Cluster Analysis for Driver Aggressiveness Identiﬁcation

Fabio Martinelli

, Francesco Mercaldo

, Vittoria Nardone

, Albina Orlando

and Antonella Santone

Istituto di Informatica e Telematica, Consiglio Nazionale delle Ricerche (CNR), Pisa, Italy

Department of Engineering, University of Sannio, Benevento, Italy

Istituto per le Applicazioni del Calcolo “M. Picone”, Consiglio Nazionale delle Ricerche (CNR), Napoli, Italy

Department of Bioscience and Territory, University of Molise, Pesche (IS), Italy

Keywords:

CAN, OBD, Automotive, Safety, Cluster Analysis, Machine Learning.

Abstract:

In the last years, several safety automotive concepts have been proposed, for instance the cruise control and the

automatic brakes systems. The proposed systems are able to take the control of the vehicle when a dangerous

situation is detected. Less effort was produced in driver aggressiveness in order to mitigate the dangerous

situation. In this paper we propose an approach in order to identify the driver aggressiveness exploring the

usage of unsupervised machine learning techniques. A real world case study is performed to evaluate the

effectiveness of the proposed method.

1 INTRODUCTION

As of 2015 there were over 263 million registered

vehicles on the roads in the United States. Of those

millions of registered vehicles, each year there are

also millions of vehicle crashes. In 2015, there were

32,166 fatalities, 1,715,000 injuries and 4,548,000 car

crashes which involved property damage. Of these

fatalities, there are far more driver deaths, than pas-

senger, pedestrian or motorcyclist deaths

. So while

many of us feel secure in vehicles, the statistics in-

dicate the importance of automobile insurance and in

most cases, auto insurance is required by law. This is

the reason why car insurance is really important be-

cause it not only covers any physical damage that may

occur in an accident, but also any damage or injury

that might be caused because of a vehicular accident

or which may be done upon oneself or ones vehicle

by another vehicle or accident, as a falling tree for

example (Marotta et al., 2017).

The insurance industry is a key component of the

economy by virtue of the amount of premiums it col-

lects, the scale of its investment and, more fundamen-

tally, the essential social and economic role it plays

by covering personal and business risks.

Usage-based insurance (UBI) also known as “pay

https://www.statista.com/topics/3087/

car-insurance-in-the-united-states/

as you drive” (PAYD) and “pay how you drive”

(PHYD) and mile-based car insurance is a type of

vehicle insurance whereby the costs are dependent

upon type of vehicle used, measured against time, dis-

tance, behavior and place (Desyllas and Sako, 2013;

Tselentis et al., 2016; Kantor and St

arek, 2014) and

they represent the emerging trend in the insurance

area.

This represents a different approach with respect

to traditional insurance, which attempts to differen-

tiate and reward “safe” drivers, giving them lower

premiums and/or a no-claims bonus. However, con-

ventional differentiation is a reﬂection of history rat-

her than present patterns of behaviour. This means

that it may take a long time before safer (or more

reckless) patterns of driving and changes in lifestyle

feed through into premiums.

UBI programs offer many advantages to insurers,

consumers and society. Linking insurance premiums

more closely to actual individual vehicle or ﬂeet per-

formance allows insurers to more accurately price

premiums (Boquete et al., 2010). This increases af-

fordability for lower-risk drivers, many of whom are

also lower-income drivers. It also gives consumers the

ability to control their premium costs by encouraging

them to reduce miles driven and adopt safer driving

habits. Fewer miles and safer driving also aid in re-

ducing accidents, congestion, and vehicle emissions,

Martinelli F., Mercaldo F., Nardone V., Orlando A. and Santone A.

Cluster Analysis for Driver Aggressiveness Identiﬁcation.

DOI: 10.5220/0006755205620569

In Proceedings of the 4th International Conference on Information Systems Security and Privacy (ICISSP 2018), pages 562-569

ISBN: 978-989-758-282-0

which beneﬁts society

Starting for these considerations, in this paper we

propose an approach able to characterize the driver

behaviour using a set of feature gathered from the

vehicle CAN bus. The proposed method considers the

unsupervised machine learning i.e., the machine le-

arning task of inferring a function to describe hidden

structure from unlabeled data to discriminate between

urban and highway roads. In order to perform this

task, we consider cluster analysis in order to group

the feature extracted from the driver under analysis:

the main assumption that will be veriﬁed in the ex-

periment is that CAN bus features gathered from the

highway path exhibits different values from the ones

gathered from urban road (and for this reason grouped

in different clusters). As a matter of fact, as demon-

strated in current literature, drivers typically exhibit

different driving style in different kind of roads (Me-

har et al., 2013; Wang et al., 2004). Furthermore, on

the basis on the cluster analysis results, we compute

an aggressiveness index of the driver under analysis in

order to boost the research into “pay how you drive”

possible risk assessment calculation.

