Journal of Intelligent & Fuzzy Systems 38 (2020) 4591–4601 4591

DOI:10.3233/JIFS-191375
IOS Press

Accident risk prediction and avoidance

in intelligent semi-autonomous vehicles
based on road safety data and driver
biological behaviours1
Shabir Ahmad∗ , Faisal Jamil2 , Azimbek Khudoyberdiev3 and DoHyeun Kim∗
Department of Computer Engineering, Jeju National University, Jeju, Korea

Abstract. Autonomous vehicles technology is an emerging area and has attracted lots of recognition in recent times. Accidents-
free driving has always been the focal point of autonomous vehicles. Autonomous vehicles have the potential to eliminate
human errors while driving, which has been argued as the predominant cause of traffic accidents. In autonomous vehicles
technologies, a variety of efforts have been made to eliminate human drivers. The full elimination of humans is not possible
at this moment, but some of the tasks can be automated to facilitate the drivers. In this paper, we investigate the leading causes
of accidents based on UK vehicle safety data of 2017-2018. We analyze the data and investigate the leading factors which
cause traffic crashes. Based on the leading features in the dataset, we then run different prediction algorithms to predict the
severity of accidents under a given input feature set. The accuracy of the model with Decision Tree classifier, Random Forest,
and Logistic Regression are compared, and it has been found that Random Forest performs best among others with 95%
accuracy. The trained random forest model is deployed on the Internet of Things server based on Arduino, and a lightweight
application is developed to get the vital data from the driver. The data is applied to the trained model to predict the risk index
of driving. This application is lightweight but yet provide a significant contribution in terms of safety in autonomous vehicles.

Keywords: Internet of things, cyber-physical systems, real-time systems, virtualisation, virtual object

1. Introduction

Internet of Things (IoT) and smart spaces are an

exciting research topic, and many studies have been
1 This research was supported by Energy Cloud R&D Pro- devoted to allow context-aware application for sus-
gram through the National Research Foundation of Korea (NRF) tainable development of the society [1–4]. Numerous
funded by the Ministry of Science, ICT (2019M3F2A1073387), sensors and actuators form a network in which the
and this research was supported by Institute for Information & information is communicated and analyzed to form a
communications Technology Planning & Evaluation (IITP) grant whole new paradigm known as the IoT.
funded by the Korea government (MSIT) (No. 2019-0-01456,
The advancement in IoT has further paved the
AutoMaTa: Autonomous Management framework based on arti-
ficial intelligent Technology for adaptive and disposable IoT). way for many technologies to embrace it and orches-
Any correspondence related to this paper should be addressed to trate the processes by leveraging human as well as
Dohyeun Kim. things capabilities. [5–8]. For instance, conventional
2 E-mail: [email protected]
real-time systems are now becoming IoT-empowered
3 E-mail: [email protected]
real-time systems (RT-IoT) with much more capabil-
∗ Corresponding authors. Shabir Ahmad and DoHyeun Kim,
ities that come with IoT [9, 10]. One such evolution
Department of Computer Engineering, Jeju National University,
Jeju, Korea. E-mail: [email protected]; ORIC ID 0000-
is the efforts towards a smart transportation sys-
0002-8788-2717 (Shabir Ahmad); E-mail: [email protected] tem. Autonomous vehicles and electric vehicles are
(DoHyeun Kim). introduced to encourage risk-free and green trans-

