Research A rticle
Dynamic Recognition of Driver’s Propensity Based on GPS
Mobile Sensing Data and Privacy Protection
Xiaoyuan Wang,
1,2
Jianqiang Wang,
2
Jinglei Zhang,
1
andJinghengWang
3
1
School of Transportation and Vehicle Engineering, Shandong U n iversity of Technology, Zibo 255049, China
2
StateKeyLaboratoryofAutomotiveSafetyandEnergy,TsinghuaUniversity,Beijing100084,China
3
Shandon g Zibo Experimental High School, Zibo 255000, China
Correspondence should be addressed to Xiaoyuan Wang; wangxiaoyuan@sdut.edu.cn
and Jianqiang Wang; wjqlws@tsinghua.edu.cn
Received March ; Revised August ; Accepted August
Academic Editor: Huaguang Zhang
Copyright © Xiaoyuan Wang et al. is is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Driver’s propensity is a dynamic measurement of driver’s emotional preference characteristics in driving process. It is a core
parameter to compute driver’s intention and consciousness in safety driving assist system, especially in vehicle collision warning
system. It is also an important inuence factor to achieve the Driver-Vehicle-Environment Collaborative Wisdom and Control
macroscopically. In this paper, dynamic recognition model of driver’s propensity based on support vector machine is established
taking the vehicle safety controlled technology and respecting and protecting the driver’s privacy as preco ndition. e experiment
roads travel time obtained through GPS is taken as the characteristic parameter. e sensing information of Driver-Vehicle-
Environment was obtained through psychological questionnaire tests, real vehicle experiments, and virtual driving experiments,
and the information is used for parameter calibration and validation of the model. Results show that the established recognition
model of driver’s propensity is reasonable and feasible, which can achieve the dynamic recognition of driver’s propensity to some
extent. e recognition model provides reference and theoretical basis for personalized vehicle active safety systems taking people
as center especially for the vehicle safety technology based on the networking.
1. Introduction
e inuence of drivers’ physiology and psychology charac-
teristics on trac safety is mainly represented as the driver’s
propensity []. Dr iver’s propensity is the attitude experience
of the drivers for the real trac conditions aected by various
dynamic factors, as well as the preference drivers show that
suits with decision or behavior value. It is the dynamic
measure of driver’s emotional preference characteristics in
driving process which guides the drivers’ intent and aects
vehicle handling behavior directly. It is a core parameter
tocomputedriver’sintentionandconsciousnessinsafety
driving assist system, especially vehicle collision warning sys-
tem. I t is also an important inuence factor to achieve
the Driver-Vehicle-Environment Collaborative Wisdom and
Control macroscopically, especially the vehicle safety con-
trolled technology. Its types can be divided into radical t ype,
common type, and conservative type. e previous studies
were mainly focused on drivers’ psychological characteristics
measurement on relatively static and macrocosmic factors
and the trac safety eect; few research works have been
conducted on microcosmic, dynamic measurement and com-
putation of driver’s emotion state in view of vehicle active
safety. Xie and Wang [] constructed a simple trac model
including car-following, lane-changing, and overtaking with
consideration of driver’s emotion under simplied road con-
ditionanddiscussedtheinuenceofthechangeofcognition
emotion on driving strategy; Wu and Hu [, ] studied the
driver’s behavior characteristics caused by anger, identied
the state of angr y driving, and researched the inuence on
driver’s physiology, psychology, and trac security aected
by angry emotion, but there was no research concerning
the microscopic and dynamic characteristics of the time-
varying emotions. Set about aective computing, Lin and
Hindawi Publishing Corporation
Mathematical Problems in Engineering
Volume 2016, Article ID 1814608, 12 pages
http://dx.doi.org/10.1155/2016/1814608
Mathematical Problems in Engineering
Feng [] analyzed and judged whether the drivers were in
the aggressive driving state or not using the speech emotion
recognition, facial expression recognition, vehicle driving
state detection methods and so on and used the vehicular
intelligent aided system to make decision of the security
alert or safety anticollision for vehicles, to improve driving
safety factor. In order to uncover whether emotional auditory
stimuli can aect risky behavior in hazardous situations,
Di Stasi et al. [] organized that forty-nine volunteers rode
a motorcycle in a virtual environment and went through
a number of preset risky scenarios, some of which were
cued by a kind of sound (beep, positive emotional sound,
or negative sound). Results showed that the beep reduced
frequency of accidents in upcoming risky situation, while
the emotional cues did not. Likewise, the beep induced
the drivers to decrease their speed and focus their gaze on
relevant areas of the visual eld, while the emotional sounds
did not. ese results suggest that auditory warning systems
for vehicles should avoid using emotion-laden sounds, as
their aective content might diminish their utility to increase
driving alertness. Trick et al. [] discussed implications
for both basic research on attention-emotion and applied
research on driving. e researchers focus on the eect of
eetingemotionsonhazardperceptionandsteeringwhile
driving. Taubman-Ben-Ari [] studied the eects of positive
emotion priming on the willingness to drive recklessly. e
research shows that positive emotions of a relaxing nature,
as well as thinking abou t the meaning of life, lowered the
willingness to engage in risky driving. Many risk factors such
as speed, drowsiness, drugs and alcohol consumption, and
stateofthecarhavebeenidentiedandhavepermittedthe
development of prevention policies. In contrast psychological
factors remain poorly studied, particularly emotional state.