We evaluate the proposed approach on a real-

world dataset gathered from a vehicle running

through several (urban and highway) roads.

The reminder of the paper is organized as follows:

Section 2 discusses the current literature, Section 3 in-

troduces the method, Section 4 illustrates the results

of the cluster analysis based experiment. Finally, con-

clusions and future works are given in Section 5.

2 RELATED WORK

In the following we section review current literature

related to the driving style and aggressiveness recog-

nition.

In the past, the automotive real-world data retrie-

ving was limited due to the difﬁculty to equip the sen-

sors in cars, since the introduction of CAN this limit

is overcome.

Authors in (Wakita et al., 2006) propose a dri-

ver identiﬁcation method that is based on the driving

behavior signals that are observed while the driver

is following another vehicle. They analyze signals,

as accelerator pedal, brake pedal, vehicle velocity,

and distance from the vehicle in front, were measu-

red using a driving simulator. The identiﬁcation rates

were 81% for twelve drivers using a driving simulator

and 73% for thirty drivers.

http://www.naic.org/cipr topics/

topic usage based insurance.htm

Data from the accelerator and the steering wheel

were analyzed by researchers in (Zhang et al., 2014).

Observing the considered features, they employ hid-

den Markov model (HMM) to model the driver cha-

racteristics. They build two models for each dri-

ver, one trained from accelerator data and one lear-

ned from steering wheel angel data. The models can

be used to identify different drivers with an accuracy

equal to 85%.

Researchers in (Kedar-Dongarkar and Das, 2012)

classify a set of features extracted from the powertrain

signals of the vehicle, showing that their classiﬁer is

able to classify the human driving style based on the

power demands placed on the vehicle powertrain with

an overall accuracy of 77%.

Van Ly et alius (Van Ly et al., 2013) explore the

possibility of using the inertial sensors of the vehicle

from the CAN bus to build a proﬁle of the driver ob-

serving braking and turning events to characterize an

individual compared to acceleration events.

Researchers in (Miyajima et al., 2007; Nishiwaki

et al., 2007) model gas and brake pedal operation

patterns with Gaussian mixture model (GMM). They

achieve an identiﬁcation rate of 89.6% for a driving

simulator and 76.8% for a ﬁeld test with 276 drivers,

resulting in 61% and 55% error reduction, respecti-

vely, over a driver model based on raw pedal opera-

tion signals without spectral analysis.

Driver behavior is described and modeled in (Choi

et al., 2007) using data from steering wheel angle,

brake status, acceleration status, and vehicle speed

through Hidden Markov Models (HMMs) and GMMs

employed to capture the sequence of driving characte-

ristics acquired from the CAN bus information. They

obtain 69% accuracy for action classiﬁcation, and

25% accuracy for driver identiﬁcation.

In reference (Meng et al., 2006) the features ex-

tracted from the accelerator and brake pedal pressure

are used as inputs to a fuzzy neural network (FNN)

system to ascertain the identity of the driver. Two

fuzzy neural networks, namely, the evolving fuzzy

neural network (EFuNN) and the adaptive network-

based fuzzy inference system (ANFIS), are used to

demonstrate the viability of the two proposed feature

extraction techniques.

A hidden-Markov-model-(HMM)-based simila-

rity measure is proposed in (Enev et al., 2016) in order

to model driver human behavior. They employ a si-

mulated driving environment to test the effectiveness

of the proposed solution.

Authors in (Kwak et al., 2016) propose a met-

hod based on driving pattern of the car. They con-

sider mechanical feature from the CAN vehicle eva-

luating them with four different classiﬁcation algo-

Figure 1: Flow diagram of the proposed approach for risk

index computation.

rithms, obtaining respectively an accuracy equal to

0.939 with Decision Tree, equal to 0.844 with KNN,

equal to 0.961 with RandomForest and equal to 0.747

using MLP algorithm.