ISSN 1064-1246/20/$35.00 © 2020 – IOS Press and the authors. All rights reserved
4592 S. Ahmad et al. / Accident risk prediction in semi-autonomous vehicles

portation systems, respectively. In the former case, The remainder of this paper is organized as follows.
the motivation is the reduction of fatalities and acci- In the next Section, the motivation and parameter
dents, which are mostly caused by drivers’ errors. It is prioritization are discussed, followed by the method-
envisioned that more than 90% of traffic crashes are ology, interaction modeling, implementation stack,
due to the drivers’ errors, such as fatigue, and thus, and execution results. Finally, the model is evaluated
their elimination can reduce the chances of accidents with respect to different parameters, and the paper
near to 10%. is concluded with a summary of key findings and
Nevertheless, the reduction of human intervention directions for future research.
is the sole aim of autonomous, which is very reward-
ing. However, at the same time, it can lead to catas-
trophic hazards if any failure occurs. AV testing and 2. Motivation
piloting have begun in various countries. By 2014,
AV testing on roadways has been legalized in four In this Section, the motivation of this paper is dis-
states in the US. In Australia, AV testing has been first cussed in detail. The primary impulse which leads to
introduced in South Australia’s roads in 2016 [11]. this work is the fact that accidents and traffic crash
The market penetration rate of AVs is estimated to be not only result in human losses but also a considerable
between 24% and 87% by 2045 [12, 13]. However, portion of economic resources are wasted in handling
no vehicle has passed level 5, which corresponds to a these crashes. For instance, in the UK, the amount
fully autonomous vehicle. Although Tesla claims that spent on avoiding fatalities is about 1.25 million in
they will introduce level 5 vehicles by 2020, it is still 2002 despite the fact that it is among the leading safest
far from reality. So, in order to contribute to the major- countries around the globe [16]. It means the money
ity of traffic crashes, automakers introduced some spent in other countries are way more than the UK.
automated features such as self-parking, automatic Similarly, in the United States, about 1.6% of annual
braking, forward collision warning, lane departure GDP is spent on handling traffic crashes, which con-
warning, and blind-spot monitoring [14, 15]. tributed around 838 billion USD in 2010, and this
The problem with these features is that they are keeps growing every year [17]. Fig. 1 shows the cost
very sophisticated at times and can facilitate driver of all individual crashes due to driver’s errors and
but near to 100% accuracy, but for every feature, external environment for drivers, which contributes
single embedded hardware is equipped with the a significant portion of the economy in handling and
vehicle, which is not only expensive but also put a avoidance crashes. Therefore, the goal of this paper
load the battery of the vehicle as well. In this paper, is to make use of newer technologies such as IoT
we model the safety of vehicle using a completely and RT-IoT to make a generic system that not only
different perspective and assume the collection avoids the traffic crashes but also contribute a sig-
of real-time tasks which must be executed within nificant deal to the economy. Other similar systems
their deadline. Unlike dedicated and expensive are accurate but not cost-effective and specific with
hardware, we use a single generic IoT hardware that a single role so multiple chips are needed to pro-
communicates with sensors to detect the driver’s vide a complete system which not only increase the
state and based the state, the risk index is predicted. cost but also the load on the vehicle. The proposed
We prioritize the tasks based on the dataset of UK system is generic and cost-effective and only a sin-
road safety, which is taken for 2017-2018. The gle general-purpose hardware need to be installed.
dataset is analyzed, and the vital factors affecting the This is the leading advantage of this work over the
safety of vehicles are investigated. Based on vital state-of-the-art solutions.
features, the data is trained to predict the risk index As part of the work, we prioritize specific param-
for a certain set of features. Furthermore, the trained eters that affect the most on road safety. Certain
model is deployed, which takes real-time data and parameters are related to the environment, such as bad
gives it to the deployed model. The model computes weather, surface friction, whereas certain parameters
the risk index. If the risk index is high, actions such are related to drivers’ biological behaviors. In this
as alerts and notifications are issued to the driver. work, we consider the driver’s biological behavior
The assurance of real-time tasks execution can be and environmental conditions which can make them
formally proved by carrying out feasibility analysis. uncomfortable and error-prune, which is considered
Once it has been guaranteed than no critical task will the leading factor for traffic crashes. Artificial intelli-
be missed, it will be deployed in a real environment. gence and Machine learning are used to find patterns
 S. Ahmad et al. / Accident risk prediction in semi-autonomous vehicles 4593

Fig. 1. Statistics of Economic Contribution towards handling traffic crashes.

