Capturing Uncertainty in Heavy Goods Vehicle Driving Behaviour

Abstract

There is a growing interest in understanding and identifying risky driving behaviours due to the numerous deaths and fatalities attributed to road traffic accidents. For Heavy Goods Vehicles (HGVs), understanding driving behaviour and its impact on road safety is a subject of interest for researchers, governments and industrial sectors, as they rely on HGVs for the delivery of goods and services. The current literature on HGV driving behaviour uses machine learning techniques to uncover core driving incident stereotypes. However, human behaviour contains different levels of uncertainty and stereotyping driving behaviour with traditional crisp methods may cause information loss and establish unfair boundaries as they do not take context into consideration. Moreover, the sensor readings also have uncertainties, and the driving stereotypes may have different subjective interpretations. In order to capture those intermediate possibilities in driver stereotyping, we propose a data-driven Fuzzy Logic system that can capture the uncertainties within driving features (data) and between driving stereotypes, and classifies drivers according to the risk of their driving styles on a scale of 0 to 100, where 0 is a low risk driver and 100 high risk. The results from telematics data show that our proposed method provides a reliable, fair and explainable approach for real-time identification of HGV driving risk level.

Full text

Capturing Uncertainty in Heavy Goods Vehicles
Driving Behaviour
Jimiama Mafeni Mase1, Utkarsh Agrawal2, Direnc Pekaslan1, Mohammad Mesgarpour3,
Peter Chapman4, Mercedes Torres Torres1, Grazziela P. Figueredo1
1School of Computer Science, The University of Nottingham
2School of Medicine, University of St Andrews
3Microlise, Farrington Way, Eastwood, Nottingham
4School of Psychology, The University of Nottingham
Abstract—There is a growing interest in understanding and
identifying risky driving behaviours due to the numerous road
fatalities attributed to them. For Heavy Goods Vehicles (HGVs),
understanding driving behaviour and its impact on road safety
is a subject of interest for researchers, the government and
industrial sectors, as they rely on HGVs for the delivery of goods
and services. The current literature on HGV driving behaviour
uses machine learning techniques to uncover core driving incident
stereotypes. However, human behaviour contains different levels
of uncertainty and stereotyping driving behaviour with tradi-
tional crisp methods may cause information loss and establish
unfair boundaries as they do not take context into consideration.
Moreover, the sensor readings also have uncertainties, and the
driving stereotypes may have different subjective interpretations.
In order to capture those intermediate possibilities in driver
stereotyping, we propose a data-driven Fuzzy Logic system that
can capture the uncertainties within driving features (data) and
between driving stereotypes, and classifies drivers according to
the risk of their driving styles on a scale of 0 to 100, where 0 is
a low risk driver and 100 high risk. The results from telematics
data show that our proposed method provides a reliable, fair and
explainable approach for real-time identification of HGV driving
risk level.
Index Terms—Fuzzy Logic Systems, Driving Risk Level, Clus-
tering, Uncertainty, Classification, Heavy Goods Vehicles, Driving
Incidents, Explainable AI
I. INTRODUCTION
Every year the World Health Organisation releases a report
with statistics on road fatalities and the report consistently
attributes human errors and violations as one of the main
causes [1]. This has prompted a growing research interest in
understanding driving styles [2]–[5], distraction triggers [6]–
[8], driver interactions with vehicle technologies [9], [10] and
psychological factors affecting safe driving [11], [12]. One
area of specific interest is understanding the different types of
incidents caused by risky driving behaviours of Heavy Goods
Vehicles (HGVs) and its impact on road safety.
The current approaches to monitor driving behaviour are
based on sensors producing large volumes of data. Effective
data-driven tools for the detection and mitigation of driving
errors as well as approaches to report and understand vehicle
incidents due to human errors or violations are therefore neces-
sary. Furthermore, better interpretation of the data produced by
HGVs can assist stakeholders to develop actions, policies and
technologies for safe driving and to minimise risks and costs
within the HGV transportation network. In a recent research by
Agrawal et al. [5], the authors uncovered eleven HGV driving
incident stereotypes using telematics1data from three years
i.e. 2014, 2015 and 2016. Their study is limited with regards
to their methodology, which considers driving features and
stereotypes as crisp sets (where a driving incident value either
belongs to a particular category or not). Such a system may
be bias and cause information loss as driving features and
stereotypes contain different levels of uncertainty produced by
sensor readings and subjective interpretations. For example,
in traditional logic,‘50’ harsh braking incidents in a day can
be considered as ‘high-risk’ while ‘49’ can be regarded as
not ‘high-risk’, when it is obvious that if ‘50’ harsh braking
incidents has some degree of ‘high risk’, ‘49’ should also have
some degree of ‘high-risk’.