The proposed method, differently from the ones

explained in this section, consider the road identiﬁca-

tion issue in order to compute two different aggres-

siveness indexes (related to the urban and to highway

roads). In addition, relating to road identiﬁcation task,

we highlight that the classiﬁcation task does not re-

quire previous knowledge about the type of road since

a cluster analysis algorithm is considered.

3 THE METHOD

In following section we describe the considered ap-

proach in order to evaluate the driving style (in terms

of aggressiveness) from a set of features extracted

from the in-vehicle CAN data and from GPS sensor.

Figure 1 depicts the ﬂow diagram of the proposed

approach for the road identiﬁcation.

As Figure 1 shows, the cluster analysis process is

concerned to OBD data (Martinelli et al., 2018) gat-

hered by the CAN bus in order to label the data as

belonging to the urban or to the highway roads.

In our analysis we considered the feature set (be-

longing to OBD and to GPS) shown in Table 1.

We considered features gathered from different

sources: the ﬁrst source is represented by the OBD

(i.e. F1, F2, F3, F4, F5, F6 and F7) while the second

one is computed by the user device GPS sensor (i.e.

F8 and F9).

The GPS sensor features are considered to add

meta information useful to have the conﬁrmation

about the kind of road (i.e., urban or highway) identi-

ﬁed by the cluster analysis using the F8 and F9 featu-

res.

As stated into the introduction in order to charac-

terize the driver style in terms of aggressiveness we

resort to an unsupervised machine learning approach

i.e., cluster analysis that, differently from the supervi-

sed machine learning does not require a labelled da-

taset to perform the classiﬁcation task (Cimitile et al.,

2017; Canfora et al., 2014; Canfora et al., 2016; Bat-

tista et al., 2016; Mercaldo et al., 2016). The cluster

analysis itself is not one speciﬁc algorithm, but the ge-

neral task to be solved: it can be achieved by various

algorithms that differ signiﬁcantly in their notion of

what constitutes a cluster and how to efﬁciently ﬁnd

them (Kaufman and Rousseeuw, 2009). Popular no-

tions of clusters include groups with small distances

among the cluster members, dense areas of the data

space, intervals or particular statistical distributions.

In this paper we consider the k-means algorithm

(MacQueen et al., 1967), one of the simplest unsuper-

vised learning algorithms that solve the well known

clustering problem (Har-Peled and Kushal, 2007).

The procedure follows a simple and easy way to

classify a given data set through a certain number of

clusters (assume k clusters, with particular regards to

the designed approach we consider k=2) ﬁxed a pri-

ori. The main idea is to deﬁne 2 centroids, one for

each cluster. These centroids should be placed in a

cunning way because of different location causes dif-

ferent result (Jain, 2010). So, the better choice is to

place them as much as possible far away from each

other. The next step is to take each point belonging to

a given data set and associate it to the nearest centroid.

When no point is pending, the ﬁrst step is completed

and an early groupage is done (Arthur et al., 2009). At

this point we need to re-calculate k new centroids as

barycenter of the clusters resulting from the previous

step. After we have these k new centroids, a new bin-

ding has to be done between the same data set points

and the nearest new centroid (MacQueen et al., 1967).

A loop has been generated. As a result of this loop we

may notice that the k centroids change their location

step by step until no more changes are done. In other

words centroids do not move any more.

We consider the k-means implementation in the

Weka data mining toolkit

i.e., SimpleKMeans. This

implementation can use either the Euclidean distance

(default) or the Manhattan distance. Euclidean, in

https://www.cs.waikato.ac.nz/ml/weka/

Table 1: Features involved in the study.

Feature Description Info OBD GPS

F1 Engine RPM Revolutions Per Minute X

F2 Mass Air Flow expressed in g/s X

F3 Instantaneous Fuel Consumption expressed in liters/100 km X

F4 Boost pressure estimation expressed in KPa/Bar/Kg X

F5 Acceleration expressed as g (gravity) X

F6 Engine power expressed in KW X

F7 Engine torque expressed in NM/Kg X

F8 Altitude expressed in degree X

F9 Longitude expressed in degree X

this study we set the SimpleKMeans algorithm with

the Euclidean distance, maximum iterations number

equal to 500 and maximum of generated clusters

equal to 2. Since the features given to the learner are

unlabeled, there is no evaluation of the accuracy of the

structure that is output by the relevant algorithm (this

is one way of distinguishing unsupervised learning

from supervised learning): for this reason we con-

sider the incorrectly clustered instances number and

percentage in order to evaluated the goodness of the

proposed method (i.e., to evaluate whether the ﬁrst

cluster contains the majority of urban while the se-

cond one contains the majority of highway ones).

