IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO.

2, JANUARY 15, 2021 869

An IoT-Based Vehicle Accident Detection and

Classification System Using Sensor Fusion
Nikhil Kumar, Debopam Acharya, and Divya Lohani , Member, IEEE

Abstract—Road accidents are a leading cause of death and In many cases, human lives are lost in road accidents due to
disability among youth. Contemporary research on accident delays in emergency medical assistance. According to Golden
detection systems is focused on either decreasing the reporting Hour Principle [3], there is a high probability that timely med-
time or improving the accuracy of accident detection. Internet-of-
Things (IoT) platforms have been utilized considerably in recent ical and surgical aid can avoid death during the golden hour,
times to reduce the time required for rescue after an accident. which is the period after the traumatic injury. A decrease in the
This work presents an IoT-based automotive accident detection response time of emergency medical care can reduce the prob-
and classification (ADC) system, which uses the fusion of smart- ability of death by one-third on an average [4]. The percentage
phone’s built-in and connected sensors not only to detect but also of people who die before reaching the hospital in low- and
to report the type of accident. This novel technique improves the
rescue efficacy of various emergency services, such as emergency middle-income countries is more than twice as compared to
medical services (EMSs), fire stations, towing services, etc., as high-income countries [2]. In the recent past, information and
knowledge about the type of accident is extremely valuable in communication technology, such as Internet of Things (IoT)
planning and executing rescue and relief operations. The emer- has been used to decrease the accident rescue time. IoT is
gency assistance providers can better equip themselves according an interconnection of a vast variety of embedded and smart
to the situation after making an inference about the injuries
sustained by the victims and the damage to the vehicle. In this devices, such as computers, smartphones, smart sensors and
work, three machine learning models based on Naïve Bayes (NB), actuators, embedded processors, etc., with the modern Internet.
Gaussian mixture model (GMM), and decision tree (DT) tech- IoT is a potential medium for tracking and control of smart
niques are compared to identify the best ADC model. Five automobiles that can link any connected physical unit to a con-
physical parameters related to vehicle movement, i.e., speed, trol server [5]. Most researchers have confined their work to
absolute linear acceleration (ALA), change-in-altitude, pitch, and
roll, have been used to train and test each candidate ADC model improving the accuracy of accident detection, estimating the
to identify the correct class of accident among collision, rollover, severity of road accidents or minimizing the rescue time post-
falloff, and no accident. NB-based ADC model is found to be occurrence of an accident [6], [7]. In addition, most systems
highly accurate with 0.95 mean F1-score. to detect and report road accidents are expensive and limited to
Index Terms—Accident detection and classification (ADC), high-end vehicles. Another drawback of the current systems is
decision tree (DT), Gaussian mixture model (GMM), Internet their inability in identifying the type of accident as a collision,
of Things (IoT), Naïve Bayes (NB), sensor fusion. rollover, or a falloff event. Merely reporting the occurrence of
an accident event severely limits the ability of the emergency
rescue workers to provide the victims with the right kind of
I. I NTRODUCTION
rescue support and medical aid. The information about the type
EATHS and disabilities by road accidents are increasing
D with each passing year. An increase in population and
per capita incomes has led to an increase in ownership and the
of accident sustained provides valuable information about the
damages sustained by the vehicle and the extent of injuries
sustained by the victims.
presence of vehicles on roads. Greater traffic volumes, over- This research work answers the following questions.
speeding, reckless and drunken driving, driver fatigue, poor 1) Can there be an inexpensive accident detection and clas-
road infrastructure, and the presence of animals on roads are sification (ADC) system that can be retrofitted to any
some reasons that are responsible for road fatalities. According vehicle?
to the World Health Organization (WHO), the percentage of 2) Which is the best suited machine-learning classification
road accident fatalities to the total number of deaths world- model that can accurately detect and classify the road
wide has increased by 2.2% [1]. Approximately 1.35 million accident type?
people die due to road accidents every year [2]. 3) How to automatically report the occurrence of road acci-
Manuscript received May 10, 2020; revised June 16, 2020; accepted dent with its type and location to the relevant agencies
June 30, 2020. Date of publication July 13, 2020; date of current version by sending the emergency notification if the accident
January 7, 2021. (Corresponding author: Nikhil Kumar.) victims are incapacitated?
Nikhil Kumar and Divya Lohani are with the Department of
Computer Science and Engineering, School of Engineering, Shiv Nadar Contribution of This Work: Research work on developing
University, Greater Noida 201314, India (e-mail: [email protected]; the mechanisms for prediction, prevention, detection, and man-
[email protected]). agement of road accidents are predominantly focused on either
Debopam Acharya is with the School of Computing, DIT University,
Dehradun 248009, India (e-mail: [email protected]). enhancing precision or reducing rescue time following the
Digital Object Identifier 10.1109/JIOT.2020.3008896 occurrence of road accidents. In this study, we have proposed
2327-4662 
c 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: VTU Consortium. Downloaded on October 18,2023 at 07:37:14 UTC from IEEE Xplore. Restrictions apply.
870 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 2, JANUARY 15, 2021