In this paper we build upon this aforementioned research
and extend it by presenting a novel data-driven Fuzzy Logic
System (FLS) for capturing the uncertainties in the available
telematics data and scoring HGV drivers according to the risk
of their driving incidents. In order to achieve this, a frame-
work consisting three steps is employed. First, a classification
framework and profile labelling algorithm (based on works in
[4] and [5]) is employed on telematics data. The data contains
driving incidents for four years, i.e., 2014, 2015, 2016 and
2017. The objective is to identify core driving stereotypes.
The driving incidents (antecedents) and driving stereotypes
(consequents) are subsequently used to obtain the fuzzy sets
for the FLS. Secondly, Wang-Mendel [13] method is employed
to learn the rules (knowledge base) that maps these antecedents
to consequents of the FLS. Lastly, a Mamdani inference [14]
approach is used to combine these rules to produce a scoring
system that can rate drivers’ risk level (between 0 and 100)
in real-time based on their driving incidents, where a score 0
represents a ‘low risk’ driving behaviour, while 100 is a ‘high
risk’ driving behaviour.
This paper is organised as follows, in Section II we re-
view the literature on understanding driving behaviour using
telematics data. Subsequently, we review FLS basic concepts
and their applications to driving behaviour. In Section III, we
1Telematics are communication systems which link to vehicles’ Controller
Area Network (CAN) databus for capturing vehicle data using sensors, e.g.,
engine revs, fuel consumption, accelerator position etc

provide an overview of the 2-stage classification framework,
profile labelling algorithm, introduce the datasets used in our
study and describe our methodology in detail. The results are
presented along with discussion in Section IV, and Section V
concludes the paper and establishes the opportunity for future
work.
II. BACKGROU ND
A. Related Work
Apart from interviews and questionaires, telematics have
been one of the main data sources for detecting driving
behaviour in modern studies. Researchers have analysed telem-
atics data using machine learning techniques to understand and
classify driving behaviour with promising results. For instance,
Constantinescu et al. [15] use Hierarchical Clustering Analysis
(HCA) to identify six profiles among 25 drivers in the city of
Bucharest, based on four telematics driving features i.e. speed,
acceleration, braking and kinetic energy. Similarly, Ellison
et al. [16] propose a driver risky index framework based
on three telematics driving incidents (i.e. over speed, harsh
acceleration and harsh braking) collected from 106 drivers
in Sidney within 25 days. These studies employ different
variations of clustering algorithms to identify groups of driving
incidents stereotypes, however, they are limited with regards to
the number of samples analysed, which is relatively small, and
their experiments are performed in controlled environments,
which is unrepresentative of real driving behaviour.
In a seminal work by Figueredo et al. [4], the authors
tackle these shortcomings by analysing driving incidents data
captured for more than 20,000 HGV drivers in the UK. The
authors use a 2-stage classification framework [17] and a
subjective profile labelling algorithm to uncover eight HGV
driving stereotypes from driving incidents data of the year
2015. Agrawal et al. [5] build upon this work to put forward
a classification system to classify HGV drivers in real time.
They use a decision tree classifier to learn driving behaviours
of drivers from the years 2014 to 2016 and test its performance
on new drivers from the year 2017. Although they achieve high
classification accuracy, their proposed model does not consider
the uncertainties within the driving incidents and stereotypes
injected by sensor readings and subjectivity, which can cause
information loss and unfair classification systems.
