicted
^
lE
tis together, then, create LE-pred.In
the last step of this phase, intervals with
^
lE
ti ¼1 are
considered as primary driving behavior intervals.
These intervals are extracted to be used in the
next phase.
To show the performance of this phase and prove
the necessity of designing the second phase, a time
series label comparison method (Figure 4) is proposed
in this paper for comparing LEand LE-pred. This
method, additionally, creates ground-truth labels to be
used as a measure of performance for training and
testing the classifiers in the next phase. A point-by-
point comparison is not reasonable to compare LE
and LE-pred, since some unintentional lags could be
observed between the real labels and where the driv-
ing event is actually happened due to the labeler
responsiveness. A sample of this lag is depicted in
Figure 5, where the yellow boxes show the time inter-
vals specified by the labeler, and they are not matched
with the events. Therefore, there is a need for a new
method to consider these situations as true positives.
According to Figure 4, the Label Comparison
Function needs two lists of real and predicted inter-
vals as input. Primarily, all the predicted intervals are
counted as False Positive (FP). Then, for the first real
event (first period with lE
ti ¼1), the method finds all
overlapped predicted intervals (periods with
^
lE
ti ¼1)
and saves them in the options list. If the options list
remains empty, the first real event will be considered
a false negative (FN) event. However, if the options
list includes one or more predicted intervals, the first
interval in this list will be considered as a True
Positive (TP), and therefore, one of the FP intervals
becomes a TP. This structure is applied for all the real
events existing in the LE:Thereafter, based on the
findings (FN, TP, and FP), precision, recall, and F1-
score of the first phase can be calculated with (12),
(13), and (14):
Precision ¼TP
TP þFP (12)
Recall ¼TP
TP þFN (13)
F1score ¼2Precision Recall
Precision þRecall (14)
Performance Improvement Phase (Phase 2)
As we stated before, the outputs of phase one may
contain some false positive events. However, the num-
ber of false negatives is not usually high in phase one.
For enhancing the precision of the model, we have
trained new classifiers to recognize the truly extracted
intervals. A sample extracted interval from the first
phase is shown by (15):
R'S
A¼fr'S,A
ðÞ
t1,r'S,A
ðÞ
t2,:::,r'S,A
ðÞ
ti ,:::,r'S,A
ðÞ
tn g(15)
where R0
ASis the data interval from the Aaxis (x, y,
or z) of sensor S (acc or gyr) captured by the previous
phase and the r0
ti S,A
ðÞ
is a single data recorded at the
ti-th time in the interval. Since R0
ASis a section of RS
A,
we can state that n <N.
According to Figure 3, in the second phase, some
statistical features are experimentally chosen to be cal-
culated from FR0
AS(Fdenoting the type of assigned
filter). The details of these features and their defini-
tions are demonstrated in Table 4. The minimum,
Figure 4. Label comparison method.
8 M. ZAREI YAZD ET AL.

maximum, and energy level of the intervals are the
statistical features used in this section. Among all the
features, the B1 and T1 features in Table 4 need some
inputs more than just a single interval. One second of
LRAcc
yor LRAcc
xbefore the extracted interval (i¼pto
q) and another one second after that (i¼p’to q’) are
necessary to compute these features. Other input fea-
tures are extracted using the intervals which are
selected in the previous phase.
After the feature derivation from multiple extracted
intervals, they are fed to the proposed classifiers. The
classifiers are expected to determine whether an inter-
val is a real event or a false positive event. For brake
intervals, a random forest and for turn intervals, a
neural network with fully connected layers is
designed. The shape of a sample input data (j-th inter-
val) for brake and turn classifiers are denoted by (16)
and (17) respectively:
brakeinput ¼fB1j,B2j,B3j,B4jg(16)
turninput ¼fT'1j,T'2j,T'3j,T'4j,T'5jg(17)
where B1j,…,B4jand T1j,…,T5jare the calcu-
lated features. It should be noted that for learning the
classifiers in this phase, real labels are assigned by the
Label comparison method provided in phase one.
Therefore, for testing the classifiers it is determined
whether a single interval refers to a true positive event
or not (L0
E¼0 or 1).
Results and discussion
For demonstrating the proficiency of the proposed
method, two different testing scenarios are considered
in this section. A sensitivity analysis is also imple-
mented to clarify the importance of different parts
used in our method. In the first scenario, we will
show the performance of our designed model using
the same approach adopted previously in the litera-
ture. Afterward, in the second scenario, we will dem-
onstrate that our model also yields satisfactory results
when it comes to the prediction of newly collected
data in different situations.