Once evaluated the k-means algorithm in order to

distinguish between feature gathered while the dri-

ver is traveling on urban roads and features gathered

while the driver is traveling on highway ones, we dis-

cuss an approach to apply this information to provide

an aggressiveness index for the PHYD car insurance.

4 EXPERIMENTAL EVALUATION

In this section we discuss the experiment we perfor-

med related to cluster analysis in order to classify be-

tween urban and highway paths.

The evaluation consists of two stages: (i) a com-

parison of descriptive statistics of the populations of

feature and (ii) an unsupervised classiﬁcation analy-

sis aimed at assessing whether the urban and highway

features are grouped in different clusters.

We realize a real-world dataset, gathering data

from the in-vehicle CAN bus. The vehicle involved

in the experiment is a Fiat Punto Evo 1.3 Diesel with

75 horsepowers with one driver.

In order to collect data, the DashCommand (OBD

ELM App)

application and Mini Bluetooth ELM327

OBD 2 Scanner were used.

OBD is available on modern car to produce the

self-diagnostic report by monitoring vehicle system in

https://play.google.com/store/apps/

details?id=com.palmerperformance.DashCommand&hl=it

terms of measurement and vehicle failure (Martinelli

et al., 2017). The data are recorded every 1 second

during driving using the DashCommand application

by an Android smartphone (i.e., a Huawei p8 lite 2017

with Android 7.0 Nougat onboard) ﬁxed in the car by

a support.

In order to label the track using the “urban” or

the “highway” label we developed a Java script able

to generates an address from a latitude and longitude

through the reverse geocoding Java wrapper

able to

query the Nominatim search engine for OpenStreet-

Map data

We collected data from the vehicle in an urban an

a highway area in Italy, in Figure 2 the urban path con-

sidered: it consists of 22 Km from the Istituto di Infor-

matica e Telematica in Pisa to Cascina, in the center

of Italy. The highway path 3 is related to the main

Italian highway (the A1, i.e. Autostrada del Sole) be-

tween the Center and the South of Italy and it consists

of 234 Km. In order to balance to traveled kilometers

between the urban and the highway paths, we have

considered 10 urban paths (i.e., ten different travels

of the urban path of 22 Km) and one highway path:

in this way we have a dataset composed of 220 Km of

urban path and 234 Km of highway path.

We represent two scatterplots with the aim to give

statistical evidence that considered feature popula-

tion exhibit different trend between the urban and the

highway populations. Similar consideration can ad-

dressed for the other features.

Figure 4 shows the scatterplot related to the En-

gine RPM (i.e., the F1 feature) and Boost pressure

estimation (i.e. the F4 feature): Engine RPM feature

is represented on the X axis while the Boost pressure

estimation one on the Y axis.

The red distribution is related to the urban path,

while the blue one is related to the highway path:

from the scatterplot it is clear the division between

the red points, mostly allocated on the center-low left

https://www.daniel-braun.com/technik/

reverse-geocoding-library-for-java/

http://nominatim.openstreetmap.org/

Figure 2: The urban path considered in the study highlig-

hted in blue: it consists of 22 Km from the Istituto di In-

formatica e Telematica in Pisa to Cascina, in the center of

Italy.

Figure 3: The highway path considered in the study high-

lighted in blue: it is related to the main Italian highway be-

tween the Center and the South of Italy and it consists of

234 Km.

side of the scatterplot, and the blue one, mostly allo-

cated on the high and low right side of the scatterplot.

Figure 5 shows the scatterplot related to the F1

feature and the F7 one i.e., the Engine torque: the F1

feature is represented on the X axis while the F7 one

on the Y axis.

As in Figure 4, also in the scatterplot in Figure 5

the red distribution is related to the urban path, while

Figure 4: Scatterplot related to the F1 feature and the F4 fe-

ature (the red distribution is related to the urban path, while

the blue distribution in related to the highway path).