B. Fuzzy Logic Systems and their applications to driving
behaviour
Fuzzy Logic was introduced by [18] to address imprecision
or uncertainties in input and output variables directly by defin-
ing them using linguistic terms with their degrees of member-
ship between 0 and 1 (fuzzy sets). Fuzzy Logic Systems (FLS)
consist of three stages: 1) Fuzzification, 2) Fuzzy inference,
and 3) Defuzzification. In fuzzification, inputs (e.g. driving
incidents) and outputs (e.g. driving stereotypes of behaviour)
also known as antecedents and consequents respectively, are
represented by linguistic variables (e.g. ‘Low’, ‘Moderate’
and ‘High’ number of harsh braking incidents). The linguistic
variables have soft boundaries that allows them to better
Fig. 1: An example of fuzzy sets for a driving feature
capture the uncertainties in inputs and outputs. In Fig. 1, ‘35’
harsh braking incidents is considered ‘Moderate’ risk with a
high degree of membership and also considered ’High’ risk
with a low degree of membership.
Fuzzy rules are defined to capture the relationships between
the input and output variables and are usually based on IF-
THEN statements, which are easy to understand. For example,
a fuzzy rule for driving behaviour can be:
IF the number of harsh braking incidents is ‘Low’
risk AND the number of over speeding incidents is ‘Low’
risk THEN the driver’s risk level is ‘Low’.
Lastly, the fuzzy inference systems combine the input fuzzy
sets using the fuzzy rules to produce output fuzzy sets, and
defuzzification maps the output fuzzy sets to a real or crisp
number (e.g. a driving score in this paper).
Recently, FLS are examined in transport research to capture
the uncertainties in driving features, and to develop fair
and explainable classification systems [19]–[22]. Aljaafreh et
al. [19] employ FLS and two driving features (acceleration and
speed) to classify drivers into four predefined driving stereo-
types (below normal, normal, aggressive, and very aggressive
profiles). In a different study, Imkamon et al. [20] use FLS
for detecting unsafe driving behaviour. They use three sensors:
an engine control unit (ECU) reader, an accelerometer and a
camera, to extract features, such as vehicle movement. The
features are then combined using a fuzzy inference system
to identify a driver’s risk level ranging from 1 to 3. In a
similar study by Boonmee et al. [21], the authors use harsh
acceleration, sudden braking, and turning quickly incidents
and a FLS for rating drivers on 1 to 10 scale (where 10 is
the most risky driver) .
The above studies on the use of FLS to describe driving
behaviours are limited as their input fuzzy sets and mem-
bership functions are defined using small number of drivers
leading to fewer driving stereotypes. Thus, this limitation can
introduce bias and make their outcome unrepresentative of
more general driving patterns. In this work, we bridge these
gaps by capturing uncertainties in HGV driving incidents using
a FLS developed from four years of telematics data i.e. input
fuzzy sets will be defined by driving stereotypes captured by
real world data. We will use the FLS to score drivers according

to the risk of their driving behaviours in the range from 0 to
100, where 0 is a low risk driver while 100 is high risk.
III. METHODOLOGY
In this section we present our novel data-driven fuzzy logic
approach for scoring HGV drivers based on driving incidents
captured using telematics. First, the fuzzy sets and membership
functions are obtained from the driving profiles uncovered
from the classification framework. Subsequently, the fuzzy
rules are generated using the Wang-Mendel method. Lastly, the
Mamdani inference method are used to score drivers according
to the risk level of their driving behaviour. An overview of our
proposed method is shown in Fig. 2. Each stage is described
in detail below.
A. Dataset Description
The dataset utilised in this study is provided by our industry
partner Microlise [23]. Information produced by their telemat-
ics solutions are transmitted and collected from the HGVs in
real time. Data is captured by sensors connected to multiple
electronic control units using a Controller Area Network
(CAN) bus. The HGV drivers must complete a minimum of 10
journeys per quarter (i.e. minimum 40 journeys yearly) each
year on any road in the UK to be considered in our analysis.
The data was collected between the first of January and the
thirty first of December for the years 2014, 2015, 2016, and
2017. The dataset consists of four driving incidents: frequency
of Harsh Braking (HB) events, Over-Speed (OS) duration
in seconds, Excessive Throttle (ET) duration in seconds and
frequency of Over Revving (ORev) events of HGV drivers.