First scenario
In this part, some different sub-datasets are extracted
from the gathered data described previously to train
Figure 5. Unintended lags between the real labels and the predicted events.
Table 4. Statistical features calculated in Phase 2.
Event Feature Definition Formula
Brake B1 Mean (LR0Acc
yÞ–Mean (1 sec before and after the LR0Acc
y)Pn
i¼1Lr0
tiðAcc:yÞ
nPq
i¼pLrðAcc:yÞ
ti þPq0
i¼p0LrðAcc:yÞ
ti
qp
ðÞ
þðq0p0Þ
B2 Min (LR0
yAcc) minðLr0Acc,y
ðÞ
ti Þ
B3 Energy (W1R0
yAccÞPn
i¼1ðW1r0Acc,y
ðÞ
ti Þ2
B4 Energy (W2R0
yAccÞPn
i¼1ðW2r0Acc,y
ðÞ
ti Þ2
Turn T1 Mean (LR0
xAcc)–Mean (1 sec before and after the LR0Acc
x)Pn
i¼1Lr0
tiðAcc,xÞ
nPq
i¼pLrðAcc,xÞ
ti þPq0
i¼p0LrðAcc,xÞ
ti
qp
ðÞ
þðq0p0Þ
T2 Max (absðLR0Acc
xÞ) maxðabsðLr0Acc,x
ðÞ
ti ÞÞ
T3 Max (absðLR0Gyr
zÞ) maxðabsðLr0Gyr,z
ðÞ
ti ÞÞ
T4 Max (absðHR0
zGyrÞ) maxðHr0Gyr:z
ðÞ
ti Þ
T5 Duration (R0
AS)n
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 9

and test the model. All the collected data in this
research are divided into six groups and every single
group is different from the others in terms of driver,
route, and smartphone type. To show the model’s per-
formance in an experiment like the literature, some
sub-datasets are generated by different combinations
of the six data groups presented in Table 2.Table 5
shows the properties of these sub-datasets (bundles).
For this analysis, we created two to five-phone bun-
dles, and another bundle containing all the collected
data. The proposed two-phased method in this paper
is trained and tested using these bundles by 5-fold
cross-validation, and the results are reported in
Table 6.
Some essential points are revealed according to
Table 6. As demonstrated, the first phase of the pro-
posed method finds the potential intervals with some
false positive events. Therefore, a high recall and a
relatively low precision are shown after this phase.
However, by extracting the statistical features and
using the second classifier, both precision and recall
improve, and as a result, a high F1-score is provided.
Additionally, the proposed two-phase method has a
better performance in turn detection compared to
brake detection. Turns are usually more sensible for
the labeler; however, brake intensity varies in different
situations, and this may confuse the labeler while con-
fronting multiple intensities. Furthermore, there are
two sensors (accelerometer and gyroscope) for turn
recognition, but there is just the accelerometer sensor
available in smartphones for brake detection.
In Table 6, the larger the bundle becomes, the
more the accuracy drops. A large and diverse dataset
contains different types of each driving event (brakes
and turns) since different smartphones, drivers, and
trip routes are covered by that. As a result, the unique
characteristics of each different event become more
uncertain employing large datasets, and this causes
accuracy reduction of the model. However, by training
and testing our structure using such a dataset against
some datasets covering just one or a few situations,
we can further address the problem of overfitting in
our solution rather than just using train-test splitting
or cross-validation. This phenomenon is mostly over-
looked by the previous papers utilizing just one data-
set collected in unique situations. In other words,
here, we demonstrated that our model still works well
(with F1-scores of more than 70 percent) not only
when cross-validation or train-test split is applied on
data with unique characteristics (like the case in previ-
ous papers), but also when the data is collected in
various situations.
Second scenario
In this scenario, we train and test the proposed model
to measure its robustness. By robustness, we mean
that our model performs well even in situations when
the train and test datasets are different, and they are
gathered in an independent time and place using dif-
ferent smartphones and drivers. As the partitioned
datasets in Table 2 are completely different from each
other, we employed them for this aim in this section.
To demonstrate the robustness, we will compare
the accuracy of our model when we use a train-test
split for a single database and when a different
Table 5. Bundles description.