a smartphone-based end-to-end IoT system architecture to

detect and notify automotive accidents. The work also focuses
on identifying the best machine-learning ADC model based on
either of three different classifiers, namely, Gaussian mixture
model (GMM), decision tree (DT), and Naïve Bayes (NB).
This classification can be extremely helpful in the planning
and execution of rescue operations. The proposed solution is Fig. 1. Architecture of the proposed IoT system.
inexpensive and efficient that can be easily retrofitted in any
class and model of the vehicle. Experiments and analysis of using GPS when an accident is detected. Fernandes et al. [14]
the results have shown that in all proposed solutions NB-based have developed an Android application, which receives inputs
model is highly accurate in the detection and classification of from the smartphone sensors to detect collisions and rollovers,
road accidents. as well as receive road hazard warnings issued by vehicles in
This article is organized as follows. Major related work in the vicinity.
the area of automatic vehicle accident detection and multi- When a good Internet connection is limited to some
sensor fusion is described in Section II. The architecture of nearby vehicles, intervehicle communication approaches, such
the proposed IoT system, workflow diagram of the system, as vehicular ad hoc network (VANET) and Internet of
model variables, and preprocessing methods are explained in Vehicles (IoV), may help vehicles carry out reporting of
Section III. The hardware and software setup of the system is accidents via each other. Fabian et al. [15] have introduced
proposed in Section IV. Next, Section V describes the possible a programmable objective function (POF) and developed an
accident scenarios, and then there is a description of the exper- architecture inspired from a software-defined network (SDN)
imental setup and data collection in Section VI. Three different approach to improve the Quality of Service (QoS) in
ADC models, based on GMM, DT, and NB classifiers, are dynamic IoV networks, which can be very important
proposed in Section VII. Results, comparative analysis of while vehicle reporting the vehicle accidents to each other.
proposed models, and discussion are presented in Section VIII Rachedi and Badis [16] have proposed a programmable and
while conclusion and future work are laid out in Section IX. hybrid architecture based on the virtual backbone (VB) formed
by selected parked vehicles. Architecture is used to reduce the
use of cellular communication link in IoV, which is based on
II. R ELATED W ORK connected dominated sets (CDSs) and software-defined vehic-
Research work related to the detection, localization, report- ular networks (SDVNs) approaches. Mekki et al. [17] have
ing, modeling, and analysis of road accidents is discussed discussed a vehicular cloud access (VCA) problem that is
in this section. Several authors have reported smartphone- modeled as an evolutionary game, and proposed two VCA
based accident detection systems. Aloul et al. [8] developed algorithms based on the evolutionary game (EG-VCA), and
an Android application that uses the accelerometer data to distributed Q-learning (QL-VCA).
develop an accident detection model based on dynamic time Acharya et al. [18] and Ibrahim et al. [19] have
warping (DTW) and hidden Markov models (HMMs). An developed systems that used accelerometer and gyro-
SMS, which includes accident location and severity, is sent scope data to detect and report vehicle rollover events.
to the emergency medical service (EMS), police, and fam- Smolka and Skublewska-Paszkowska [20] presented
ily on detection of an accident. Bhatti et al. [9] have used a smartphone-based collision detection system using
built-in speed, location, pressure, sound, and g-force sensors accelerometer, magnetometer, and GPS module to reduce the
of a smartphone to develop a low cost, portable solution that number of false alarms. Sadeky et al. [21] have used real-time
detects and reports an accident to the nearest hospital. video footage of traffic surveillance systems to construct the
Shaik et al. [10] proposed an IoT-based vehicle colli- histogram of flow gradient (HFG). The HFG features are used
sion detection system that uses a ADXL345 accelerometer as inputs to develop the statistical logistic regression model
to detect the collision and GPS to detect and send the acci- to predict road accidents. None of these systems reports the
dent location to the nearest ambulance via Internet. Sharma ability to determine the type of accident, which can be useful
and Sebastian [11] proposed an algorithm for vehicle accident in planning the rescue type and estimating the number of
detection that uses ADXL345 accelerometer and vibration sen- emergency services required.
sor to detect the vehicle crash, and heart rate sensor to monitor The fusion of sensor data for detection and analysis of road
health of the passenger, and send an emergency SMS or voice accidents is not reported in any major work, which can help to
call to the family. An accelerometer and an ultrasonic sensor improve the ADC performance. Felisberto et al. [22] presented
are used by Priya et al. [12] in an AtMega162-microcontroller- a system, which used the fusion of data from a network of
based accident detection system to notify the nearest first aid wireless sensors to detect accidents in elderly people. The
center about the accident location via SMS. authors claimed that the proposed solution is nearly 100%
To reduce delay in reporting aroad accident, Dar et al. [13] accurate in detecting normal falls. Damousis and Tzovaras [23]
have proposed a fog computing-based approach to build a low- have used a fuzzy expert system for the fusion of the most
cost smartphone-based system. The system uses an Android significant features extracted from the eyelid activities of vehi-
application that takes inputs from smartphone sensors to detect cle drivers. The system is found to be efficient in predicting
an accident. The system locates and notifies a nearby hospital sleep onset and accidents. Zhang et al. [24] have attempted the

Authorized licensed use limited to: VTU Consortium. Downloaded on October 18,2023 at 07:37:14 UTC from IEEE Xplore. Restrictions apply.
KUMAR et al.: IoT-BASED VEHICLE ADC SYSTEM USING SENSOR FUSION 871

Fig. 2. Workflow diagram of the accident classification system.

fusion of multisource data by combining traffic metrics and model includes five parameters, namely, speed, absolute lin-
social media tweets for real-time detection of road accidents. ear acceleration (ALA), change-in-altitude, pitch, and roll, to
Using support vector machines (SVMs) as the classification detect and classify a vehicle accident. The system is trained
model and implementing fivefold cross-validation, the authors and a knowledge base having four accident classes is cre-
claim an increase in prediction accuracy by the integration of ated by providing the training-data set to the machine learning
these two data sources. classifier. After training, the system model is tested against the
Moulik and Majumdar [25] have developed an IoT system testing data set; and then accuracy is measured to ascertain the
that uses the fusion of multiple ultrasonic sensors. The system delivery of the valid notification to the emergency services, in
uses infrared transmitter–receiver pairs and a fuzzy inference case of an accident.
system to detect the accidental fall of humans. The multisen-
sor fusion is claimed to achieve an improvement of 16% as C. Model Variables
compared with existing approaches.
1) Speed: It is quite logical that when a car experi-
Although, previous works have proposed some systems for
ences a serious accident, whether it is a collision, a rollover
collision/rollover detection and notification, to the best of our
or a falloff, its speed would become zero ultimately.
knowledge there is no system that classifies these accidents as
Smartphone’s GPS has been used to measure the speed
collision, falloff, rollover, and no accident. This work presents
of the vehicle. Each GPS device receives National Marine
a novel smartphone sensor fusion-based solution for detection
Electronics Association (NMEA) sentence from satellites con-
and categorization of accident.
taining information related to the location, velocity, and time
of the device [27]. Each class of GPS devices does have its
III. A RCHITECTURE OF THE S YSTEM
own NMEA sentence. The most common NMEA sentence that
A. Architecture of Our Proposed IoT System is used by most of the Android devices is $GPRMC. Following
An IoT architecture is proposed in Fig. 1 to address the is an example of the $GPRMC NMEA sentence, where the
problem of vehicle accident classification. We have used velocity is 018.8 knots:
a modern smartphone and a prefabricated sensor platform
termed as Sensordrone [26] to obtain the values of dif- $GPRMC,152844,A,1831.335,N,07734.331,E
ferent physical parameters relevant to vehicle motion. The 018.8,054.4,030220,003.1,W*6A.
Android smartphone contains several integrated sensors, such
as accelerometer, gyroscope, magnetometer, GPS, etc., which 2) Absolute Linear Acceleration: Whenever a running vehi-
can be used to determine the speed, direction, rotation, and cle drop from a height, the orientation of the vehicle will not
g-force, etc., of the vehicle. The Sensordrone remains attached be the same as it was on the road before falling. In such a case,
to the smartphone via Bluetooth connection to send the data to the vehicle cannot remain parallel to the gravitational axes (X,
the smartphone. Most of the processing is handled by the Y, and Z) and deceleration due to the impact that can be dis-
smartphone. Processing in smartphone substantially decreases tributed in more than one axis. Therefore, it is very difficult
the Internet resource (e.g., bandwidth, etc.) usage by trans- to assess the precise value of the deceleration when it is dis-
mitting only relevant details, such as venue, name, nature of tributed in X, Y, and Z components. To address this problem,
the accident, etc., to the IoT server. The IoT server delivers ALA (or signal magnitude vector [28]) is calculated with the
emergency alerts to various emergency services, such as EMS, help of X, Y, and Z components of deceleration. ALA is
local police station, and fire service, and other receivers, such independent of the vehicle’s orientation during falloff or col-
as relatives, insurers, blood donor, towing service, etc., after lision directions (whether the vehicle is moving or steady).
assessing the situation. ALA shows the acceleration characteristic parallel to the direc-
tion of impact. This absolute quantity is a resultant vector of
X, Y, and Z components of deceleration. ALA is determined
B. Operating Workflow Diagram as
The operating flow diagram of the proposed ADC system 
is depicted in Fig. 2. The feature vector of our classification ALA = (DECX )2 + (DECY )2 + (DECZ )2 (1)
Authorized licensed use limited to: VTU Consortium. Downloaded on October 18,2023 at 07:37:14 UTC from IEEE Xplore. Restrictions apply.
872 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 2, JANUARY 15, 2021