Fig. 3. Selected features with the leading correlation with accident

risk.

with the random forest as compared to other coun-

Fig. 2. System Model for Risk Index Prediction based on Trained terparts algorithms. The output is a risk index which
Model and Drivers Biological Data. computes the accident chance and their severity. Fig.
3 shows the top 5 factors having the strongest correla-
in historical data. Different factors are contributing tion with accident severity. It has been clear that two
to road safety according to recent research studies. of the factors are related to the driver’s state, whereas
In order to systematically investigate the parameters the topmost culprit for accidents is the weather con-
which cater the most for road hazards and accidents, ditions.
we analyze the standard dataset of UK road safety
for the year 2017-2018 [18]. The data are collected
by the UK police forces from the occurrence of the 3. Methodology
accident. The data field contains records for accident
severity, weather, speed, age of the driver, drowsiness In this paper, we consider three different aspects
of the driver, number of fatalities, and vehicle type. to model a semi-autonomous vehicle, which is safe
Fig. 2 shows the architecture for data analysis and for users, roads, and other vehicles on the road. The
parameter prioritization. primary motivation of this work is the task modeling
The model accuracy is 95.67 based on Random for- of all the scenarios, which can potentially lead to haz-
est, which on analysis further revealed that the dataset ardous situations. Based on these scenarios, actions
features which are preprocessed have a better match are performed to avoid such hazardous situations. The
4594 S. Ahmad et al. / Accident risk prediction in semi-autonomous vehicles

Fig. 4. Methodology of Proposed Safety Modeling based on Task Management.

methodology of the system is exhibited in Fig. 4. 3.1. Task modeling

First, we generate tasks based on two scenarios, i.e.,
environment monitoring and driver biological moni- IoT applications modeling on a task level provides
toring. For these scenarios, tasks are generated. Tasks flexibility and a process-aware approach. One bene-
for the first scenario are getting temperature, get- fit of this is the involvement of the stack holders in
ting humidity, getting light conditions, and getting the process. In this work, we also consider modeling
weather conditions, while for the following scenar- a safe semi-autonomous vehicle based on real-time
ios, tasks, namely getting ECG, getting a pulse, and system theory of tasks. A task is said to be a sin-
getting oxygen saturation, are generated. These tasks gle function or a combination of function, which is
are mapped on their respective sensors based on the performed on sensors and actuators. Thus, in the IoT
correlation with the sensor. These tasks mapping are context, a task can be classified into two broad cate-
stored in a central repository from which it has been gories; sensor tasks and actuator tasks. Sensors tasks
consumed by IoT devices. In this setup, Arduino is are commonly periodic tasks, while actuator tasks are
used for monitoring driver biological behavior such control tasks that are triggered on the occurrence of
as drowsiness and fatigue, while Raspberry PI is some events. In IoT systems, tasks are not different
employed to tackle the environmental conditions. The from conventional real-time systems; thus, the design
contextual data sensed from sensors are published of tasks is influenced by conventional tasks in a real-
to the central repository from where it is consumed time system. A task can have attributes such as start
by intelligent control. Intelligent control checks the time, execution time, type, urgency. These attributes
values of the contextual data to predict the risk of acci- are summarized in Table 1. A task τ is mapped on a
dents. If the risk is above some predefined threshold, virtual object Vo to perform some action on a sensor
it triggers control tasks. The control tasks are also or actuator commonly called an IoT node.
mapped on the corresponding actuators and follow One advantage of vehicle modeling on task level is
the same pattern. Eventually, the tasks are deployed also to have a flexible approach to model the energy
on actuators connected with IoT devices. consumption and the power consumption and to
 S. Ahmad et al. / Accident risk prediction in semi-autonomous vehicles 4595