Brake detection Turn detection
Bundle number Bundle phones Bundle number Bundle phones
2 Xiaomi and Huawei 2 Asus and Huawei
3 Xiaomi, Huawei, and other 3 Asus, Nokia, and Huawei
4 Xiaomi, Nokia, Huawei, and other 4 Asus, Xiaomi, Nokia, and Huawei
5 Samsung, Xiaomi, Nokia, Huawei, and other 5 Samsung, Asus, Xiaomi, Nokia, and Huawei
6 Samsung, Asus, Xiaomi, Nokia, Huawei, and other 6 Samsung, Asus, Xiaomi, Nokia, Huawei, and other
Table 6. Brake and turn detection results by 5-fold cross validation.
Event type # Bundle
First phase results Improved results (second phase)
Precision % Recall % F1-score % Precision % Recall % F1-score %
Brake 2 64.13 96.47 77.04 87.52 89.56 88.53
3 54.11 95.2 69.00 79.78 79.88 79.83
4 49.56 92.57 64.56 79.13 68.65 74.14
5 46.98 89.2 61.55 76.16 66.14 70.79
6 45.83 89.37 60.59 75.30 66.46 70.60
Turn 2 75.86 95.65 84.61 93.18 94.25 93.71
3 78.53 93.85 85.51 93.35 91.59 92.46
4 76.4 91.71 83.36 89.52 86.28 87.87
5 73.49 88.6 80.34 87.32 80.43 83.73
6 71.5 88.81 79.22 84.99 79.00 81.89
10 M. ZAREI YAZD ET AL.

validation dataset is considered. Each dataset in Table
2is first split into 80 percent training and 20 percent
testing batches. Then, we trained and tested our
model six times with these created train-test splits.
Additionally, we also considered each row of Table 2
as a training set and the other rows as a testing set,
and again, we trained and tested our method another
six times. The results of these experiments are
depicted in Figure 6.
The first point about Figure 6 is the considerable
drop in F1-score in most of the cases when we use a
validation dataset different from the training set. To
the best of our knowledge, this experiment is never
done in previous studies. In the literature, all the gen-
erated models are fitted to a single dataset, and this
will increase the probability of overfitting even if the
collected dataset is large. Although by using a valid-
ation dataset our method does not perform as well as
using train-test split, the F1-scores yielded in Figure 6
remained acceptable and relatively high. From Figure
6, we can infer that by employing a small dataset (one
of the six datasets in Table 2) as a training set and
testing the model by an approximately five-time larger
and unseen dataset (other five datasets), the accuracy
remains reasonable and relatively high (about 70 and
80 percent of F1-score in brake and turn detection).
Another interesting observation in Figure 6 is that
the performance is somehow related to the training
size. This means that datasets containing a lower
number of events lead to lower F1-scores. For
instance, when the method is trained by Asus or
Xiaomi datasets, specifically in brake detection, the
validation results are 66 and 68 percent of F1-score,
respectively. As shown in Table 2, these two datasets
(Asus and Xiaomi) include 129 and 589 brake samples
which is a low number compared to the other data-
sets. However, the dataset size is not the only reason
affecting the results. Other factors such as labeling
error, smartphone type, and environmental situation
should also be considered for inferring the results. As
seen, the Huawei dataset contains 467 brake samples,
but the achieved F1-score is higher than Asus
and Xiaomi.
A detailed report of the results, choosing the
Huawei dataset as the training set and the other data-
sets as the test set, is shown in Table 7. The model
recognizes brakes and turns by using one part of a
six-part dataset with an accuracy of 71 and 82 percent,
respectively. The impact of the second phase intro-
duced in our method in terms of improving the preci-
sion is also observable in this scenario, Table 7.In
summary, the result indicates that although there are
Figure 6. Results of testing the model by second scenario.
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 11

different smartphones, drivers, and rout conditions in
reality affecting the motion data of vehicles, it is pos-
sible to train a model based on the data collected in a
few situations and use it where there is no historical
data of another new particular participant (e.g., differ-
ent driver or smartphone). This challenge becomes
more visible when exploiting widely used smartphones
instead of precise but pricy OBD devices for driving
behavior monitoring as there are several brands and
types in the world. However, the result in this paper
shows that our structure is capable of handling it
based on our experiments.