precisely, but not azimuth. Collectively, all sensors can calcu-

late rotational speed and rotational angle, about all three axes.
At high frequency, short-term estimates of the combination of
the accelerometer and magnetometer are not accurate owing
to a few small external factors, except the g-force. A low-
pass filter (only makes low-frequency signals from 0 Hz to
Fig. 3. (Left) yaw, pitch, and roll; (right) complementary filter.
its threshold level) is required for correction because the com-
bination of both of these inertial sensors works well at low
where DECX , DECY , and DECZ are the decelerations at x-axis, frequency.
y-axis, and z-axis, respectively. Smartphone’s accelerometer Another sensor that can calculate the angular momentum
has been used to measure the ALA whose measurement unit about the x-, y- and z-axis is the gyroscope, but it is also
is g (≈9.80665 m/s2 ). It is mentioned in the literature that exposed to strong drift. Gyroscope operates well at a high rate,
when a vehicle collided with an object (or any solid surface) so high-pass filtering (allows only high-frequency signals from
at the speed of 23 km/h or more, the intensity of deceleration its threshold frequency) is required to fix the results because
always cross 5g [29]. the measurements start drifting in long run.
3) Change-in-Altitude in 1 s: Knowledge of altitude shift Each sensor has its pros and cons, but their fusion can
enables the system to predict the falloff of the vehicle. During improve the performance anyway [31]. In this research, a com-
the falloff, the altitude of the vehicle is the most accountable plementary filter [32] is used, consisting of both a low-pass
feature that shifts drastically. Although GPS can also be used filter and a high-pass filter [Fig. 3(right)]. The complemen-
to determine the altitude, GPS is not acceptable when the posi- tary filter provides data that are more precise by applying the
tion stays the same while the altitude shifts. No specialized weight fractions to the outputs of low-pass and high-pass fil-
instrument is present that can determine altitude solely, but it ters. If Agyro , Pgyro , and Rgyro are the corresponding azimuth,
can be determined using atmospheric pressure P and tempera- pitch, and roll provided by the high-pass filter, and Aacc_mag ,
ture. In this experiment, the altitude is calculated by using the Pacc_mag , and Racc_mag are the corresponding azimuth, pitch,
pressure P measured by the Sensordrone’s barometric altime- and roll provided by the low-pass filter, then the resultant
ter. The atmospheric temperature is neglected because it does azimuth, pitch, and roll determined by the complementary
not substantially change at same coordinates. If P0 is reference filter are as follows:
pressure recorded at the sea level, final formula of calculating
Azimuth = α ∗ Agyro + (1 − α) ∗ Aacc_mag (3)
the altitude would come out as provided by Ye et al. [30]
  Pitch = α ∗ Pgyro + (1 − α) ∗ Pacc_mag (4)
Altitude = 44330.77 ∗ 1 − (P/P0 )0.190263 . (2) Roll = α ∗ Rgyro + (1 − α) ∗ Racc_mag . (5)
4) Roll and Pitch: The rotation angles of an object around The gyroscope gives more stable value than accelerometer
the x- and y-axis are known as roll and pitch, respectively. Roll and magnetometer collectively, and weightage of gyroscope
and pitch are calculated by the fusion of rotation information must be near 0.95 while using complementary filter [33]. We
provided by the gyroscope, accelerometer, and magnetometer have tried to calibrate the combination for different values of
using a complementary filter. The rotation matrix rotates the α and find that this combination was showing smoother curve
Euclidean space counterclockwise around the center of the of pitch, roll, and azimuth at α = 0.98.
Cartesian coordinate system by angle θ . We found that if either 2) Maximum Deceleration Using 10-ms Moving Maximum:
roll or pitch is greater than 90◦ , a rollover happened. The The standard signal generation rate of an accelerometer might
longitudinal axis is always supposed to be x-axis and the lateral be more than 2000 Hz. It is quite challenging to process
axis is always supposed to be y-axis [Fig. 3(left)]. the data and report the maximum ALA value at such a high
frequency, especially when we need accurate decisions and
D. Preprocessing Methods our device has limited resources and processing capabilities.
1) Sensor Fusion Using Complementary Filter: The direc- Iyoda et al. [34] have used a 10-ms moving average technique
tion and orientation of the vehicle can be determined by pitch, to deal with this problem; however, the moving average can
roll, and azimuth (yaw). The longitudinal, lateral, and verti- diminish the ALA’s peak value. To attain the highest possi-
cal axes of the vehicle’s body are expressed by X, Y, and Z, ble ALA, we are using a 10-ms Moving Maximum method
respectively. Roll, pitch, and azimuth (yaw) are the rotating that registers the peak value for each and every 10-ms time
angles around the x-, y-, and z-axis, respectively [Fig. 3(left)]. window as . . . , S − 2, S − 1, S, S + 1, . . . (Fig. 4).
Three inertial sensors, viz., gyroscope, accelerometer, and
magnetometer have been used to determine the orientation of
IV. H ARDWARE AND S OFTWARE S ETUP OF THE S YSTEM
the vehicle.
Despite their respective properties, owing to their short- A. Hardware Setup
comings and drift in measurements, no stand-alone inertial In this work, a complete hardware setup comprises
sensor is adequate to calculate the orientation effectively. The a SAMSUNG Galaxy S8 Android smartphone, a Sensordrone,
magnetometer can measure azimuth precisely, but not pitch and a 1:12 scaled RC car. The hardware setup is depicted in
and roll, whereas accelerometer can calculate pitch and roll the inset figure of Fig. 6.

Authorized licensed use limited to: VTU Consortium. Downloaded on October 18,2023 at 07:37:14 UTC from IEEE Xplore. Restrictions apply.
KUMAR et al.: IoT-BASED VEHICLE ADC SYSTEM USING SENSOR FUSION 873

Fig. 4. 10-ms moving maximum.

TABLE I
RC C AR S PECIFICATIONS

Fig. 5. (Left) data collection; (center) user dashboard; (right) notification.