It is important to note that these incidents were chosen for
our study because they are the most relevant incidents present
in all of the HGVs, which are related to the risk of crashes
and vehicle costs. In total, 15,893 drivers in year 2014, 21,234
drivers in year 2015, 34,675 drivers in year 2016, and 35,432
drivers in year 2017 were collected using this criteria. The
drivers were grouped into three subgroups based on their daily
average mileage travelled as uncovered by [4] including short
subgroup drivers who cover an average daily mileage upto
136.70 miles, medium subgroup drivers who cover an average
daily mileage between 136.70 and 217.48 miles, and long
subgroup drivers who covers an average daily mileage more
than 217.48 miles. Table I presents the distribution of the
drivers for the four years within the three driving subgroups.
TABLE I: Distribution of drivers among Average Daily
Mileage groups for years 2014, 2015, 2016 and 2017
Average daily Distribution of Distribution of Distribution of Distribution of
mileage groups drivers in 2014 drivers in 2015 drivers in 2016 drivers in 2017
Short 3,327 5,076 16,281 10,745
Medium 6,419 8,392 10,232 14,833
Long 6,147 7,766 8,162 9.854
Total 15,893 21,234 34,675 35,432
B. Data Pre-processing
For unbiased results, we normalise the number of driving
incidents produced by each driver by dividing with the total
driving time (in seconds). To obtain stable driving stereotypes,
the drivers who were consistently present across all the years
and who did not change subgroups across these years are
considered for the analysis. For example, if a driver is present
in short average daily mileage subgroup in the year 2014,
that driver should also be in the short average daily mileage
subgroup for the years 2015, 2016 and 2017 respectively. As a
result, 2,462 drivers are considered across the four years with
569, 964 and 929 drivers distributed between short, medium
and long average daily mileage subgroups respectively.
C. Generating Driving Stereotypes for FLS
To generate Membership Functions (MFs) and fuzzy rules,
the first step is to obtain the driving stereotypes. The driving
stereotypes are obtained using 2-stage classification framework
followed by a profile labelling algorithm (introduced in [4],
[5]) on the datasets from four years. The classification frame-
work is a two step process: 1) Multiple clustering algorithms
are run to group the data. Using a consensus approach, most of
the data are assigned to one of the identified groups (profiles),
while some of the data remain unclustered due to lack of
consensus. 2) An ensemble of classification algorithms is
trained on the clustered data in the previous step. The trained
model is then run on the unclustered data to assign them to one
of the groups uncovered during clustering. The classification
framework uncovers eleven clusters. Subsequently, a profile
labelling algorithm sub-divides each driving incident into three
categories or linguistic terms (fuzzy sets); ‘low’, ‘moderate’
and ‘high’. Fig. 4 (a), (b), (c) and (d) show the boxplot
distributions of the four driving incidents HB, OS, ET and
ORev respectively, for short mileage subgroup (due to space
constraints, we only show the distributions of short mileage
subgroup). It is important to note that the boxplot for ORev
has no distribution for ‘high’ as the profile labelling algorithm
produced only two categories for ORevs of short mileage
drivers due to their relatively smaller values compared to the
other subgroups. The clusters obtained above represent the
eleven driving stereotypes, as shown in Table II. The first
column represents the driving profile number, columns two to
five represent the labels of the occurrence of incidents in each
profile and the last column indicates the mileage subgroups
in which these profiles are present. A more elaborate descrip-
tion of the classification framework and the profile labelling
algorithm can be found in [4] and [5].
TABLE II: HGV driving stereotypes and their feature labels
uncovered across 2014, 2015, 2016 and 2017
Profile Harsh Over-Speed Excessive Number of Driving
Number Braking duration Throttle Over Revs Subgroup
1 Low Low Low Moderate S,M,L
2 Low Low High Low M
3 Low Low High Moderate S
4 Moderate Moderate High Low L
5 Moderate Moderate High Moderate S,M,L
6 Moderate Moderate High High M
7 Moderate High Moderate Low S
8 Moderate High High Low L
9 High Moderate Moderate Low M
10 High High Moderate Low S,M,L
11 High High High Low M,L

Fig. 2: Our proposed Data-Driven Fuzzy Logic architecture to score HGV driving behaviour
Fig. 3: Driving profiles (Consequents) membership functions
D. Generating fuzzy sets and membership functions
For all subgroups, the quartiles of the boxplots for each driv-
ing incident are used to construct their respective membership
functions (antecedents MFs). For example, Fig. 5 shows the
membership functions for the four driving features of short
daily average mileage subgroup generated from their boxplot
distributions shown in Fig. 4. It is important to note that for
each feature, the three fuzzy sets represent the corresponding
low, moderate and high distributions respectively. The eleven
driving profiles i.e. the consequents of the FLS obtained
by the classification framework, are represented in Fig. 3.