Feature sensitivity analysis
In this section, the significance of each part of the
model such as inputs and features are analyzed. As
discussed earlier, accelerometer and gyroscope data,
along with different filters and statistical features are
utilized to detect driving maneuvers. Therefore, their
marginal contribution in accuracy improvement
worth further investigation. For this analysis, 80-20
train-test split is implemented with the Huawei data-
set, and each part of the method is systematically
dropped to find the effect of that part on
the accuracy.
The sensitivity analysis of the first phase is demon-
strated in Table 8. For brake and turn detection, first,
we implemented our completed method. Then, differ-
ent parts of the method like filters in brake detection
and sensors in turn detection are removed. The aim
in the first phase is to get a high recall, and therefore,
it is observable that by neglecting each part in this
phase the recall metric also becomes lower. Besides,
the most drop in the recall is recorded when all filters
in turn detection part are removed and raw data is
used instead. This shows the significance of noise
reduction and using filters in our method. Although
this is not the case for brake detection in this experi-
ment, a huge recall reduction was observed in another
experiment exploiting different datasets for training
and testing (different from the results demonstrated
in Table 8). In this new experiment, we used one of
the previous validation-sets (as discussed in scenario
2) as a test set and the recall reduction for brake
detection was about 37 percent which is considerable.
This observation showed that the intensity of noise
produced by sensors differs from one type of smart-
phone to another and this may affect the result of
event detection. Since the employed filters in our
approach reshape the data from time-domain to fre-
quency-domain and then removes the noise based on
frequency, we can state that our method is also robust
to the amount of noise, and this is also proved based
on these observations.
For studying the sensitivity of our model in the
second phase to the different input parameters, we
used the first phase with no reduction, as it has the
highest recall, and then we dropped some different
features of the second phase. The neglected features
and the achieved accuracy parameters after the
second phase are reported in Table 9. As we can see,
in the brake detection section, removing energy-
related features and other statistical features affect
the F1-score and decrease it by approximately 2 to 4
percent. Since during brakes a reduction of acceler-
ometer data is expected, energy related features help
sign some milder brakes and then the combination
of both statistical and energy related features reveals
Table 7. The performance of our method trained by a single
dataset and tested on the whole other data.
Event type First phase results Second phase results
Precision Recall F1-score Precision Recall F1-score
Brake 43.99 88.35 58.74 71.06 71.52 71.29
Turn 72.97 89.38 80.35 81.7 81.4 81.56
Table 8. First phase feature sensitivity analysis.
Dropped item Description Precision Recall F1-score
Nothing –65.51 97.64 78.41
Brake
detection
W1RAcc
yNo wavelet filter 83.08 95.71 88.95
LRAcc
yNo low-pass filter 75.68 95.28 84.36
Filters Just raw data is used 87.33 95.93 91.43
Turn
detection
Nothing –70.35 98.35 82.03
RAcc
xNo accelerometer data 62.71 97.36 76.28
RGyr
zNo gyroscope data 74.32 90.46 81.60
Filters Just raw data is used 64.67 81.91 72.28
Table 9. Second phase feature sensitivity analysis.
Dropped item Description Precision Recall F1-score
Brake detection Nothing –93.68 94.68 94.18
B3 and B4 No energy related feature 85.57 94.68 89.89
B1 Removing B1 91.66 93.61 92.63
B2 Removing B2 89.89 94.68 92.22
Turn detection Nothing –97.82 91.84 94.74
Duration (T5) Not using T5 feature 96.38 81.63 88.39
T3 and T4 No gyroscope feature 95.89 71.42 81.87
T1 and T2 No accelerometric feature 93.54 88.78 91.10
12 M. ZAREI YAZD ET AL.

the best result. In the turn detection section, the
marginal contribution of the features is higher than
the brake section. This can be justified by the fact
that there are different types of turns in reality. For
example, some of them are simply right or left turn
while others may be U-turns. Therefore, the model
embedded in our proposed structure may use each of
the features for the detection of any specific type of
event. Furthermore, it is observed that gyroscopic
features are more vital to get the F1-score high
rather than accelerometric features and the duration
feature (T5) is an essential part of the second phase
as its absence caused a 6.5% reduction in the F1-
score. Some of the turn samples always contain a
brake in themselves. Consequently, it would be frus-
trating for the model to solely rely on the accelerom-
eter information. In these situations, the duration of
the events and the gyroscope data would be benefi-
cial for detection.