Figure 5: Scatterplot related to the F1 feature and the F7 fe-

ature (the red distribution is related to the urban path, while

the blue distribution in related to the highway path).

the blue one is related to the highway path. In this

case both the red and the blue distributions are allo-

cated in the down side of the graph, but also in this

case their division between is clear: the red points are

in the left and middle part of the scatterplot, while the

blue points are most allocated in the left side.

From the considerations related to Figures 4 and

5 we state that the features under analysis can be use-

ful to discriminate between urban and highway paths

and, consequently, good candidates for the cluster

analysis phase.

Relating the unsupervised classiﬁcation, we com-

pute the incorrectly clustered instances number and

percentage in three different scenarios (i.e., we per-

form three different clustering experiments) with fol-

lowing instances:

• C1: instances related only to the urban path;

• C2: instances related only to the highway path;

• C3: instances related to the urban and highway

path (i.e., the full dataset);

The aim on this experiment is to demonstrate that

the cluster analysis is useful to discriminate between

urban and highway path (the C3 scenario) producing

two different clusters, while the performances obtai-

ned in the unsupervised classiﬁcation related to ur-

ban (i.e., C1) and highway (i.e., C2) paths are not

able to obtain a good clustering (i.e., if we ﬁx the

number of clusters to generate as two, when we con-

sider only urban/highway instances the cluster algo-

rithm is not able to create the ﬁrst cluster with all the

urban/highway instances and to generate the second

cluster with few instances).

We consider three difference instance set (i.e., C1,

C2 and C3) with the aim to demonstrate that the more

appropriate clusters are obtained using the C3 instan-

ces (related to the urban and highway path).

Table 2 shows the results of the C1, C2 and C3

unsupervised classiﬁcations.

As shown in Table 2, the C1 experiment (with only

urban path instances) obtains an Incorrectly cluste-

Table 2: Results of the C1, C2 and C3 experiments.

Exp. ICI % time

C1 5551 63.4545% 0.06

C2 8735 83.5437% 0.09

C3 444 2.4636% 0.06

red instances value equals to 5551 (i.e., 63% of the

instances considered); the C2 experiment (with only

highway path instances) obtains an Incorrectly clus-

tered instances (ICI) value equal to 8735 (i.e., 83%

of the instances considered) while the C3 experiment

(with both urban and highway paths instances) obtain

an Incorrectly clustered instances value equals to 444

with a percentage of incorrectly clustered instances

equal to 2%.

These results demonstrate that the adoption of the

unsupervised machine learning techniques is promi-

sing: as a matter of fact, considering the different

driving styles that should be adopted in urban and

highway roads, we can consider the Incorrectly clus-

tered instances value as an estimator of the driving

style. In case this value is low, the driver exhibits

a different driving style between urban and highway

paths and this is the result of the different driving style

that should be adopted in different roads. From the ot-

her side, whether the Incorrectly clustered instances

value exhibits an high value (for instance, in the C1

and C2 experiment), as we demonstrated cluster ana-

lysis is not able to correctly deﬁne the clusters (C1

and C2 experiment), and this is symptomatic that the

driver under analysis exhibits a driving style pretty si-

milar in urban and highway roads and the feature set

considered is representative of the kind of traveled ro-

ads.

Once obtained the clusters with regards to the ur-

ban and to the highway path, in order to compute

the driver aggressiveness index in urban and highway

paths we consider the acceleration feature (i.e., F5)

variation: this is the reason why we resort to the

standard deviation statistical dispersion index i.e., an

estimate of the variability of a data population or a

random variable (in this case the variable is represen-

ted by the F5 feature).