Table 1 Table 2
Task Modelling and Parameters Description Tasks Generation and Mapping in Electric Vehicle Safety
Modeling
Task Description Applicability
Parameter Task ID Task Name Vo IDs
Start time The arrival of tasks, also called Sensors, 1 GetTemperatureData Temperature 01
or arrival phase of the tasks. Actuators 2 GetHumidity Humidity BMP 01
time 3 GetPressure Pressure BMP 01
Execution The time it requires to execute Sensors, 4 GetLightCondition Light Sensor
time during its period in other words, Actuators 5 getWeather Weather Server
The execution time within its 4 GetECG Electro Cardio Gram
period. (ECG)-e-Health Libillium
Period The time after which the tasks Sensors, 5 GetPulse Pulse-e-Health Libillium
repeat itself. This parameter is Actuators 6 GetOximeter Oximeter e-Health Libillium
specific to periodic tasks only and 7 FindIrregularPattern All Vo
for sporadic tasks it is irrelevant. 8 ControlFan Fan Actuator
Type This indicate the type of task, N/A
whether the task is periodic or
event-driven task
where V is the voltage supplied to the node and R
Urgency Urgency is the quantitative Sensors,
measure of the priority of the Actuators is resistance induced to prevent from burning due to
tasks. 0 means Normal Periodic, excessive current. The total power is given by
1 means highly urgent Periodic 2
means normal event-driven and 3
means high urgent event-driven 
i=n
task. total
ψm = ψi,m (3)
Deadline The maximum time limit in Sensors,
i=o
which the task must have to be Actuators
executed otherwise will be The power and energy must be within the permis-
considered invalid. For event
tasks the deadline is considered 0.
sible range to avoid any circuit breakage, which can
Sensing Every task is inherently Sensors cause hazards in many cases. In this paper, tasks con-
Output associated with a physical device. sidered vital for safety are summarized along with
Be its sensor or actuator, in either their mapped virtual objects in Table 2.
case it generates some data which
can be stored as a task’s
parameters.
Energy Every task is when deployed on Sensors 4. Interactive modeling
Consumed an IoT node, consume some
energy which depends on the
amount of data it is shipped with In this Section, the interaction and flow of the
and the period of the task. system have been described in detail. The proposed
system first analyzes the safety scenarios, which are
described in a service requirement. Once the scenar-
optimize it accordingly. For instance, if a task τi with ios are analyzed, the tasks are generated for them.
a period P is deployed on a physical object vm , the These tasks are then mapped on virtual objects and
energy consumed is given by allocated to the corresponding IoT node. The data
collected from sensor nodes are persisted in a central
 φ+P repository from where it is consumed by a module
Dm
Ei,m = dt (1) named intelligent control. The module checks the val-
t=φ P ues of the sensors’ data and classifies them as normal
where Dm is the output data from the node vm , and φ is or abnormal. In the case of abnormal values, some
the arrival time of the task. The formula suggests that actions need to be taken, which is in the form of con-
the longer the period and lesser the data, the little will trol task generation and deployment. For instance,
be the energy consumed and vice versa. Similarly, the if the ECG values suggest that the driver is tired or
power is directly proportional to the energy, and for drowsy, a control task with the sound notification is
the same task, the power ψi,m is given by generated and deployed on the corresponding actua-
tor. The flowchart of this process is shown in Fig. 5
 to portray the above-mentioned processes.
V φ+P Dm The sequence of operations in the proposed work
ψi,m = dt (2)
R t=φ P is exhibited in the form of a UML sequence diagram,
4596 S. Ahmad et al. / Accident risk prediction in semi-autonomous vehicles

5. Implementation detail

In this Section, the implementation stack is illus-

trated. The machine learning and model training
part of the work is implemented in Pandas and
TensorFlow libraries of Python. For model training
using different algorithms, Scikit Learn (SKlearn)
library. This library has built-in modules for different
algorithms, i.e., Decision tree, Random Forest, and
Logistic Regression, to name a few. Once the model is
trained, it is deployed on a web application developed
in Python mini-framework named Flask. For visual
appearance, Bootstrap 3 and Javascript are used. We
use Arduino, Raspberry PI, and sensors such as tem-
perature sensors, humidity sensors, wind sensors, to
name a few, and actuators such as buzzer and LEDs
for alerts. The implementation stack is summarized
in Table 3.