Summary and conclusion
In this paper, a two-phase method using Random
Forest and Artificial Neural Network (Multi-Layer
Perceptron) along with high-pass, low-pass, and wave-
let filters is established to detect driving brakes and
turns. Suspicious intervals from filtered time series
data are first extracted with a high recall in our
method. Then, false positive samples are dropped by
feeding statistical features of the suspicious intervals
into specific classifiers. A label comparison method is
also developed to compare real labels of the events
with the predicted ones to overcome some limitations
of the previous sliding windows. The robustness of
this method is demonstrated by different scenarios
employing six distinct datasets collected by different
drivers and smartphones in multiple routes.
Additionally, the marginal contribution of input fea-
tures in this method, specifically different filters, is
quantified to show the significance of each part in
achieving high performance.
The proposed methodology in this paper can be
used as a monitoring system to detect driving maneu-
vers and provide feedbacks for drivers about their
driving styles. Our experiments show that this method
reveals acceptable results not only when the train and
test data are collected in similar situations but also
when the test data is different from the training set.
Results in this paper demonstrated that although the
performance usually drops by transferring a trained
model into a different situation, the proposed method
in this paper does not experience a high reduction in
accuracy while being used for different drivers, smart-
phones, and routes. It is shown that by training our
two-phase method with a dataset containing enough
number of events (e.g., 300 samples per event type),
we can detect braking and turning of other drivers in
different situations with an average F1-score of 71%
and 81% respectively. Using filters for data denoising
also showed a significant impact on accuracy and
raised the recall in the first phase of brake and turn
detection about 2% and 16%, respectively.
This research venue may be extended in several
ways such as developing an unsupervised or semi-
supervised approach for detecting driving behaviors
and testing them by new unseen datasets. Collecting
unsupervised or partially supervised data is less time
and labor consuming. Furthermore, another area
worth exploring is also investigating the effects of the
smartphones’placement on data quality.
Disclosure statement
No potential conflict of interest was reported by
the authors.
ORCID
Iman Taheri Sarteshnizi http://orcid.org/0000-0002-
7798-8788
References
Bates, L., Soole, D., & Watson, B. (2012). The effectiveness
of traffic policing in reducing traffic crashes. In
T. Prenzler (Ed.), Policing and security in practice (pp.
90–109). Springer.
Bejani, M. M., & Ghatee, M. (2018). A context aware system
for driving style evaluation by an ensemble learning on
smartphone sensors data. Transportation Research Part C:
Emerging Technologies,89(February), 303–320. https://
doi.org/10.1016/j.trc.2018.02.009
Carlos, M. R., Gonzalez, L. C., Wahlstrom, J., Ramirez, G.,
Martinez, F., & Runger, G. (2020). How smartphone
accelerometers reveal aggressive driving behavior? The
key is the representation. IEEE Transactions on Intelligent
Transportation Systems,21(8), 3377–3387. https://doi.org/
10.1109/TITS.2019.2926639
Carvalho,E.,Ferreira,B.V.,Jr.,Souza,J.F.,De,C.,Carvalho,
H. V., Suhara, Y., Pentland, A. S., Pessin, G., & Behavior,
A. D. (2017). Exploiting the use of recurrent neural net-
works for driver behavior profiling. In Y. Choe, C. Jayne, B.
Hammer, & I. King (Eds.), International Joint Conference
on Neural Networks (IJCNN) (pp. 3017–3021). IEEE.
Chan, T. K., Chin, C. S., Chen, H., & Zhong, X. (2020). A
comprehensive review of driver behavior analysis utilizing
smartphones. IEEE Transactions on Intelligent
Transportation Systems,21(10), 4444–4475. https://doi.
org/10.1109/TITS.2019.2940481
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 13

Chhabra, R., Verma, S., & Krishna, R. (2018). Detecting
aggressive driving behavior using mobile smartphone. In
C. Rama Krishna, R. Kumar, & M. Dutta (Eds.),
Proceedings of 2nd International Conference on
Communication, Computing and Networking (pp.
513–522). Springer.
Daptardar, S., Lakshminarayanan, V., Reddy, S., Nair, S.,
Sahoo, S., & Sinha, P. (2015, November 1–4). Hidden
Markov model based driving event detection and driver
profiling from mobile inertial sensor data [Paper presenta-
tion]. 2015 IEEE Sensors. Busan, Korea (South). https://
doi.org/10.1109/ICSENS.2015.7370312
Eftekhari, H. R., & Ghatee, M. (2018). Hybrid of discrete
wavelet transform and adaptive neuro fuzzy inference
system for overall driving behavior recognition.