Considering u

the value of the i-th urban path

occurrence of the F5 feature, N

the total number

of urban path occurrences of the F5 feature (with

1 ≤ i ≤ N

) we deﬁned the driver aggressiveness in-

dex σ

urban

in urban path as follows:

urban

∑

i=1

− ¯x

urban

)

where ¯x

urban

represents the arithmetic mean of F5

feature urban path distribution and it is deﬁned as:

¯x

urban

∑

i=1

Relating to the driver aggressiveness index

highway

in highway path, considering h

the value of

the k-th highway path occurrence of the F5 feature, N

the total number of highway path occurrences of the

F5 feature (with 1 ≤ i ≤ N

), we deﬁne the σ

highway

index as follows:

highway

∑

i=1

− ¯x

highway

)

where ¯x

highway

represents the arithmetic mean of

F5 feature highway path distribution and it is deﬁned

as:

¯x

highway

∑

i=1

We compute the driver aggressiveness indexes:

with regards to the urban path σ

urban

we obtained fol-

lowing value:

urban

= 6.4734

while relating to the driver aggressiveness index

in highway path we obtain following value:

highway

= 2.4519

From these results we deduce that the driver under

analysis exhibit a drive style more aggressive in the

urban path (with σ

urban

= 6.4734) than in the highway

one (i.e., σ

urban

= 2.4519).

This behaviour can be considered as normal, as

a matter of fact, typically urban roads require more

accelerations and decelerations if compared to the

highway ones.

The opposite behavior would be considered highly

aggressive.

5 CONCLUSION AND FUTURE

WORK

In this paper, starting from the consideration that

depending on the type of road run (i.e., urban and

highway) drivers adopt different driving style, we pro-

pose an approach to compute the driver aggressive-

ness. The aim of the proposed approach is to identify

the kind of road traveled through unsupervised ma-

chine learning in order to identify the driver aggressi-

veness index is urban and highway paths. In order to

evaluated the cluster analysis method to discern bet-

ween urban and highway data, we use a set of features

extracted from the CAN bus of real-world car while

traveling in different roads (i.e., urban and highway)

in the center and south of Italy. As future work we

plan to adopt formal veriﬁcation techniques aimed to

identify whether a driver can be classiﬁed in several

predeﬁned categories (for instance: the young driver,

the ruthless driver, the cautious driver) in order to pro-

pose a risk index considering the category to which a

driver belongs.

ACKNOWLEDGMENTS

This work has been partially supported by H2020

EU-funded projects NeCS and C3ISP and EIT-Digital

Project HII and PRIN “Governing Adaptive and Un-

planned Systems of Systems” and the EU project Cy-

berSure 734815.

REFERENCES

Arthur, D., Manthey, B., and R

oglin, H. (2009). k-means

has polynomial smoothed complexity. In Foundations

of Computer Science, 2009. FOCS’09. 50th Annual

IEEE Symposium on, pages 405–414. IEEE.

Battista, P., Mercaldo, F., Nardone, V., Santone, A., and

Visaggio, C. A. (2016). Identiﬁcation of android mal-

ware families with model checking. In ICISSP, pages

542–547.

Boquete, L., Rodr

ıguez-Ascariz, J. M., Barea, R., Cantos,

J., Miguel-Jim

enez, J. M., and Ortega, S. (2010). Data

acquisition, analysis and transmission platform for a

pay-as-you-drive system. Sensors, 10(6):5395–5408.

Canfora, G., Medvet, E., Mercaldo, F., and Visaggio, C. A.

(2016). Acquiring and analyzing app metrics for ef-

fective mobile malware detection. In Proceedings of

the 2016 ACM on International Workshop on Security

And Privacy Analytics, pages 50–57. ACM.

Canfora, G., Mercaldo, F., Visaggio, C. A., and Di Notte, P.

(2014). Metamorphic malware detection using code

metrics. Information Security Journal: A Global Per-

spective, 23(3):57–67.

Choi, S., Kim, J., Kwak, D., Angkititrakul, P., and Han-

sen, J. H. (2007). Analysis and classiﬁcation of dri-

ver behavior using in-vehicle can-bus information. In

Biennial Workshop on DSP for In-Vehicle and Mobile

Systems, pages 17–19.

Cimitile, A., Martinelli, F., and Mercaldo, F. (2017). Ma-

chine learning meets ios malware: Identifying mali-

cious applications on apple environment. In ICISSP,

pages 487–492.

Desyllas, P. and Sako, M. (2013). Proﬁting from business

model innovation: Evidence from pay-as-you-drive

auto insurance. Research Policy, 42(1):101–116.

Enev, M., Takakuwa, A., Koscher, K., and Kohno,

T. (2016). Automobile driver ﬁngerprinting.

Proceedings on Privacy Enhancing Technologies,

2016(1):34–50.

Har-Peled, S. and Kushal, A. (2007). Smaller coresets for

k-median and k-means clustering. Discrete & Com-

putational Geometry, 37(1):3–19.