The consequent membership functions are equally distributed
between 0 and 100, which represents the driving risk level.
E. Fuzzy rule generation, inference and defuzzification
The Wang-Mendel method [13], one of the most commonly
used rule generation technique, is employed to generate the
rules of the FLS. In our dataset, the four different features
and output (profile) pairs are denoted as:
(x1
HB , x1
OS , x1
ET , x1
ORev ;y1
P R),
(x2
HB , x2
OS , x2
ET , x2
ORev ;y2
P R),
.
.
..
.
.
(xN
HB , xN
OS , xN
ET , xN
ORev ;yN
P R),
(1)
where Nis the number of input-output data pairs and each
abbreviation denotes the corresponding feature (e.g. HB =
Harsh Braking incident).
By using the 2-stage classification framework (see details in
Section III-C), the domain and three Membership Functions
(MFs) i.e. Low, Medium, High, are defined for each feature in
the datasets. Then, the output domain is defined by splitting
11 MFs as shown in Fig. 3 and each driver is assigned by a
profile number between 1-11. After constructing the MFs and
consequents for all the input-output pairs, each pair is assigned
to the corresponding MFs so that the model rules are generated
along with their weights. By following the practice from
[13], the rule weights are utilised to implement rule reduction
procedure on the conflicting rules. Table III shows the rules for
the short mileage subgroup generated in a data-driven manner
(due to space constraints, we only show fuzzy rules of short
mileage subgroup). For a more elaborate description of the
Wang-Mendel rule generation method, please refer [13].
Next, we use the Mamdani inference technique [24], a
commonly used rule-based inference method, to combine
the antecedents fuzzy sets using the fuzzy rules and fuzzy
operators (min and max for T-norm and T-conorm operators) to
obtain output fuzzy sets. The output fuzzy sets are combined to
calculate a crisp value (driver’s score) using centroid defuzzi-
fication [25] (a common and fast defuzzification technique).
IV. RES ULT S AN D DISCUSSION
A. Driving stereotypes
The 2-stage classification framework and profile labelling
algorithm uncovers eleven HGV driving incident stereotypes
(Table II), which are used for generating the membership
functions (Fig. 3 and 5) to capture the uncertainties in the
driving incidents and stereotypes. Profile 1 represents a ‘very
low’ risk level stereotype with low incidents, except for over
revving that has a moderate number of incidents. Profiles 2 and
3 can be considered ‘low’ risk level stereotypes because of low
harsh braking and over-speeding incidents. Profiles 4, 5 and 6
are similar except for their economic behaviour (over revving
incidents) and can be described as ‘moderate’ risk level
stereotypes with moderate harsh braking and over-speeding
incidents. Profiles 7 and 8 are ‘high’ risk level and speedy

Fig. 4: Boxplots for the driving features of short mileage drivers
Fig. 5: Driving incidents membership functions for short mileage drivers
profiles with high duration of over speeding incidents while
profile 9 represents a ‘high’ risk level and aggressive profile
due to high harsh braking and excess throttling incidents.
Profile 10 can be observed to be a ‘very high’ risk level driving
stereotype with high harsh braking and over-speeding incidents
and profile 11 can be described as ’extremely’ risky and the
most aggressive of all the profiles with high harsh braking,
over-speeding and excessive throttling. The output domain (0
to 100) is defined by placing these 11 profiles according to
our subjective risk level interpretations from ‘very low’ to
‘extremely’ risky profiles.