required to be installed on accident rescuers’ Android smart-

phone so that timely rescue operation can be initiated after
locating the place.
1) SNUSense Android Application: An Android applica-
tion, namely, SNUSense, has been developed for capturing
1) Samsung Galaxy S8 Android Smartphone: In this work, and analyzing various parameters, such as rotation angle (roll,
besides of external sensors, the SAMSUNG Galaxy S8 Android pitch, and azimuth), speed of the vehicle, noise, deceleration
smartphone’s inbuilt 9-axis inertial sensor and GPS sensor at different axes, atmospheric pressure, altitude, temperature,
have been used. Inertial sensor, whose measuring range is latitude, longitude, humidity, CO2 , CO, light intensity, etc.,
±179◦ and ±16g, is used to measure the pitch, roll, and by utilizing Android smartphone sensors and Sensordrone
deceleration, and a GPS sensor is used to determine the vehi- [Fig. 5(left)]. It continuously processes the streamed data com-
cle’s speed and location. 1.9-GHz clock speed of Exynos ing from the Sensordrone and smartphone sensors and makes
8895 microprocessor and 4 GB of RAM makes it a suitable an estimate using the vehicle’s five characteristics, that is,
computing device for our experiments. speed, ALA, change-in-altitude, pitch, and roll. After a posi-
2) Sensordrone: Sensordrone is a tiny sensor hub that tive detection of the accident event, SNUSense sends the name
includes seven separate in-house sensors to assess twelve of the victim, type, and location of the vehicle accident to
different environmental variables, such as temperature, humid- the SNUSense IoT server (Google Firebase [36]) using 4G
ity, CO, atmospheric pressure, light, proximity, etc., [26]. Internet connection. We can see in the user dashboard that in
Sensordrone is a programmable device, which has a UART case of false alarm, victim can abort the notification by press-
port to attach an additional sensor (e.g., CO2 sensor) and ing the STOP button [Fig. 5(center)]. IoT server further sends
Bluetooth to establish a connection with any processing the notification (using Firebase Cloud Messaging service) to
device. In this research, only barometric altimeter sensor (mea- the intended recipients, such as EMS, police, family, etc., after
surement range: 26kPa to 126kPa) of Sensordrone is used to analyzing the data (such as the severity of the accident based
determine the altitude of the vehicle. on the force of impact and the speed). It also collects all
3) RC Car: A 1:12 scaled toy Radio Controlled (RC) sensors’ data in a CSV (comma-separated values) file in the
Car [35] is used to mimic real-life accident scenarios since smartphone’s storage.
conducting the accident experiments with an actual car is 2) SNUAlertApp Android Application: Another Android
neither viable nor budget effective. This RC car is made application, SNUAlertApp, has been created to receive accident
of heavy-performance ABS material and has metallic shock alerts as shown in Fig. 5(right). The alert notification contains
absorbers. The configuration of RC car is specified in Table I. the coordinates of the accident, the name of the driver, and the
type and severity of the accident. A simple touch on the noti-
fication launches the Google Maps with a tracker icon on it
B. Software Setup so that rescuers can conveniently locate the place of accident.
Two Android applications, SNUSense and SNUAlertApp
have been developed for sending and receiving emergency V. ACCIDENT S CENARIOS
notifications. SNUSense is required to be installed on the acci- Depending on the five variables described in Section III-C,
dent victim’s smartphone to collect and process the sensor’s this research focuses only on the identification and classifica-
data. All the preprocessing methods, such as ALA calcula- tion of four different vehicle accident situations: 1) collision;
tion, complementary filter (sensor fusion), moving-maximum 2) falloff; 3) rollover; and 4) no accident. Table II describes the
function, and machine learning classifiers (feature fusion) are actions and criteria of five factors for each accident situation
programmed in the SNUSense application. SNUAlertApp is to be classified.

Authorized licensed use limited to: VTU Consortium. Downloaded on October 18,2023 at 07:37:14 UTC from IEEE Xplore. Restrictions apply.
874 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 2, JANUARY 15, 2021

TABLE II
ADC C RITERIA

Fig. 7. Collision.

cannot be foreseen, the deceleration measurement would be

more accurate if the direction of impact is same as the direc-
tion of the accelerometer sensor installation. Similarly, if the
Fig. 6. (Left) falloff setup; (center) rollover setup; (right) collision setup.
impact is on the opposite side of the sensor position, due
to the size and build of the vehicle, the deceleration output
may be less than the actual. The National Highway Traffic
Thresholds of speed and roll/pitch are identified by perform-
Safety Administration (NHTSA), the nodal agency of the U.S.
ing accident tests multiple times with our RC car setup. If any
Department of Transportation, recommends that inertial sen-
fatal accident occurs, the ideal speed of vehicle becomes zero
sors be placed near the CG [37], for all crash tests. As shown
instantly. Threshold for speed is considered 2 km/h because
in Fig. 6, we have also followed this recommendation for our
GPS readings does not stabilize immediately and slightly
experimental RC car setup and tied up the smartphone on the
fluctuates around for a while. In the literature, for rollover, dif-
chassis of the RC car. For the rest of the sensors used in our
ferent researchers have considered different angles of roll/pitch
experiment, their placement within a vehicle does not have
to define the rollover. In this work, we are considering the
a significant impact on the accuracy of their data values.
rollover for angles greater than or equal to 90◦ , as we are
All the tests were performed at the speed of 35.1 km/h
considering not partial but complete rollover events. For colli-
because, in the literature, minimum threshold speed to detect
sion, it is mentioned in the literature that if a vehicle collides
collision (or airbag deployment) of vehicle is ≈23 km/h [29].
with a solid obstacle with a speed of more than 23 km/h (i.e.,
It is considered that speed below 23 km/h will not be enough
21 ft/s) then it will generate a deceleration of more than 5g
to create a severe collision impact. As speed will increase,
and the collision is said to be severe [29]. From the follow-
the severity of accident will increase proportionally. Because
ing equation of velocity, we can observe that if a vehicle falls
this research does not focus on severity monitoring, con-
from 8 ft with an initial speed of 0, it achieves the speed of
sidering the speed similar to the minimum speed would be
more than 21 ft/s in almost 0.8 s, which is minimum speed to
sufficient to perform the accident tests. Another reason for
generate 5-g deceleration:
considering experimental speed 35.1 km/h is the RC car’s
velocityfinal = velocityinitial + g ∗ t. (6) limitations. Although RC car’s maximum speed is 53 km/h
(which cannot be set to a specific speed), but due to the
So, threshold of parameter, i.e., change-in-altitude, is consid- weights of smartphone and Sensordrone, it comes down to
ered a minimum 8 ft. Although there may be other subclas- 35.1 km/h.
sification of the accidents, this study is restricted to collision,
falloff, and rollover events. During the implementation of the
B. Data Collection
classification, the lag between the reading timings of different
parameters is well thought out. To collect the head-on collision data, we have used a plastic
road barricade as a barrier [Fig. 6(right)], and the RC car
setup is crashed into it. Fig. 7 is showing the change in ALA,
VI. E XPERIMENTAL S ETUP AND DATA C OLLECTION
speed, and change-in-altitude during collision event. After the
A. Experimental Setup collision, a sudden decrease in the speed and increase in the
To accurately detect collision and falloff, the smartphone ALA is clearly noticeable in the graph. The drawn red line
is placed near the center of gravity (CG) of the vehicle (in reflects the moment when the collision occurred.
our case, on the armrest between the front seats) as both The falloff experiment is conducted on the elevated run-
collision and falloff events leads to deceleration, and it is ning track of the Indoor Sports Complex (ISC) at Shiv Nadar
measured with the help of smartphone’s built-in accelerom- University, India [Fig. 6(left)]. The RC car setup was dropped
eter sensor. Although the direction of impact on the vehicle onto the wooden tennis court from the running track elevated

Authorized licensed use limited to: VTU Consortium. Downloaded on October 18,2023 at 07:37:14 UTC from IEEE Xplore. Restrictions apply.
KUMAR et al.: IoT-BASED VEHICLE ADC SYSTEM USING SENSOR FUSION 875

VII. ACCIDENT D ETECTION AND C LASSIFICATION

M ODELS
In this work, three different classification models, namely,
GMM (a clustering model), DT (a logic-based model), and
NB (a probabilistic model) have been compared to find best
learning model to address the vehicle’s ADC problem. Five
parameters of training data set, that is, change-in-altitude,
pitch, roll, speed, and ALA has been used to train every classi-
fication model. Separate testing data set of each accident event
has been used to assess the ADC performance of the models.
90% observations are used for training the models and 10%
observations are used for testing the models.