Jain, A. K. (2010). Data clustering: 50 years beyond k-

means. Pattern recognition letters, 31(8):651–666.

Kantor, S. and St

arek, T. (2014). Design of algorithms

for payment telematics systems evaluating driver’s

driving style. Transactions on Transport Sciences,

7(1):9.

Kaufman, L. and Rousseeuw, P. J. (2009). Finding groups

in data: an introduction to cluster analysis, volume

344. John Wiley & Sons.

Kedar-Dongarkar, G. and Das, M. (2012). Driver classiﬁ-

cation for optimization of energy usage in a vehicle.

Procedia Computer Science, 8:388–393.

Kwak, B. I., Woo, J., and Kim, H. K. (2016). Know your

master: Driver proﬁling-based anti-theft method. In

PST 2016.

MacQueen, J. et al. (1967). Some methods for classiﬁcation

and analysis of multivariate observations. In Procee-

dings of the ﬁfth Berkeley symposium on mathematical

statistics and probability, volume 1, pages 281–297.

Oakland, CA, USA.

Marotta, A., Martinelli, F., Nanni, S., Orlando, A., and

Yautsiukhin, A. (2017). Cyber-insurance survey.

Computer Science Review.

Martinelli, F., Mercaldo, F., Nardone, V., Orlando, A., and

Santone, A. (2018). Who’s driving my car? a machine

learning based approach to driver identiﬁcation. In

ICISSP.

Martinelli, F., Mercaldo, F., Nardone, V., and Santone, A.

(2017). Car hacking identiﬁcation through fuzzy lo-

gic algorithms. In Fuzzy Systems (FUZZ-IEEE), IEEE

International Conference on. IEEE. IEEE.

Mehar, A., Chandra, S., and Velmurugan, S. (2013). Speed

and acceleration characteristics of different types of

vehicles on multi-lane highways. European Trans-

port, 55:1825–3997.

Meng, X., Lee, K. K., and Xu, Y. (2006). Human driving

behavior recognition based on hidden markov models.

In Robotics and Biomimetics, 2006. ROBIO’06. IEEE

International Conference on, pages 274–279. IEEE.

Mercaldo, F., Nardone, V., Santone, A., and Visaggio,

C. A. (2016). Download malware? no, thanks. how

formal methods can block update attacks. In For-

mal Methods in Software Engineering (FormaliSE),

2016 IEEE/ACM 4th FME Workshop on, pages 22–

28. IEEE.

Miyajima, C., Nishiwaki, Y., Ozawa, K., Wakita, T., Itou,

K., Takeda, K., and Itakura, F. (2007). Driver mo-

deling based on driving behavior and its evaluation

in driver identiﬁcation. Proceedings of the IEEE,

95(2):427–437.

Nishiwaki, Y., Ozawa, K., Wakita, T., Miyajima, C., Itou,

K., and Takeda, K. (2007). Driver identiﬁcation based

on spectral analysis of driving behavioral signals. In

Advances for In-Vehicle and Mobile Systems, pages

25–34. Springer.

Tselentis, D. I., Yannis, G., and Vlahogianni, E. I. (2016).

Innovative insurance schemes: pay as/how you drive.

Transportation Research Procedia, 14:362–371.

Van Ly, M., Martin, S., and Trivedi, M. M. (2013). Driver

classiﬁcation and driving style recognition using iner-

tial sensors. In Intelligent Vehicles Symposium (IV),

2013 IEEE, pages 1040–1045. IEEE.

Wakita, T., Ozawa, K., Miyajima, C., Igarashi, K., Ka-

tunobu, I., Takeda, K., and Itakura, F. (2006).

Driver identiﬁcation using driving behavior signals.

IEICE TRANSACTIONS on Information and Systems,

89(3):1188–1194.

Wang, J., Dixon, K., Li, H., and Ogle, J. (2004). Nor-

mal acceleration behavior of passenger vehicles star-

ting from rest at all-way stop-controlled intersecti-

ons. Transportation Research Record: Journal of the

Transportation Research Board, (1883):158–166.

Zhang, X., Zhao, X., and Rong, J. (2014). A study of

individual characteristics of driving behavior based

on hidden markov model. Sensors & Transducers,

167(3):194.