B. Uncertainties in driving incidents
The MFs produced by the 11 profiles capture the uncertain-
ties within driving incidents caused by sensor readings and
multiple subjective interpretations (e.g., subjectivity produced
by the profile labelling algorithm). Such uncertainty can be
captured and visualised in the over revving incident MF in
Fig. 6 with the fuzzy sets: Low, Moderate and High. The
uncertainty of whether the over revving value, 0.01, should
be considered ‘Low’, ‘Moderate’ or ‘High’ is represented by
the MFs as follows:

TABLE III: Fuzzy rules for short mileage drivers
Rule Harsh braking Over speeding Excessive Throttle Over Revs Profile
No No
1 LOW AND (LOW OR MODERATE) AND (LOW OR MODERATE) AND (LOW OR MODERATE) 1
2 MODERATE AND LOW AND (LOW OR MODERATE) AND (LOW OR MODERATE) 1
3 MODERATE AND MODERATE AND (LOW OR MODERATE) AND MODERATE 1
4 HIGH AND LOW AND (LOW OR MODERATE) AND (LOW OR MODERATE) 1
5 HIGH AND MODERATE AND LOW AND MODERATE 1
6 LOW AND MODERATE AND HIGH AND LOW 3
7 (MODERATE OR HIGH) AND LOW AND HIGH AND MODERATE 3
8 LOW AND (LOW OR MODERATE) AND HIGH AND MODERATE 5
9 MODERATE AND MODERATE AND HIGH AND (LOW OR MODERATE) 5
10 (LOW OR MODERATE) AND LOW AND HIGH AND LOW 5
11 HIGH AND MODERATE AND HIGH AND LOW 5
12 LOW AND HIGH AND MODERATE AND LOW 7
13 LOW AND HIGH AND HIGH AND MODERATE 7
14 LOW AND HIGH AND LOW AND LOW 10
15 MODERATE AND MODERATE AND (LOW OR MODERATE) AND LOW 10
16 (MODERATE OR HIGH) AND HIGH AND (LOW OR MODERATE OR HIGH) AND (LOW OR MODERATE) 10
17 HIGH AND MODERATE AND (LOW OR MODERATE) AND LOW 10
18 HIGH AND MODERATE AND (MODERATE OR HIGH) AND MODERATE 10
Note: Profiles 2, 4, 6, 8, 9 and 11 are found in the other driving subgroups (i.e., Medium and Long) as shown in Table II.
Fig. 6: Over revving membership function produced by our
Fuzzy Logic approach
ORLow (0.01) = 0.75
ORM oderate(0.01) = 0.35
ORH igh(0.01) = 0
These fuzzy values are used during rule evaluation to
determine the driver’s risk level or score. For example, if we
consider Rule 7 of short mileage drivers in Table III:
IF HB is either ‘MODERATE’ OR ‘HIGH’ AND OS
is ‘Low’ AND ET is ‘HIGH’ AND ORev is ‘MODER-
ATE’ THEN the driver’s risk level is ‘LOW’
The degree of membership 0.35 represented by
ORModer ate(0.01) as well as those obtained from the
other driving incidents will be used to evaluate the firing
strength of the rule and the rules will be used to obtain the
driving score.
C. Scoring HGV driving behaviour
The Mamdani inference system combines the antecedents
(driving incidents) fuzzy sets using the knowledge base
(rules) generated by the Wang-Mendel method to produce
rule strengths. The rule strengths are then applied to the
consequents to produce inferred output fuzzy sets. These
output fuzzy sets are combined using centroid defuzzification,
which considers the center of the area of the output fuzzy sets
to produce the driving risk score. The scores or risk levels
range from 0 to 100, where 0 represents low risk drivers and
100 is high risk drivers. Fig. 7 shows an example of output
fuzzy sets obtained from the FIS. The shaded regions in the
figure represent the fired consequents i.e., for certain values
of the driving incidents, the rule strengths for the consequent
fuzzy sets ‘profile 5’ and ‘profile 10’ (denoted by the purple
and blue shades) are 0.8 and 0.05 respectively. This means
that the driver’s risky behaviour belongs to profile 5 with
a degree of 0.8 and to profile 10 with a degree of 0.05.