Fo r this phenomenon, M’bailara et al. [, ] researched
the relationship between accident and emotional state. e
results revealed that emotional reactivity is signicantly
associated with the drivers’ responsibility, suggesting that
emotional hypo- or hyperactivity is a signicant source of
accidents. Arnau-Sabat
´
es et al. [] studied those emotional
abilities as predictors of risky driving behavior. e risky
driving attitudes and emotional abilities of future driving
instructors were measured. e results demonstrate that risky
attitudes correlate negatively with emotional abilities. e
results obtained can be used to formulate the corresponding
prevention programs to reduce risky driving behaviors. Chan
and Singhal [] analyzed the eects of emotional distraction
on driving. is purpose was achieved using a driving
simulator and three dierent types of emotional i nformation:
neutral, negative, and positive emotional words. Participants
also responded to target words while driving and completed
a surprise-free r ecall task of all the words at the end of the
study. e ndings suggest that emotional distraction can
modulate attention and decision-making abilities and have
adverse impacts on driving behavior.
In order to ll the vacancy of related research about
microcosmic, dynamic measurement and calculation of
driver emotion, Zhang et al. [, ] from the perspectives of
vehiclesafedrivingsupportsystem,especiallyintheautomo-
tive collision warning system driver’s intention, emotion, and
other psychological eects associated with coupling aware-
ness computing core research scientic issues, collaborative
applications car GPS, car laser radar and other dynamic data
acquisition board video sensor system, variable data capture
vehicles environment, freedom of driving, the car following
a complex and multilane vehicle under the cluster grouping
of specic trac scene when the driver becomes law emotion
and collaborate deduction, especially for emotional dynamic
measure, online real-time identication and characterization
of key scientic issues, such as exploratory research. However,
the study also found that, for the LiDAR data acquisition, due
to its precision optical characteristics LiDAR is expensive, its
installation is relatively complex, seismic immunity is weak,
and it is prone to some error so the penetration rate is not
high and experiments and subsequent data process are more
cumbersome.
With GPS users increasing, especially the popularity
of mobile GPS, researchers make it possible to excavate
breadth and depth of trac information GPS data. GPS has
the advantages of low cost, high penetration, being easy to
carry, zero invasive eect essentially for drivers, antivibration,
and strong anti-interference. Data is paid more and more
attention by trac researchers. However, GPS data (such as
location, trajectory) is easy to leak travelers hobbies, behavior
patterns, habits, and other personal privacy. Date of GPS is
oen used by malicious attackers to detect and analyze the
current position of the traveler, the last visited location, home
address, working place, income levels, health status, political
aliation, etc.; it even can cause threat to personal safety of
travelers. us, many scholars at home and abroad strengthen
the protection of privacy on the GPS basic research in the
promotion of GPS applications at the same time.
GPS, especially the popularity of mobile phones and
portable GPS, makes the concept of synergy based on vehicle
road trac management and control possible and provides
secure innovative ideas for the dangerous state detection
based on driver and vehicle dynamic security technology of
shared control by vehicle network. is technology can make
use of nonvehicle personnel (such as trac police and the
driver relatives) of the vehicle to reach remote sharing and
control interventions (such as advice to or warning of the
driver, and the implementation of open warning light on
the vehicle, limiting-velocity, trac controls, pull-over, and
other remote auxiliary controls), to solve the trac safety
problem when the driver is in danger. Meanwhile, real-time
network-based storage and sharing driver status information
can provide reference for accident analysis and tracking
accountability to reduce dangerous driving behaviors of the
drivers.