Transportation Research Part F: Traffic Psychology and
Behaviour,58, 782–796. https://doi.org/10.1016/j.trf.2018.
06.044
Eftekhari, H. R., & Ghatee, M. (2019). A similarity-based
neuro-fuzzy modeling for driving behavior recognition
applying fusion of smartphone sensors. Journal of
Intelligent Transportation Systems,23(1), 72–83. https://
doi.org/10.1080/15472450.2018.1506338
Ferreira, J., Pentland, A., Ferreira, B. V., Pessin, G.,
Carvalho, E., de Souza, C., & Suhara, Y. (2017). Driver
behavior profiling: An investigation with different smart-
phone sensors and machine learning. PloS One,12(4),
e0174959. https://doi.org/10.1371/journal.pone.0174959
Haque, M. E., Khan, M. N. S., & Sheikh, M. R. I. (2016).
Smoothing control of wind farm output fluctuations by
proposed low pass filter, and moving averages. ICEEE
2015 –1st International Conference on Electrical and
Electronic Engineering (pp. 121–124). https://doi.org/10.
1109/CEEE.2015.7428234
Haykin, S. (1994). Neural networks: A comprehensive foun-
dation. Prentice Hall.
Ho, T. K. (1995). Random decision forests. Proceedings of
3rd International Conference on Document Analysis and
Recognition (Vol. 1, pp. 278––282).
Jamson, S. L., Hibberd, D. L., & Jamson, A. H. (2015).
Drivers’ability to learn eco-driving skills; effects on fuel
efficient and safe driving behaviour. Transportation
Research Part C: Emerging Technologies,58(PD), 657–668.
https://doi.org/10.1016/j.trc.2015.02.004
Johnson, D. A., & Trivedi, M. M. (2011). Driving style rec-
ognition using a smartphone as a sensor platform. IEEE
Conference on Intelligent Transportation Systems (ITSC)
(pp. 1609–1615). https://doi.org/10.1109/ITSC.2011.
6083078
Kazemeini, A., Taheri, I., & Samimi, A. (2022). A GPS-
based algorithm for brake and turn detection.
International Journal of Intelligent Transportation Systems
Research,20(2), 433–445. https://doi.org/10.1007/s13177-
022-00301-9
Ma, Y., Zhang, Z., Chen, S., Yu, Y., & Tang, K. (2019). A
comparative study of aggressive driving behavior recogni-
tion algorithms based on vehicle motion data. IEEE
Access,7, 8028–8038. https://doi.org/10.1109/ACCESS.
2018.2889751
Malghan, P. G., & Hota, M. K. (2020). A review on ECG fil-
tering techniques for rhythm analysis. Research on
Biomedical Engineering,36(2), 171–186. https://doi.org/
10.1007/s42600-020-00057-9
Nguyen, T., Lu, D., Nguyen, D., & Nguyen, H. (2020).
Dynamic basic activity sequence matching method in
abnormal driving pattern detection using smartphone
sensors. Electronics,9(2), 217. https://doi.org/10.3390/
electronics9020217
Nuswantoro, F. M., Sudarsono, A., & Santoso, T. B. (2020).
Abnormal driving detection based on accelerometer and
gyroscope sensor on smartphone using artificial neural
network (ANN) algorithm. 2020 International Electronics
Symposium (IES) (pp. 356–363). https://doi.org/10.1109/
IES50839.2020.9231851
Ouyang, Z., Niu, J., & Guizani, M. (2018). Improved vehicle
steering pattern recognition by using selected sensor data.
IEEE Transactions on Mobile Computing,17(6),
1383–1396. https://doi.org/10.1109/TMC.2017.2762679
Saleh, K., Hossny, M., & Nahavandi, S. (2017). Driving
behavior classification based on sensor data fusion using
LSTM recurrent neural networks. 2017 IEEE 20th
International Conference on Intelligent Transportation
Systems (ITSC) (pp. 1–6). https://doi.org/10.1109/ITSC.
2017.8317835
S
anchez, S. H., Pozo, R. F., & Hern
andez G
omez, L. A.
(2018). Estimating vehicle movement direction from
smartphone accelerometers using deep neural networks.