Fig. 8. Falloff.
A. Gaussian Mixture Model-Based ADC Model
GMMs are among the most statistically matured methods
for clustering and density estimation [38], [39]. They model
the probability density function (PDF) of observed data points
using a multivariate Gaussian mixture density. Mixture mod-
els are a form of density model that consists of several
components, usually Gaussian in nature. These functions of
components are combined to give a multimodal density. In
this work, GMMs are developed to capture the information
about each accident type, i.e., collision, rollover, and falloff.
The number of Gausses in the mixture model is known as
the number of components. They indicate the number of clus-
ters in which data points are to be distributed. The number of
components in each GMM is optimized based on the number
of training data points. The components within each GMM
capture finer level details among the feature vectors of each
type of accident. Depending on the number of data points,
Fig. 9. Rollover.
the number of components may be varied in each GMM. If
the GMM has a few components and is trained using a large
by 15-ft height by diverting it to the left side with a 45◦ angle number of data points, it may lead to more generalized clus-
on the track. Fig. 8 reflects the falloff case, with changes ters that fail to capture specific details related to each class.
in ALA, speed, and altitude. The gradual decrease in speed On the other hand, overfitting of the data points may happen,
is showing the duration of fall, 15 ft change-in-altitude (the if too many components represent a few data points. Also, the
difference between the altitude before the fall and after the complexity of the models increases if they contain a higher
fall), and sudden increase in ALA is showing the percussion number of components.
on the tennis court after the fall. The dotted red line indi- The decision regarding the category of accident is taken
cates the moment of the percussion. 1-s sliding window has based on its probability of coming from feature vectors of the
been used to record the change-in-altitude because in 1 s the specific model. Given a set of inputs, GMM refines the weights
falling object can attain the final speed of more than 30 km/h, of each distribution through the expectation–maximization
which is far enough to create the impact force of more algorithm. Once a model is generated, conditional probabil-
than 5 g. ities can be computed for test patterns (unknown data points).
To conduct the rollover experiments, we have used a 4-cm The ADC events are modeled using Gaussian PDFs, explained
thick wooden board that was inclined 30◦ with the flat ground by the mean vector and the covariance matrix. For a feature
surface [Fig. 6(center)]. Fig. 9 demonstrates changes in speed, vector xt , the mixture density for an accident type is defined
roll, and pitch angles during the rollover event. The sudden as the weighted sum of N component’s Gaussian densities as
increase in the roll angle is the perfect indicator of rollover 
N
incident. The red dotted line in the graph shows the instant of P(xt |) = Wi Pi (xt ) (7)
the rollover activity. i=1
Every sort of accident experiment is repeated 30 times to where Pi (xt ) is the component densities and wi are the weights.
collect the data and 1167 observations has been recorded. Each For a D-dimensional feature vector, each component density
plot in Figs. 7–9 are plotted using raw data samples received is a D-variate Gaussian function, i.e.,
from the sensors, which are showing multivariate time series   
to observe the trends of different model parameters during 1 − 12 (xt −mi ) −1
i (xt −mi )
Pi (xt ) = D 1
e (8)
different type of accidents. (2π ) 2 |Ri | 2

Authorized licensed use limited to: VTU Consortium. Downloaded on October 18,2023 at 07:37:14 UTC from IEEE Xplore. Restrictions apply.
876 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 2, JANUARY 15, 2021

Fig. 10. GMM workflow for accident classification.

where Ri is the covariance matrix for the ith component and Fig. 11. DT for accident classification.
mi is a mean vector. The mixture weights satisfy the constraint

N
where pi is the probability of taking a particular value and i
Wi = 1. (9)
is the number of options. The DT divides the input variable
i=1
M into a number of subsets: M1 , M2 , . . . , Mn , using a splitter.
Fig. 10 illustrates the overall methodology of a GMM- The expected unpredictability of the n outcomes of the input
based accident classification system. The process is divided variable M is measured by using EE, defined by the equation
into two parts: 1) feature extraction phase and 2) accident
detection/classification phase. In the first phase, training mod- 
n
ci
els produce accident classification models by using a feature EE = −pi log2 pi (11)
c
i=1
vector derived from the identified accident data set by feeding
the accident parameters to the GMM splitting and optimization where ci is the number of observations of the input vari-
block which uses the expectation–maximization algorithm. able in each subset M1 , M2 . . . , Mn , and c is total number of
In the second phase, the testing (assessment) of the trained observations in parent node M.
models is conducted using an unidentified accident data set. The information gain (I) is the difference between EE and
Features of all unidentified accidents are provided to all trained actual entropy E(M)
models. The models then calculate the probability of uniden-
tified vector features corresponding to a particular model. The I = E(M) − EE. (12)
most probable model is viewed as a potential accident case
for that feature vector. The node with information gain of 0 is treated as the
terminal node, which cannot be split further. A saturated tree
is obtained with the recursive application of previous steps.
B. Decision Tree-Based ADC Model
Five variables of our SNUSense database namely, speed, ALA,
DTs are prediction and classification tool with a tree-like roll, pitch, and change-in-altitude have been chosen as predic-
structure, where a test is performed on an attribute at each tor variables, and collision, rollover, falloff, and no accident
node, each branch holds a result of the test and each terminal have been chosen as response classes (Fig. 11). The input data
node contains the class label. DTs are best suited for classifi- set is divided into two data sets for training and testing. When
cation as they are strictly nonparametric, and does not require the DT model fits the training data, the recursive algorithm of
any information about the distribution of input parameters. the DT model goes on to split the input data until it ends up
DTs are capable of handling both numeric and categorical with pure sets. The testing data set is then used for testing the
inputs, modeling nonlinear relationships between features and performance of the trained model.
classes, allowing for missing values [40], [41]. DTs have an
intuitive appeal because of the classification structure being
explicit and easily interpretable. C. Naive Bayes-Based ADC Model
The first in the development of a DT is tree growth. All Another classification technique, NB, is used for ADC,
input data are concentrated in the root node in the beginning. which is based on Bayes’ theorem. NB method is highly scal-
The data set is then broken down into child nodes, with the able and scales linearly with the number of predictors [42].
application of a series of splitting variables (splitters). The DT It is favored over other classification techniques because of
algorithm measures the entropy, expected entropy (EE), and its computational simplicity and its ability to be trained very
the gain in information to decide if an input variable must be fast [43]. It is robust to noisy data. The technique is based on
selected as the splitter and if the node can be split further or the presumption that the predictor variables are independent,
not. Entropy is the measure of the amount of uncertainty of an i.e., a particular feature present in a class is not related to the
event. Let M is an input variable for accident classification, presence of other features.
with n distinct values. The entropy of the input variable is The NB model contains the multiple input (or predictor)
calculated by variables (change-in-altitude, pitch, roll, speed, and ALA) and

n target variables (types of accident) as model outputs. Let T
E(M) = −pi log2 pi (10) be the state or class of the target variable and vector X =
i=1 (x1 , x2 , . . . , xn ) be the states of n input features. To estimate

Authorized licensed use limited to: VTU Consortium. Downloaded on October 18,2023 at 07:37:14 UTC from IEEE Xplore. Restrictions apply.
KUMAR et al.: IoT-BASED VEHICLE ADC SYSTEM USING SENSOR FUSION 877