To meet the needs of vehicle road collaborative innova-
tion technology, especially in shared controlled vehicle study
security innovation vehicle networking, the GPS semantic
mining and privacy preserving are collaboratively consid-
ered, and travel time is selected as measure basis; drivers’
tendency dynamic identication theory is studied in this
paper. We get dynamic data of human vehicle and environ-
ment which is corresponding to dierent propensity (such
as radical type, common type, and conservative type) driver
through the psychological test, virtual driving experiment,
Mathematical Problems in Engineering
and real vehicle experiment, analyze data and extract the
travel time as the parameter to build dynamic recogni-
tion model of driver’s propensity based on support vector
machine, and use experimental data for model parameter
calibration, verication, and analysis.
2. Dynamic Recognition of Driver’s Propensity
2.1. Support Vector Machine. Support vector machine (SVM)
is a kind of new machine learning method proposed by
Vapnik [–], based on the statistical learning theory,
has a complete foundation of statistical learning theory
and excellent learning performance, and is the youngest
content of statistical learning theory and the most practical
part. C ontrasting with the heuristic neural network learning
method and the imp lementation, SVM has a more rigorous
theoretical and mathematical foundation; local minimum
problem does not exist. e technology can solve the practical
problems of small sample, high dimension, nonlinear and
local minimum points, and so forth and keep good general-
ization capability in the small sample conditions, successfully
applied in signal processing, regression analysis, function
approximation, and other elds [–].
2.1.1. e Basic Idea of Support Vector Machine. e basic
idea of SVM method is based on structural risk minimization
(structural r isk minimization, SRM) principle. rough a
specic nonlinear mapping, the sample space is mapped to
a high dimension as feature space of innite dimensional
(Hilbert space). In the feature space, to nd the optimal
classication or regression linear hyperplane, the plane is
taken as the classication decision surface, so as to solve
theproblemsinsamplespaceofnonlinearclassicationand
regression.
e learning p rocess and model selection phase are two
important aspec ts of the SVM algorithm. Reference []
“Tikhonov, Ivanov and Morozov regularization for Support
Vector Machine Learning” introduces the learning method
of training support vector machine in consideration of
structural risk minimization, comparing the advantages and
disadvantages of three kinds of regularization algorithm;
this paper chooses the appropriate regularization algorithm
to nish the learning process of SVM based on achieve-
ments of [] for achieving unication of the algorithm
eectiveness and operability. Reference [] “In-Sample and
Out-of-Sample Model Select ion and Error Estimation for
Support Vector Machines” details common methods of SVM
model selection phase, introducing the dierence between in-
sample and out-of-sample and application condition of the
two methods. is paper perfects the SVM model selection
phase of the dynamic recognition of driver’s propensity based
on travel time in order to make the model more reasonable
and reliable.
e decision function of SVM is only determined by a
few support vectors; the computational complexity depends
on the number of support vectors, rather than the dimension
of the sample space, which avoids the “dimension disaster in
some sense.”
H1
H
H2
sv
sv
sv
sv
sv
sv
sv
Margin =
2
‖𝜔‖
F : Optimal classication plane diagram of support vector
machine.
2.1.2. Support Vector Machine Model. If there is a hyperplane
suchthattwokindsofdatacanbeclassiedandthedistance
between the date and the hyperplane is the largest, the plane
is the optimal hyperplane.
(1) Linear Optimal Classication Hyperplane. e training
sample consists of two categories of a given group (
1
,
1
),
(
2
,
2
),...,(
𝑛
,
𝑛
),where
𝑖
∈
𝑛
and
𝑖
∈{−1,1},if
𝑖
belongs to the rst category, so
𝑖
is labeled as positive
(
𝑖
=1);otherwise,
𝑖
=−1and = 1,2,...,. Specic
ideas are shown in Figure . In Figure , circular and square,
respectively, represent two types of samples, H is classication
hyper plane, H and H, respectively, represent the various
types and are samples nearest to H and the plane is parallel
to H, and the distance between them is called classication
interval (margin). e so-called optimal classication face
is required to correctly separate two kinds with farthest
classication interval. e sample point of H and H is
support vector.