The defuzzification method computes the center of the area
between the shaded regions which represents the driver’s risk
level i.e., 39. Furthermore, the logical rule knowledge base
(IF-THEN rules) coupled with the linguistic fuzzy sets make
the proposed system explainable to end-users and stakeholders
in the HGV community. This framework can be applied to the
development of driver alert systems by providing feedback to
drivers about the risk level of their driving incidents when they
exceed predefined risk scores.
V. CONCLUSION AND FUTURE WORK
This study captures the uncertainties within driving features
and stereotypes produced by sensor readings and varying
subjective interpretations using a data-driven Fuzzy Logic
System. Our proposed framework provides a risk level for
driving incidents within the range 0 to 100, where 0 is a low
risk driver and 100 is a high risk driver.
To generate the antecedent and consequent MFs, the afore-
mentioned 2-stage classification framework is employed on
four years of HGV driving incidents data (number of harsh
braking events, over speeding duration, excessive throttling
duration and number of over revving events) in the UK i.e.
2014, 2015, 2016 and 2017. A total of eleven stereotypes
were uncovered. Subsequently, the Wang-Mendel method was

Fig. 7: A sample result of our FIS to determine the risk level
of a certain driving behaviour
employed to generate the fuzzy rules and the Mamdani infer-
ence system for scoring HGV drivers in the range 0 to 100.
Although the proposed framework requires further validation
in real time using experts, it can be incorporated into driver
assistance and alert systems for monitoring and providing
feedback to the drivers in real time when they exhibit risky
driving styles.
Lastly, some limitations and scope for future works are
discussed.
Limited number of features: This study is limited in
the number of driving features analysed (only four driving
features). In the future, we plan on including more driving
features such as harsh cornering, lane changing and close
following, which seem to be relevant for road safety.
Other facets of driver behaviour: This study only explores
telematics data. In future, we plan on capturing other facets
of driving behaviour such as driver distraction, workload,
situation awareness, stress, fatigue and attention, to provide
a more reliable holistic view of driving behaviour to assist
stakeholders in the decision making process.
ACK NOW LE DG EM EN T
The first author is supported by the Horizon Centre for
Doctoral Training at the University of Nottingham (UKRI
Grant No. EP/L015463/1) and by Microlise.
REFERENCES
[1] Who. global status report on road safety: World health organization.
2018. Last Accessed: 2-3-2020.
[2] R. Kalsoom and Z. Halim. Clustering the driving features based on data
streams. In Multi Topic Conference (INMIC), 2013 16th International,
pages 89–94, Dec 2013.
[3] C. Saiprasert, S. Thajchayapong, T. Pholprasit, and C. Tanprasert. Driver
behaviour profiling using smartphone sensory data in a v2i environment.
In 2014 Int Conf on Connected Vehicles and Expo, pages 552–557, 2014.
[4] G. P. Figueredo, U. Agrawal, J. M. Mase, M. Mesgarpour, C. Wagner,
D. Soria, J. M. Garibaldi, P.O. Siebers, and R. I. John. Identifying heavy
goods vehicle driving styles in the united kingdom. IEEE Transactions
on Intelligent Transportation Systems (in press), tbc:tbc, 2018.
[5] Utkarsh Agrawal, Jimiama Mafeni Mase, Grazziela P Figueredo, Chris-
tian Wagner, Mohammad Mesgarpour, and Robert I John. Towards real-
time heavy goods vehicle driving behaviour classification in the united
kingdom. In 2019 IEEE Intelligent Transportation Systems Conference
(ITSC), pages 2330–2336. IEEE, 2019.
[6] Jimiama M.M. Mafeni, Peter Chapman, Grazziela P. Figueredo, and
Mercedes T. Torres.
[7] Mohammed S Majdi, Sundaresh Ram, Jonathan T Gill, and Jeffrey J
Rodr´
ıguez. Drive-net: Convolutional network for driver distraction
detection. In 2018 IEEE Southwest Symposium on Image Analysis and
Interpretation (SSIAI), pages 1–4. IEEE, 2018.
[8] Jimiama M.M. Mafeni, Peter Chapman, Grazziela P. Figueredo, and
Mercedes T. Torres. A hybrid deep learning approach for driver
distraction detection. unpublished, 2020.