TABLE III
the value of T based on X, we need to calculate the conditional D ETECTION AND C LASSIFICATION P ERFORMANCE U SING NB
probability of T given X
p(X|T)p(T)
p(T|X) = (13)
p(X)
where, p(T) and p(X) are the constants that are directly derived
from the data. To obtain the value of p(X|T), it is factorized as
n TABLE IV
p(X|T) = p(x1 , x2 , . . . , xn |T) = p(xi |T). (14) D ETECTION AND C LASSIFICATION P ERFORMANCE U SING GMM
i=1
Combining (13) and (14), we get
n
p(T)
p(T|X) = p(xi |T). (15)
p(X)
i=1
The conditional distribution of T given X is calculated
from (15). The value of target variable p(T|X) is classification ROC curve for various threshold settings of the param-
outcome, which is the state of T with highest probability. eter. TPR [=TP/(TP + FN)] measures the proportion
of actual positive incidents that are correctly identified.
VIII. R ESULTS AND A NALYSIS The FPR [=FP/(FP + TN)] is measured as the ratio
The performance of the ADC models based on GMM, NB, between the number of falsely classified negative inci-
and DT classifiers are discussed in this section. Outcomes have dents as positive (false positives) and the overall number
been obtained using the SNUSense database. All 1167 exper- of TN incidents (irrespective of classification). TPR and
imental observations of the SNUSense database are shuffled FPR, both are different criteria which are compared in
using the random function, out of them 1050 (90%) obser- ROC curve. In our case, the area under the curve is an
vations are used for training the ADC models while 117 indicator of how well a function can distinguish a vehicle
(10%) are used for testing. By an observation here we mean accident type with the rest of the others.
a data point that is the vector of values read from the sen- The detection and classification performance of the NB
sors at a particular time during accident tests. The data set is technique is summarized in Table III. The mean of precision,
balanced and large enough to train and test the model. The recall, and F1-score for the NB-based classification model are
classification performance of the proposed models is assessed 0.94, 0.95, and 0.95, respectively. The model performs best
using precision, recall, F1-score, and receiver operating char- with collisions, where the F1-score is 0.97, which is followed
acter (ROC) curve [44]. These quantities are described as by a rollover, no accident, and falloff events.
following. Accident classification performance of GMM models for
1) Precision: It is the ratio of the number of true posi- the four accident classes using the SNUSense database has
tive (TP) predictions to all positive predictions. In this been computed. Using the 1050 observations and 16, 32, 64,
work, it represents the proportion of accident classifica- and 128 components for building each GMM, performances
tions that are actually right, i.e., are observed to be 89.75%, 87%, 95.5%, and 90.5%, respec-
tively. During the training period, it is observed that the highest
true_positive_samples accident classification performance of 95.50% is achieved with
precision =
positively_predicted_samples 64 components and 100 iterations. If more than 64 components
TP are used, the performance deteriorates.
= .
(TP + FP) Table IV is a summary of the ADC performance of the
2) Recall: It is the ratio of number of TPs to all the actual GMM technique on testing data. The average precision, recall,
positive records. It represents the proportion of accident and F1-score for the GMM-based model are observed to be
classifications that can be predicted by model, i.e., 0.91, 0.92, and 0.91, respectively. In the accident events, the
model performs best with collisions with an F1-score of 0.92,
true_positive_samples TP
recall = = . followed by falloff and rollover. The nonoccurrence of a road
actual_positive_samples (TP + FN) accident is reported with an F1 value of 0.94.
3) F1-Score: It takes both precision and recall into consid- Table V describes the detection and classification
eration. It is the harmonic mean of precision and recall, performance of the DT model. The model performs best with
i.e., rollover events with an F1-score of 0.92, followed by falloff
and collision. The mean precision, recall, and F1-score of the
precision ∗ recall
F1 − score = 2 ∗ . DT model are observed to be 0.88, 0.89, and 0.88, respectively.
precision + recall It is evident that the NB model is the best performing model
4) ROC Curve: The probability of detection, i.e., TP with an average F1-score of 0.95, which is followed by the
rate (TPR) is plotted as a function of the probabil- GMM and DT models with an F1-score of 0.91 and 0.88,
ity of false alarm, i.e., false-positive rate (FPR) in the respectively.

Authorized licensed use limited to: VTU Consortium. Downloaded on October 18,2023 at 07:37:14 UTC from IEEE Xplore. Restrictions apply.
878 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 2, JANUARY 15, 2021

TABLE V
D ETECTION AND C LASSIFICATION P ERFORMANCE U SING DT

Fig. 14. ROC curve for DT.

NB-based model is implemented in the SNUSense application.

We have used the “JNBC” Java library [45] for NB classifi-
cation, which is provided by NamSor.com. Although, every
ML algorithm is hard to implement in smartphones because
of their processing toll and availability of a suitable library. For
Fig. 12. ROC curve for NB.
this purpose, the concept of IoT server is used in the proposed
architecture to implement the ML algorithm on the server side
and to establish publish/subscribe communication.
In case of rollover and falloff accidents, RC cars can mimic
the real car accidents completely. As far as the intensity of
deceleration generated due to the impact is concerned, it may
differ due to the build, size, and weight of a real car, thus
deceleration threshold could be set accordingly while training
the system for real cars.

IX. C ONCLUSION
The proposed detection system works frequently whether
there is an accident or not and reports the incident to
predefined emergency services and family in case of an acci-
dent. The system classifies accidents into four classes, i.e.,
Fig. 13. ROC curve for GMM. collisions, rollovers, falloffs, and no accidents, so that the
best possible rescue operations can be undertaken. Five train-
ing variables namely, change-in-altitude, pitch, roll, speed, and
ROC curves of NB-, GMM-, and DT-based models are plot- ALA are used as input variables to train and test the system.
ted in Figs. 12–14, respectively. Actually, the ROC curve is The proposed ADC system uses smartphone sensors, and
a graphical plot that demonstrates a binary classifier system’s Sensordrone sensors to measure the values of model variables.
diagnostic capability. Because our problem is a multiclassifica- The system can be retrofitted in any type of vehicle.
tion problem, each ROC curve of a particular model is drawn As far as the accuracy of the proposed ADC system is
by using the One-versus-Rest scheme where each accident sce- concerned, three different GMM-, NB-, and DT-based classi-
nario is plotted against all other accident types, which makes it fication models are evaluated and compared to determine the
binary classification problem for ROC curves of every accident most accurate ADC model. The NB model is found to outper-
types. With the help of positive and negative outcomes of ADC form others with an average F1-score of 0.95. It is the most
models, we have used binary logistic regression and saved the accurate for collision with an F1-score of 0.97, followed by
prediction probabilities. Then for generating the multivariable rollover, no accident, and falloff events, respectively.
ROC curve for a particular accident class, saved prediction If someone wants to reduce the time of automatic notifi-
probabilities as the test variables are utilized. cation after the incident or compare the system with other
To train, test, and identify the best ADC model among NB-, automatic notification systems, every delay should be as short
GMM- and DT-based models, first the “scikit-learn” Python as possible whether it is algorithm’s execution time or time
library and SNUSense data set are used on a separate com- taken in notification. But this work focuses on reducing over-
puter. After identifying that the NB-based model is the best all reporting time using technology, compared to just manual
ML model over GMM and DT for our ADC problem, the reporting or nonreporting of an incident. To achieve this, we