If the sample is linearly separable, there is a hyperplane
H:
⋅+=0.
()
Make
⋅
𝑖
+≥1,
𝑖
=1,
⋅
𝑖
+≤−1,
𝑖
=−1.
()
e formula () is normalized, the linearly separable data
meet:
𝑖
[
(
⋅
)
+
]
≥1, =1,2,...,.
()
According to the optimal hyperplane denition, classi-
cation interval can b e expressed as
= min
{𝑥
𝑖
,𝑦
𝑖
=1}
⋅
𝑖
+
+ min
{𝑥
𝑗
,𝑦
𝑗
=−1}
⋅
𝑗
+
=
2
,
=1,2,...,.
()
Mathematical Problems in Engineering
Make the max optimal 2/,so(1/2)or (1/2)
2
is
the min. A linear support vector machine is transformed into
the problem of solving t he following two convex programing
problems:
min
1
2
2
Constraint conditions :
𝑖
⋅
𝑖
+≥1,
=1,2,...,.
()
e optimal solution can be obtained by the following
Lagrange function:
(
,,
)
=
1
2
2
−
𝑛
𝑖=1
𝑖
𝑖
⋅
𝑖
+−1, =1,2,...,,
()
where
𝑖
≥0(=1,2,...,)is the Lagrange multiplier.
To obtain the optimal solution for the above problem as a
classication function,
(
)
= sgn
(
⋅+
)
= sgn
𝑥
𝑖
⊂SV
𝑖
𝑖
𝑖
⋅+
. ()
(2) e Generalized Optimal Classication Hyperplane. e
optimal hyperplane is discussed that the linear problems can
be divided, in the training sample set linear inseparable case;
some training samples cannot satisfy condition () and then
can join a relaxation factor in conditions
𝑖
≥0,whichis
𝑖
[
(
⋅
)
+
]
≥1−
𝑖
.
()
e objective function is to nd the minimum value of
(1/2)+
∑
𝑛
𝑖=1
𝑖
,whereis the penalty function; a larger
indicates more punishment misclassications.
For the nonlinear problem, make the nonlinear problem
() :
𝑛
→, is a high dimensional inner product space
called the feature space, and () is the feature mapping.
en one constructs the generalized optimal hyperplane in
. One does not need to consider its exact form structure
and only needs to carry on the inner product computation
in the high dimensional space; kernel function (
𝑖
,
𝑖
)can
be introduced. As long as the kernel function is satisfying
the conditions of the inner product, it corresponds to a
transformation space.
e nonlinear decision function is constr u cted in the
input space:
(
)
= sgn
(
⋅
(
)
+
)
= sgn
𝑖
𝑖
𝑖
,
+
.
()
Type (
𝑖
,
𝑗
)=(
𝑖
)⋅(
𝑗
) is called the kernel
function; 0≤
𝑖
≤ ( = 1,2,...,) is the Lagrange
multiplier .
Learning machine that can construct the decision-
making function is called support vector machine. e
structure of support vector machine is given in Figure .
(3) Kernel Function. OneofthefeaturesofSVMistheintro-
duction of kernel function. e low dimensional space vector
setisusuallydiculttoclassify,thesolutionistobemapped
into a high dimensional space, but this approach increases
the complexity of calculation, and the kernel function cleverly
solves the problem. At present, the kernel function is studied
mainly in the following forms:
() e linear kernel function: (,
𝑖
)=⋅
𝑖
.
() Polynomial kernel function: (,
𝑖
)=[(⋅
𝑖
)+1]
𝑑
.
() Gauss radial basis kernel function: (,
𝑖
)=
exp{−|,
𝑖
|
2
/2
2
}.
() Neural network kernel function: (,
𝑖
)=tanh(V(⋅
𝑖
)+).
isthenumberofpolynomials;is width parameter to
RBF function; and V and are constants.
2.2. Experiment Design
2.2.1. Psychological Test. According to the questionnaire and
the method in [, ], investigation and evidence collection
to the driver show the driver’s propensity types preliminar-
ily. e content of psychological questionnaire reects the
psychological characteristics of drivers. options questions
in the table have been given scores in accordance with
incremental numbers referring to the classic psychological
scale. e larger the value, the greater the possibility that
selecting this option on behalf of the driver is conservative.
Scores of – are radical types, – are the common
types, and – are the conservative type. In order to explain
the content of t he questionnaire, but being limited by the
space, the following lists part of the subject and the option
of the psychological questionnaire.