6. Execution results

In this Section, the execution results of the pro-

posed system are discussed. The main components
Fig. 5. Flowchart of Proposed Safety Modeling based on Task
described in the earlier sections are implemented
Management. using the technologies mentioned earlier. As dis-
cussed, there are two planes embedded plane and task
orchestration plane. The deployment is shown in Fig.
as shown in Fig. 6. We have two main planes; IoT 7. Conceptually, there is a client application, which, in
Embedded Plane and Task Management and Orches- this case, is exposing safety services in a typical elec-
tration Application Plane. IoT Embedded Plane hosts tric vehicle. The client application consumes services
an IoT server which at first initializes and registers which are based on task management and application
physical nodes connected with it. These nodes can be platform. The platform is the mean by which the IoT
either sensors or actuators. The registry information devices interact with the client. The IoT devices, as
is passed on to the application. The device infor- described earlier, are Rasberry PI and Arduino. The
mation is sent to the virtualizer module, where the physical setup is shown in Fig. 7b. The central focus
virtual objects are formed. Virtual objects (VO) are of the Fig. is the task manager and virtual object man-
the in-system representation of physical devices. Task ager. Task manager generates tasks and persists them
generator creates the tasks, and the tasks and VOs are in tasks database. Similarly, the virtual object man-
passed to Task orchestrator module. Once tasks are ager also generates virtual objects. Virtual objects and
created, the data is loaded to the mapper. The phys- tasks are supplied to the mapper module, which maps
ical device data is also loaded to the mapper, where the tasks on virtual objects. The tasks are then allo-
the tasks are mapped on virtual objects. The mapping cated in the embedded plane. It can be seen from Fig.
information is sent to the monitoring center. The mon- 7b that Arduino has a pulse oximeter and ECG sen-
itoring center deploys sensing tasks periodically and sor attached to the Libillium sensor platform shield.
assesses the risk index. At times, when the risk index In contrast, fan motors, BMP sensors are attached to
exceeds the threshold, it sends an acknowledgment to Raspberry PI.
the task generator. Task generator generates control
tasks and deploys it to mitigate the risk by notifying 6.1. Execution flow
with alerts. The information is visualized on a remote
simulator. Sensor data are visualized and can be seen Task Management module is responsible for task
without being on-premises. adding, editing, deleting, and storing it to the
 S. Ahmad et al. / Accident risk prediction in semi-autonomous vehicles 4597

Fig. 6. Sequence of Interaction among different components.