Sensors (Switzerland),18(8), 2624. https://doi.org/10.3390/
s18082624
Shinde, P. P., & Shah, S. (2018, August 16–18). A review of
machine learning and deep learning applications.
Proceedings –2018 4th International Conference on
Computing, Communication Control and Automation,
ICCUBEA 2018. IEEE. https://doi.org/10.1109/ICCUBEA.
2018.8697857
Siami, M., Naderpour, M., & Lu, J. (2021). A mobile tele-
matics pattern recognition framework for driving behav-
ior extraction. IEEE Transactions on Intelligent
Transportation Systems,22(3), 1459–1472. https://doi.org/
10.1109/TITS.2020.2971214
Singh, G., Bansal, D., & Sofat, S. (2017). A smartphone
based technique to monitor driving behavior using DTW
and crowdsensing. Pervasive and Mobile Computing,40,
56–70. https://doi.org/10.1016/j.pmcj.2017.06.003
Tanvir, S., Chase, R. T., & Roupahil, N. M. (2021).
Development and analysis of eco-driving metrics for nat-
uralistic instrumented vehicles. Journal of Intelligent
Transportation Systems,25(3), 235–248. https://doi.org/
10.1080/15472450.2019.1615486
Toledo, T., Musicant, O., & Lotan, T. (2008). In-vehicle
data recorders for monitoring and feedback on drivers’
behavior. Transportation Research Part C: Emerging
Technologies,16(3), 320–331. https://doi.org/10.1016/j.trc.
2008.01.001
Wang, R., Xie, F., Zhao, J., Zhang, B., Sun, R., & Yang, J.
(2021). Smartphone sensors-based abnormal driving
behaviors detection: Serial-feature network. IEEE Sensors
Journal,21(14), 15719–15728. https://doi.org/10.1109/
JSEN.2020.3036862
World Health Organization. (2018). Global status report on
road safety 2018.
Wu, Z., Liu, P., Liu, Q., & Wang, Y. (2018). MEMS-based
IMU assisted real time difference using raw
14 M. ZAREI YAZD ET AL.

measurements from smartphone. Proceedings of the 31st
International Technical Meeting of the Satellite Division of
the Institute of Navigation (ION GNSSþ2018) (pp.
445–454). https://doi.org/10.33012/2018.15928
Xie, J., Hilal, A. R., & Kulic, D. (2018). Driving maneuver
classification: A comparison of feature extraction meth-
ods. IEEE Sensors Journal,18(12), 4777–4784. https://doi.
org/10.1109/JSEN.2017.2780089
Xie, J., Kulic, D., & Hilal, A. R. (2018, October 7–10).
Driver distraction recognition based on smartphone sen-
sor data. 2018 IEEE International Conference on Systems,
Man, and Cybernetics (SMC). Miyazaki, Japan. https://
doi.org/10.1109/SMC.2018.00144
Yao, Y., Zhao, X., Wu, Y., Zhang, Y., & Rong, J. (2021).
Clustering driver behavior using dynamic time warping
and hidden Markov model. Journal of Intelligent
Transportation Systems,25(3), 249–262. https://doi.org/
10.1080/15472450.2019.1646132
Yi, D., Su, J., Liu, C., Quddus, M., & Chen, W. H. (2019).
A machine learning based personalized system for driving
state recognition. Transportation Research Part C:
Emerging Technologies,105, 241–261. https://doi.org/10.
1016/j.trc.2019.05.042
Yu, J., Chen, Z., Zhu, Y., Chen, Y., Kong, L., & Li, M.
(2017). Fine-grained abnormal driving behaviors detec-
tion and identification with smartphones. IEEE
Transactions on Mobile Computing,16(8), 2198–2212.
https://doi.org/10.1109/TMC.2016.2618873
Yuksel, A. S., & Atmaca, S. (2021). Driver’s black box: A
system for driver risk assessment using machine learning
and fuzzy logic. Journal of Intelligent Transportation
Systems,25(5), 482–500. https://doi.org/10.1080/
15472450.2020.1852083
Zhang, J., Wu, Z., Li, F., Luo, J., Ren, T., Hu, S., Li, W., &
Li, W. (2019). Attention-based convolutional and recur-
rent neural networks for driving behavior recognition
using smartphone sensor data. IEEE Access.7,
148031–148046. https://doi.org/10.1109/ACCESS.2019.
2932434
JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS 15