Part of the Subject and the Option
of the Psychological Questionnaire
() Gender
Male (1),Female(2)
() Age
< years old (1),–yearsold(2),–
years old (3), > years old (4)
() Driving years
<years(1), – years (2),–years(3), >
years (4)
() Driving speed will exceed the speed limit or not when
there is no other vehicles interference
Oen (1),Occasionally(2),Never(3)
() e driver oen follows the vehicle ahead or not
Yes (1),Maybe(2),Notusually(3)
() e driver always wants to overtake
Yes (1),Maybe(2),Notusually(3)
Mathematical Problems in Engineering
f(x)
𝛼
1
y
1
𝛼
2
y
2
𝛼
s
y
s
K(x
1
·x)
K(x
2
·x)
···
···
···
x
1
x
2
x
n
Weight: 𝛼
i
y
i
Intermediate layer: nonlinear transformation
K(x
s
·x)
based on support vectors x
1
,x
2
,...,x
s
Input layer: vectors x
1
,x
2
,...,x
n
Output layer: f
(
x
)
= sgn[∑
x
𝑖
⊂SV
𝛼
i
y
i
(x
i
,x)+b]
F : e structure of support vector machine.
() e driver oen makes urgent acceleration or decel-
eration
Yes (1),Maybe(2),Notusually(3)
()Driversmayfeelangrysometimeswhentheyare
overtaken
Yes (1),Maybe(2),Notusually(3)
() e driver will overtake when the headway is narrow
Yes (1),Maybe(2),Notusually(3)
() e mood will be agitated when there is a trac jam
Yes (1),Maybe(2),Notusually(3)
() e driver will speed up to pass the intersection
during signal alternating
Oen (1),Occasionally(2),Never(3)
() e driver will speed up to go through corners
Yes (1),Maybe(2),Notusually(3)
() Drivers will improve the speed imperceptibly when
there are friends in their cars
Yes (1),Maybe(2),Notusually(3)
() Drivers change driving route at the same time that
they make a turn light
Yes (1),Maybe(2),Notusually(3)
e solving process of each subject’s scoring average,
standard deviation, coecient of correla tion is as follows.
Organize that drivers do the psychological questionnaire
in accordance with the requirements, choose the options that
are the closest ones to the feelings, thoughts, and behavior
mostly, getting each subject’s score of each driver in the
psychological test. Using the rst subject as instance, we can
get the scoring average by using the rst subject’s total divided
by headcount; we can get the standard deviation by using the
score average according to the standard deviatio n formula
andthenwiththescoreaverageandstandarddeviationwe
can get the coecient of correlation.
Coecient of internal con sistency (Cronbach’s Alpha
coecient) mainly reects the reliability of the relationship
between the subjects within the test, reviewing whether each
subjectofthetestmeasuresthesamecontentandtrait.
Coecient of internal consistency being greater than .
means that the reliability of the scale is higher.
Shandong Jiaoyun group is commissioned to investigate
through questionnaire in drivers. Take the test scores
as a sample, use SPSS. soware for statistical analysis
to evalua te the reliability (stability of psycho logical mea-
surement tools) and validity (eectiveness of measurement
tool) of the questionnaire [], and in t ernal consistency
coecient is 0.836 > 0.8,whichindicatesthatthescalehas
high reliability homogeneity. Table corresponds to mean,
standard deviation, and the total score Pearson correlation
coecient in each item; about % pr oblems of the score
andtotalscoreinthe.and.levelsaresignicantly
correlated, indicating that the scale has good content validity
and can carry out test.
2.2.2. Vehi cle Experimen t. In the vehic le experiment, the
research perspective is based on the vehicle safety controlled
technology, respecting and protecting the driver privacy as a
precondition. e experiment roads travel time is taken as the
characteristic parameter obtained through GPS. e dynamic
recognition model of driver’s propensi ty is established based
on support vector machine. In this paper the permeability of
%, %, %, %, %, %, and % under the driver’s
propensity dynamic identication is studied, space is limited,
and take the permeability of % as an exam ple.
(1) Experimental Equipment. In city road environment,
dynamic Driver-Vehicle-Environment information acquisi-
tion system (as shown in Figure , including SG-GPS
noncontact multifunction speedometer, BTM--
laser range sensor, HD camera, MiniVcap monitoring system,