Table 3 and the configuration is persisted. Finally, the task-vo

Implementation Stack pair is deployed on physical devices. Fig. 8 shows the
Component Description execution flow of the process mentioned above.
Hardware Raspberry PI, Arduino PC
Virtual object manager deals in virtual object gen-
Operating System Raspbian, Window, Arduino 10
Memory 1 GB, 8 GB eration based on the device registry information.
Server Flask, ArduinoHTTP Webserver Virtual objects have been assigned tags and match-
Libraries Jinja, HTTPClient, Bootstrap, ing names, which can help in the mapping process.
Javascript, HTML5 and CSS3
IDE PyCharm, IDLE, Arduino Studio
Similarly, the task name can also be named in such
Core Programming Python 3.5 a way that it can say the type. For instance, for sen-
Language sors, tasks names are prepended with get. The virtual
Model Training Pandas, TensorFlow, SKlearn object manager module interfaces are shown in Fig. 8
Libraries
Once the tasks are deployed on virtual objects, the
deployment starts, and it checks the risk index is com-
puted periodically based on the model output. If the
database. Firstly, tasks are added by the task manage-
risk index exceeds 2, the vehicle and driving are con-
ment module and stored in the database. Secondly, the
sidered in the hazardous state. At this point, control
virtual objects are also generated by virtual objects
tasks should be triggered to inform the driver about
manager. The data from the device registry is avail-
the associated risk of accidents. Fig. 9 illustrates it
able to this module, which uses the data to generate a
pictorially.
virtual object for the corresponding physical device.
The virtual objects are also persisted in the database.
Next, the mapping is done in mapping plane in which 7. Model evaluation
on one end, tasks are populated whereas, on the
other end, virtual objects are populated. The tasks In this Section, we evaluate the performance of the
are mapped using the drag-and-drop JSPlumb library, model based on three algorithms. We use Random
4598 S. Ahmad et al. / Accident risk prediction in semi-autonomous vehicles

Fig. 7. Detailed Schematic Representation of Sub-modules of proposed framework.

Forest, Decision Tree and Linear Regression. The overall accuracy of Random Forest is also the high-
model output is risk severity classes. There are three est among other algorithms. The micro-average for
classes. 1 represents fatality, 2 represents casualties, Logistic regression is highest, which means this algo-
and 3 represents no harm. The input is the data param- rithm is less accurate, but its performance is balanced
eters. The data is normalized prior to giving it to the across all classes.
training phase. Table 4 shows the summary of model Fig. 10 shows the graph of true predictions and
evaluation in terms of various attributes. The recall false predictions. The value of true predictions is
and precision of different classes in Random For- comfortably higher than the value of false predic-
est are more than its counterpart’s algorithms. The tions. Similarly, for every respective algorithm, the
 S. Ahmad et al. / Accident risk prediction in semi-autonomous vehicles 4599

Fig. 8. Execution flow and pictorial representation of various interfaces.

crash, driver’s biological condition, and speed of the

vehicle. Many research tests have been performed
on Real-world data in the recent studies [21–24]
and found that the less the human intervention and
the more automation, the less will be the crash
probabilities.
It is predicted that human offloading from
autonomous vehicles will make the transportation
smart, and the chance of crash will be much lower;
Fig. 9. Vehicles Safety Monitoring and Control.
however, the fully automated car is yet to be deployed
by the organization. Therefore, nowadays, warning
systems and driver assistance systems are in high
true predictions of Random forest are slightly higher demand, which facilitates drivers with alerts and
than the counterparts’ algorithms. notifications by executing real-time tasks remotely
from edge nodes [25–27]. These devices can detect
numerous states of drivers and environments such
8. Related work as drowsiness, alcohol, and drugs and fatigue of
the driver [28]. Commercial systems such as LiDar,
The related work of this work is interlinked with anti-lock braking systems (ABS), and adaptive head-
the motivation, architecture, and design of this work. lights are among few notable systems; however, these
Numerous efforts have been made on the road safety systems are dedicated to a single job and are very
of autonomous vehicles. Fagnant et al. [19] consid- costly [27]. In this paper, a generic approach to
ered the near elimination of error linked with human finding the most pressing parameters linked with
interventions. These human errors form 90% of the crashes are investigated, and a lightweight sys-
fatalities. In another study, Rau et al. [20] proposed tem is developed to give alerts and notifications
a mechanism and categorized AV functions into four remotely in a real-time manner and this, to the best
steps in order to identify the cause of fatalities. These of the authors’ knowledge, is the first-ever such
functions are the pre-crash scenario, location of the attempt.
4600 S. Ahmad et al. / Accident risk prediction in semi-autonomous vehicles

support
4111
38151
264697
306959
306959
306959
f1-score
0.0412
0.171
0.858

0.357
0.761
0.75
Decision Tree

85.67
Recall
0.046
0.186
0.845

0.359
0.752
0.75
Precision

0.158
0.871

0.355
0.771
0.03

0.75
support
4111
38151
264697
306959
306959
306959
Analysis of Model Training Results of various algorithms

f1-score

0.308
Logistic Regression

0.01
0.01
0.92
0.93

0.79
92.4
Recall
0.001
0.001
0.99
0.99
0.33
0.86
Table 4

0.866541
Precision

Fig. 10. Model Evaluation in terms of Confusion Matrix.