Authorized licensed use limited to: VTU Consortium. Downloaded on October 18,2023 at 07:37:14 UTC from IEEE Xplore. Restrictions apply.
KUMAR et al.: IoT-BASED VEHICLE ADC SYSTEM USING SENSOR FUSION 879

have tried to reduce the time taken for notification after acci- [9] F. Bhatti, M. A. Shah, C. Maple, and S. Ul Islam, “A novel
dent detection by implementing machine learning algorithms Internet of Things-enabled accident detection and reporting system
for smart city environments,” Sensors, vol. 19, no. 9, p. 2071, 2019,
on the smartphone itself rather than transmitting data of all doi: 10.3390/s19092071.
sensors after every 10 ms. [10] A. Shaik et al., “Smart car: An IoT based accident detection system,”
The system is low cost in comparison to factory fitted in Proc. IEEE Global Conf. Internet Things (GCIoT), 2019, pp. 1–5,
doi: 10.1109/GCIoT.2018.8620131.
systems because they substantially increase the costs of the
[11] S. Sharma and S. Sebastian, “IoT based car accident detection and notifi-
cars. According to [46], there were more than 3.2 billion cation algorithm for general road accidents,” Int. J. Elect. Comput. Eng.,
smartphone users in the world in 2019. Hence, we have vol. 5, no. 3, pp. 4424–4430, 2017, doi: 10.11591/ijece.v9i5.4020-4026.
used a smartphone and its built-in sensors to develop our [12] E. K. Priya et al., “IoT based vehicle tracking and accident detec-
tion system,” Int. J. Innov. Res. Comput. Commun. Eng., vol. 5, no. 3,
system. Apart from the smartphone, only a barometric altime- pp. 4020–4026, 2017.
ter and a 4G Internet connection are needed to implement the [13] B. K. Dar, M. A. Shah, S. U. Islam, C. Maple, S. Mussadiq,
system. and S. Khan, “Delay-aware accident detection and response system
To the best of our knowledge, our system is the only one- using fog computing,” IEEE Access, vol. 7, pp. 70975–70985, 2019,
doi: 10.1109/ACCESS.2019.2910862.
of-a-kind system that classifies accidents as collision, rollover, [14] B. Fernandes, M. Alam, V. Gomes, J. Ferreira, and A. Oliveira,
falloff, and no accident. Although system is highly accurate “Automatic accident detection with multi-modal alert system imple-
and have several advantages over other systems, it also has mentation for ITS,” Veh. Commun., vol. 3, pp. 1–11, Jan. 2016,
doi: 10.1016/j.vehcom.2015.11.001.
some limitations, i.e.: 1) system needs continuous Internet [15] P. Fabian, A. Rachedi, and C. Guéguen, “Programmable objective
connection to send the emergency alerts; 2) placement of the function for data transportation in the Internet of Vehicles,” Trans.
smartphone would be predefined and user cannot put his smart- Emerg. Telecommun. Technol., vol. 31, no. 5, 2020, Art. no. e3882,
doi: 10.1002/ett.3882.
phone at any other place, such as pocket, bag, etc., as otherwise
[16] A. Rachedi and H. Badis, “BadZak: An hybrid architecture based
chances of FP (in case of no accident) and FN (in case of on virtual backbone and software defined network for Internet
an accident) would be very high; and 3) if smartphone gets of Vehicles,” in Proc. IEEE Int. Conf. Commun., 2018, pp. 1–7,
ejected outside the vehicle, or hardware setup breakdowns, doi: 10.1109/ICC.2018.8422759.
[17] T. Mekki, I. Jabri, A. Rachedi, and M. Ben Jemaa, “Vehicular
results may be impacted, and system may fail. cloud networking: Evolutionary game with reinforcement learning-based
As a future work, we intend to add another accident class, access approach,” Int. J. Bio Inspired Comput., vol. 13, no. 1, pp. 45–58,
i.e., fire/explosion, to our ADC model. To identify the most 2019, doi: 10.1504/IJBIC.2019.097730.
appropriate classification model for our ADC system, we are [18] D. Acharya, V. Kumar, and G. M. Gaddis, “A mobile system for
detecting and notifying vehicle rollover events,” in Proc. 15th Int.
going to implement some validation schemes, such as K-fold Conf. Adv. Comput. Commun. (ADCOM), Dec. 2007, pp. 268–275,
cross validation, and we will also consider the time taken by doi: 10.1109/ADCOM.2007.102.
the classification algorithm. In this work, both Android apps [19] H. A. Ibrahim, A. K. Aly, and B. H. Far, “A system for vehicle collision
and rollover detection,” in Proc. Can. Conf. Elect. Comput. Eng., 2016,
are function oriented, but in future we are also going to merge pp. 1–6, doi: 10.1109/CCECE.2016.7726720.
them in a single app having a user-friendly and interactive [20] J. Smolka and M. Skublewska-Paszkowska, “A method for collision
interface to make a unified solution for our ADC system. detection using mobile devices,” in Proc. 9th Int. Conf. Human Syst.
Interact. (HSI), 2016, pp. 126–132, doi: 10.1109/HSI.2016.7529620.
[21] S. Sadeky, A. Al-Hamadi, B. Michaelis, and U. Sayed, “Real-time auto-
R EFERENCES matic traffic accident recognition using HFG,” in Proc. Int. Conf. Pattern
Recognit., 2010, pp. 3348–3351, doi: 10.1109/ICPR.2010.817.
[1] World Health Statistics 2014: A Wealth of Information of Global [22] F. Felisberto, F. Fdez-Riverola, and A. Pereira, “A ubiquitous and
Public Health, WHO, Geneva, Switzerland, 2014. [Online]. Available: low-cost solution for movement monitoring and accident detection
https://apps.who.int/iris/handle/10665/112739 based on sensor fusion,” Sensors, vol. 14, no. 5, pp. 8961–8983, 2014,
[2] Global Status Report on Road Safety 2018: Summary, WHO, doi: 10.3390/s140508961.
Geneva, Switzerland, 2018. [Online]. Available: https://www.who.int/ [23] I. G. Damousis and D. Tzovaras, “Fuzzy fusion of eyelid activ-
violence_injury_prevention/road_safety_status/2018/English-Summary- ity indicators for hypovigilance-related accident prediction,” IEEE
GSRRS2018.pdf Trans. Intell. Transp. Syst., vol. 9, no. 3, pp. 491–500, Sep. 2008,
[3] E. B. Lerner and R. M. Moscati, “The golden hour: Scientific fact or doi: 10.1109/TITS.2008.928241.
medical ‘urban legend’?” Acad. Emerg. Med., vol. 8, no. 7, pp. 758–760, [24] Z. Zhang, Q. He, J. Gao, and M. Ni, “A deep learning approach for
2001, doi: 10.1111/j.1553-2712.2001.tb00201.x. detecting traffic accidents from social media data,” 2018. [Online].
[4] R. Sánchez-Mangas, A. García-Ferrrer, A. De Juan, and A. M. Arroyo, Available: arXiv.abs/1801.0.
“The probability of death in road traffic accidents. How important [25] S. Moulik and S. Majumdar, “FallSense: An automatic fall detec-
is a quick medical response?” Accident Anal. Prev., vol. 42, no. 4, tion and alarm generation system in IoT-enabled environment,”
pp. 1048–1056, 2010, doi: 10.1016/j.aap.2009.12.012. IEEE Sens. J., vol. 19, no. 19, pp. 8452–8459, Oct. 2019,
[5] M. A. Razzaque, M. Milojevic-Jevric, A. Palade, and S. Cla, doi: 10.1109/JSEN.2018.2880739.
“Middleware for Internet of Things: A survey,” IEEE Internet Things J.,
[26] D. Lohani and D. Acharya, “Real time in-vehicle air quality monitoring
vol. 3, no. 1, pp. 70–95, Feb. 2016, doi: 10.1109/JIOT.2015.2498900.
using mobile sensing,” in Proc. IEEE Annu. India Conf. (INDICON),
[6] G. Ponte, G. A. Ryan, and R. W. G. Anderson, “An estimate of the effec-
2017, pp. 1–6, doi: 10.1109/INDICON.2016.7839099.
tiveness of an in-vehicle automatic collision notification system in reduc-
ing road crash fatalities in South Australia,” Traffic Inj. Prev., vol. 17, [27] NMEA 0183—Standard for Interfacing Marine Electronic Devices, Nat.
no. 3, pp. 258–263, 2016, doi: 10.1080/15389588.2015.1060556. Marine Electron. Assoc., Middletown, RI, USA, 2002.
[7] Y. Chung and W. W. Recker, “A methodological approach for estimating [28] Y. Lee, H. Yeh, K. H. Kim, and O. Choi, “A real-time fall detection
temporal and spatial extent of delays caused by freeway accidents,” IEEE system based on the acceleration sensor of smartphone,” Int. J. Eng. Bus.
Trans. Intell. Transp. Syst., vol. 13, no. 3, pp. 1454–1461, Sep. 2012, Manag., vol. 10, pp. 1–8, Jan. 2018, doi: 10.1177/1847979017750669.
doi: 10.1109/tits.2012.2190282. [29] J. Kendall and K. A. Solomon, “Airbag deployment criteria: Institute of
[8] F. Aloul, I. Zualkernan, R. Abu-Salma, H. Al-Ali, and Risk and Safety Analyses,” Forensic Exam., pp. 1–21, 2014.
M. Al-Merri, “IBump: Smartphone application to detect car [30] H. Ye, K. Dong, and T. Gu, “HiMeter: Telling you the height
accidents,” Comput. Elect. Eng., vol. 43, pp. 66–75, Apr. 2015, rather than the altitude,” Sensors, vol. 18, no. 6, p. 1712, 2018,
doi: 10.1016/j.compeleceng.2015.03.003. doi: 10.3390/s18061712.