[9] Jeffrey S Hickman and Richard J Hanowski. Use of a video monitoring
approach to reduce at-risk driving behaviors in commercial vehicle
operations. Transportation research part F: traffic psychology and
behaviour, 14(3):189–198, 2011.
[10] Jennifer L Bell, Matthew A Taylor, Guang-Xiang Chen, Rachel D
Kirk, and Erin R Leatherman. Evaluation of an in-vehicle monitoring
system (ivms) to reduce risky driving behaviors in commercial drivers:
Comparison of in-cab warning lights and supervisory coaching with
videos of driving behavior. Journal of safety research, 60:125–136,
2017.
[11] Hannah J Foy and Peter Chapman. Mental workload is reflected in driver
behaviour, physiology, eye movements and prefrontal cortex activation.
Applied ergonomics, 73:90–99, 2018.
[12] Jimiama M.M. Mafeni, Shazmin Majid, Mohammad Mesgarpour, Mer-
cedes T. Torres, Grazziela P. Figueredo, and Peter Chapman. Evaluating
the impact of heavy goods vehicle driver monitoring and coaching to
reduce risky behaviour. unpublished, 2020.
[13] L-X Wang and Jerry M Mendel. Generating fuzzy rules by learning from
examples. IEEE Trans. Syst., Man, Cybern., 22(6):1414–1427, 1992.
[14] Arshdeep Kaur and Amrit Kaur. Comparison of mamdani-type and
sugeno-type fuzzy inference systems for air conditioning system. In-
ternational journal of soft computing and engineering, 2(2):323–325,
2012.
[15] Zoran Constantinescu, Cristian Marinoiu, and Monica Vladoiu. Driving
style analysis using data mining techniques. Int J of Comp Communi-
cations & Control, 5(5):654–663, 2010.
[16] Adrian B Ellison, SP Greaves, and Rhonda Daniels. Profiling drivers’
risky behaviour towards all road users. In A safe system: expanding the
reach: Australasian College of Road Safety national conference, 2012.
[17] Utkarsh Agrawal, Daniele Soria, Christian Wagner, Jonathan Garibaldi,
Ian O Ellis, John MS Bartlett, David Cameron, Emad A Rakha, and
Andrew R Green. Combining clustering and classification ensembles:
A novel pipeline to identify breast cancer profiles. Artificial Intelligence
in Medicine, 2019.
[18] Lotfi A Zadeh. Fuzzy sets. Information and control, 8(3):338–353,
1965.
[19] Ahmad Aljaafreh, Nabeel Alshabatat, and Munaf S Najim Al-Din.
Driving style recognition using fuzzy logic. In 2012 IEEE International
Conference on Vehicular Electronics and Safety (ICVES 2012), pages
460–463. IEEE, 2012.
[20] T Imkamon, P Saensom, P Tangamchit, and P Pongpaibool. Detection of
hazardous driving behavior using fuzzy logic. In 2008 5th International
Conference on Electrical Engineering/Electronics, Computer, Telecom-
munications and Information Technology, volume 2, pages 657–660.
IEEE, 2008.
[21] Sakchai Boonmee and Poj Tangamchit. Portable reckless driving
detection system. In 2009 6th International Conference on Electrical
Engineering/Electronics, Computer, Telecommunications and Informa-
tion Technology, volume 1, pages 412–415. IEEE, 2009.
[22] Rui Ara´
ujo, ˆ
Angela Igreja, Ricardo De Castro, and Rui Esteves Araujo.
Driving coach: A smartphone application to evaluate driving efficient
patterns. In 2012 IEEE Intelligent Vehicles Symposium, pages 1005–
1010. IEEE, 2012.
[23] Microlise. Microlise (telematics, transport and fleet management solu-
tions). Available at: http://www.microlise.com/, Last accessed 30th April
2018.
[24] Ebrahim H Mamdani and Sedrak Assilian. An experiment in linguistic
synthesis with a fuzzy logic controller. International journal of man-
machine studies, 7(1):1–13, 1975.
[25] Antonio Ginart, Gustavo Sanchez, I Links, and G Back. Fast defuzzi-
fication method based on centroid estimation. Applied Modelling and
Simulation, 58(1):20–25, 2002.