0.01
0.01

0.86
0.28
0.74

9. Conclusion

In this paper, we have proposed an Internet

support
4111
38151
264697
306959
306959
306959

of things application to facilitate semi-autonomous

vehicles and helps them reduce crashes and fatali-
ties. This paper considers the human driving job as a
f1-score

collection of tasks that have to be performed within a

0.0905
0.916

0.801
0.51

0.84
0.34

span of time. If the real-time tasks can be guaranteed

Random Forest

to be executed in the allocated time, a warning sys-

95.67

tem could be developed to alert drivers for making

0.0563
Recall

an error, which could lead to a potential accident. We

0.972

0.845
0.51

0.84
0.34

have used Safety data of the UK and trained mod-

els based on a variety of algorithms and deployed
it on the embedded IoT application. The model also
0.866541
Precision

helped in prioritizing the parameters which affect the

0.776
0.23

0.84
0.38

most on fatalities. We have used weather conditions

0.6

and driver biological conditions and deploy different

sensors which periodically sense these scenarios and
Overall Accuracy
weighted average

compute the risk of driving. If the risk of driving is

macro average
micro average

more than some threshold, it will give alerts and noti-

Parameters
Algorithm

fications. In the end, the model is evaluated, and the

execution results are illustrated with interfaces. The
results show that this paper can be the best starting
1
2
3
 S. Ahmad et al. / Accident risk prediction in semi-autonomous vehicles 4601