Authorized licensed use limited to: VTU Consortium. Downloaded on October 18,2023 at 07:37:14 UTC from IEEE Xplore. Restrictions apply.
880 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 2, JANUARY 15, 2021

[31] N. Kumar, A. Barthwal, D. Acharya, and D. Lohani, “Modeling IoT Nikhil Kumar is currently pursuing the Ph.D.
enabled automotive system for accident detection and classification,” in degree in computer science and engineering with
Proc. IEEE Sensors Appl. Symp. (SAS), Kuala Lumpur, Malaysia, 2020, Shiv Nadar University, Greater Noida, India.
pp. 1–6. His research interests are context-aware comput-
[32] V. Kubelka and M. Reinstein, “Complementary filtering approach to ori- ing, wireless sensor networks, vehicular accident
entation estimation using inertial sensors only,” in Proc. IEEE Int. Conf. detection and classification modeling, and Internet
Robot. Autom., 2012, pp. 599–605, doi: 10.1109/ICRA.2012.6224564. of Things.
[33] S. Hajdu, S. T. Brassai, and I. Szekely, “Complementary fil-
ter based sensor fusion on FPGA platforms,” in Proc. Int.
Conf. Optim. Elect. Electron. Equip. (OPTIM) Int. Aegean
Conf. Elect. Mach. Power Electron. (ACEMP), 2017, pp. 1–50,
doi: 10.1109/OPTIM.2017.7975076.
[34] M. Iyoda, T. Trisdale, R. Sherony, D. Mikat, and W. Rose, “Event
data recorder (EDR) developed by Toyota Motor Corporation,”
SAE Int. J. Transp. Saftey, vol. 4, no. 1, pp. 187–201, 2016,
doi: 10.4271/2016-01-1495.
[35] GP TOYS. (2020). Foxx S911 RC Truck. Accessed: Apr. 5, 2020.
[Online]. Available: https://g-p.hk/gptoys-foxx-s911.html
[36] Google Developers. (2020). Google Firebase. Accessed: Apr. 5, 2020.
[Online]. Available: https://firebase.google.com/ Debopam Acharya received the Ph.D. degree from
[37] “Crashworthiness research of prototype hydrogen fuel cell vehicle: Task the University of Missouri–Kansas City, Kansas
order 7 report,” Batelle Memorial Inst., Washington, DC, USA, Rep. City, MO, USA, in 2006.
DOT HS 812 112, 2015. He is currently a Professor and the Dean of the
[38] K. S. Rao and S. G. Koolagudi, “Robust emotion recognition using School of Computing, DIT University, Dehradun,
speaking rate features,” in Robust Emotion Recognition Using Spectral India. He has worked with the Location-Based
and Prosodic Features (Springer Briefs in Electrical and Computer Services Group, Garmin Inc., Olathe, KS, USA,
Engineering). New York, NY, USA: Springer, 2013, pp. 85–94, and tenure track Faculty with Georgia Southern
doi: 10.1007/978-1-4614-6360-3_5. University, Statesboro, GA, USA. His research has
[39] Y. Sun and B. Yuan, “Hierarchical GMM to handle sharp changes in been funded by several public and private sec-
moving object detection,” Electron. Lett., vol. 40, no. 13, pp. 801–802, tor organizations like BlackBerry, Dell, and the
Jun. 2014, doi: 10.1049/el:20040552. Ministry of Electronics and IT, Government of India. His research interests
[40] Z. Zheng, P. Lu, and D. Tolliver, “Decision tree approach to accident are in the areas of Internet of Things, mobile sensing, and smart environment.
prediction for highway-rail grade crossings: Empirical analysis,” Transp.
Res. Rec., vol. 2545, no. 1, pp. 115–122, 2016, doi: 10.3141/2545-12.
[41] J. Zheng, S. Yang, X. Wang, X. Xia, Y. Xiao, and T. Li, “A deci-
sion tree based road recognition approach using roadside fixed
3D LiDAR sensors,” IEEE Access, vol. 7, pp. 53878–53890, 2019,
doi: 10.1109/ACCESS.2019.2912581.
[42] T. Šingliar and M. Hauskrecht, “Learning to detect incidents from
noisily labeled data,” Mach. Learn., vol. 79, pp. 335–354, Sep. 2009,
doi: 10.1007/s10994-009-5141-7.
[43] B. Yang, Y. Lei, and B. Yan, “Distributed multi-human location algo-
rithm using Naive Bayes classifier for a binary pyroelectric infrared
sensor tracking system,” IEEE Sens. J., vol. 16, no. 1, pp. 216–223,
Divya Lohani (Member, IEEE) received the Ph.D.
Jan. 2016, doi: 10.1109/JSEN.2015.2477540.
degree from the Indian Institute of Information
[44] Z. Omary and F. Mtenzi, “Machine learning approach to iden-
Technology Allahabad, Allahabad, India, in 2016.
tifying the dataset threshold for the performance estimators
She is serving as an Assistant Professor with the
in supervised learning,” Int. J. Infonomics, to be published,
Department of Computer Science and Engineering,
doi: 10.20533/iji.1742.4712.2010.0034.
Shiv Nadar University, Greater Noida, India. Her
[45] E. Carsenat. (2020). Java-Naïve-Bayes-Classifier-JNBC. [Online].
research interests include Internet of Things, sen-
Available: https://github.com/namsor/Java-Naive-Bayes-Classifier-JNBC
sor networks, sensor data analytics, context-aware
[46] S. O’Dea. (2020). Smartphone Users Worldwide 2016–2021. Accessed:
computing, mobile computing, and cybersecurity.
Apr. 5, 2020. [Online]. Available: https://www.statista.com/statistics/
330695/number-of-smartphone-users-worldwide/

Authorized licensed use limited to: VTU Consortium. Downloaded on October 18,2023 at 07:37:14 UTC from IEEE Xplore. Restrictions apply.