point to redefine the mechanism of driver assistance [13] P. Bansal and K.M. Kockelman, Forecasting americans’
systems and get rid of the commercial and costly long-term adoption of connected and autonomous vehicle
technologies, Transportation Research Part A: Policy and
devices which only do a dedicated job. Practice 95 (2017), 49–63.
[14] I.J. Reagan, J.B. Cicchino, L.B. Kerfoot and R.A. Weast,
Crash avoidance and driver assistance technologies–are they
References used? Transportation Research Part F: Traffic Psychology
and Behaviour 52 (2018), 176–190.
[15] J. Bierstedt, A. Gooze, C. Gray, J. Peterman, L. Raykin
[1] R. Lukkien, T. Ozcelebi, J.J. Lukkien, O. Uzun, et al., and J. Walters, Effects of next-generation vehicles on travel
Discovery, monitoring and management in smart spaces demand and highway capacity, FP Think Working Group
composed of low capacity nodes, IEEE Transactions on (2014), 10–11.
Consumer Electronics 56(2) 2010. [16] M.A. Quddus, Time series count data models: An empir-
[2] N. Dimakis, J.K. Soldatos, L. Polymenakos, P. Fleury, ical application to traffic accidents, Accident Analysis &
J. Curı́n and J. Kleindienst, Integrated development of Prevention 40(5) (2008), 1732–1741.
context-aware applications in smart spaces, IEEE Pervasive [17] K.H. Janstrup, Road Safety Annual Report 2017, 2017.
Computing 7(4) (2008), 71–79. [18] data.gove.uk, Road safety data, https://data.gov.uk/dataset/
[3] H.-n. Lee, S.-h. Lim and J.-h. Kim, Umons: Ubiquitous cb7ae6f0-4be6-4935-9277-47e5ce24a11f/road-safety-data,
monitoring system in smart space, IEEE Transactions on [Online; accessed 01-January-2020].
Consumer Electronics 55(3) 2009. [19] D.J. Fagnant and K. Kockelman, Preparing a nation for
[4] S. Ahmad, L. Hang and D.H. Kim, Design and implemen- autonomous vehicles: Opportunities, barriers and policy
tation of cloud-centric configuration repository for diy iot recommendations, Transportation Research Part A: Policy
applications, Sensors 18(2) (2018), 474. and Practice 77 (2015), 167–181.
[5] S.D.T. Kelly, N.K. Suryadevara and S.C. Mukhopadhyay, [20] P. Rau, M. Yanagisawa and W.G. Najm, Target crash pop-
Towards the implementation of iot for environmental con- ulation of automated vehicles, in In 24th International
dition monitoring in homes, IEEE Sensors Journal 13(10) Technical Conference on the Enhanced Safety of Vehicles
(2013), 3846–3853. (ESV) (2015), 1–11.
[6] X. Li, R. Lu, X. Liang, X. Shen, J. Chen and X. Lin, Smart [21] V.V. Dixit, S. Chand and D.J. Nair, Autonomous vehicles:
community: An internet of things application, IEEE Com- Disengagements, accidents and reaction times, PLoS One
munications Magazine 49(11) 2011. 11(12) p.e0168054, 2016.
[7] S. Childress, B. Nichols, B. Charlton and S. Coe, Using [22] B. Schoettle and M. Sivak, A preliminary analysis of real-
an activitybased model to explore the potential impacts of world crashes involving self-driving vehicles, University of
automated vehicles, Transportation Research Record: Jour- Michigan Transportation Research Institute, 2015.
nal of the Transportation Research Board 2493 (2015), [23] P. Bhavsar, P. Das, M. Paugh, K. Dey and M. Chowd-
99–106. hury, Risk analysis of autonomous vehicles in mixed traffic
[8] L. Da Xu, W. He and S. Li, Internet of things in industries: A streams, Transportation Research Record: Journal of the
survey, IEEE Transactions on Industrial Informatics 10(4) Transportation Research Board 2625 (2017), 51–61.
(2014), 2233–2243. [24] F.M. Favarò, N. Nader, S.O. Eurich, M. Tripp and
[9] S. Ahmad, S. Malik, I. Ullah, M. Fayaz, D.-H. Park, K. N. Varadaraju, Examining accident reports involving
Kim and D. Kim, An adaptive approach based on resource- autonomous vehicles in california, PLoS One 12(9)
awareness towards power-efficient real-time periodic task p. e0184952, 2017.
modeling on embedded iot devices, Processes 6 (7) (2018), [25] I.R. Johnston, C. Muir and E.W. Howard, Eliminating seri-
90. ous injury and death from road transport: A crisis of
[10] S. Ahmad, S. Malik and D.-H. Kim, Comparative analysis complacency. CRC Press, 2013.
of simulation tools with visualization based on real-time [26] A.E. Curry, J. Hafetz, M.J. Kallan, F.K. Winston and D.R.
task scheduling algorithms for iot embedded applications, Durbin, Prevalence of teen driver errors leading to seri-
International Journal of Grid and Distributed Computing ous motor vehicle crashes, Accident Analysis & Prevention
11(2) (2018), 1–10. 43(4) (2011), 1285–1290.
[11] A. Daly, Privacy in automation: An appraisal of the emerg- [27] S.E. Shladover, Cooperative (rather than autonomous)
ing australian approach, Computer Law & Security Review vehicle-highway automation systems, IEEE Intelligent
33(6) (2017), 836–846. Transportation Systems Magazine 1(1) (2009), 10–19.
[12] T. Litman, Autonomous vehicle implementation predic- [28] O.A. Basir, J.P. Bhavnani, F. Karray and K. Desrochers,
tions, Victoria Transport Policy Institute Victoria, Canada, Drowsiness detection system, 23 (2004), US Patent
2017. 6,822,573.