Content uploaded by Haotian Cao
Author content
All content in this area was uploaded by Haotian Cao on Aug 15, 2018
Content may be subject to copyright.
1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2865105, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, CYBER-PHYSICAL SYSTEMS IN GREEN TRANSPORTATION 1
Shared Control Driver Assistance System Based on
Driving Intention and Situation Assessment
Mingjun Li, Haotian Cao, Xiaolin Song, Yanjun Huang, Jianqiang Wang, Zhi Huang
Abstract—This paper presents a shared control driver assis-
tance system based on the driving intention identification and
situation assessment to avoid obstacles. A constrained linear-time-
varying model predictive controller (LTV-MPC) is designed to
follow the obstacle-avoidance path which is obtained by artificial
potential method in real time. A human driver’s driving intention
and the desired maneuver are recognized by the inductive multi-
label classification with unlabeled data (iMLCU) approach that is
trained based on the lateral offset and lateral velocity to the road
center line. In addition, the situation assessment of collision risk
is represented by the time to collision (TTC) and the performance
evaluation is designed according to lateral deviation. All of them
are employed for the design of the shared control fuzzy con-
troller. The cooperative coefficient, denoting the control authority
between the controller and a human driver, is determined by
three fuzzy controllers in different conditions, which are the
consistent, the advanced inconsistent, and the lagged inconsistent
fuzzy controller, respectively. More importantly, there are two
scenarios studies are provided to verify the proposed system.
The results prove that the shared control driver assistance system
can successfully help drivers to avoid obstacles and obtains great
vehicle stability performance in different scenarios.
Index Terms—Shared control, driving intention, situation as-
sessment, cooperative coefficient, model predictive control.
I. INT ROD UC TI ON
TRAFFIC safety reports from Nation Highway Traffic
Safety Administration show that about 33,000 injury
crashes have been caused by the drowsy drivers. Over 3,400
people died and another 391,000 people injured due to the
distracted drivers in 2015 [1][2]. However, it is a challenge to
decrease the traffic accidents caused by human errors (e.g.
drowsiness, distraction, fatigue) since these factors usually
happen in a stochastic way and also individual-dependent.
As a result, the advanced driver assistance systems (ADAS)
appear and draw a large amount of attention. Generally,
ADAS are designed to improve human drivers’ performance
especially under the abnormal conditions by executing parts
This work was supported by the National Natural Science Foundation
of China (grant number 51575169), the Joint Laboratory for Internet of
Vehicles, Ministry of Education-China Mobile Communications Corporation
(grant number ICV-KF2018-01), the Natural Science Foundation of Hunan
Province (grant number 2017JJ2032), and China Scholarship Council (CSC)
(grant number 201606130075). (Corresponding author: Xiaolin Song.)
Mingjun Li, Xiaolin Song, Haotian Cao, Zhi Huang are with the State Key
Laboratory of Advanced Design and Manufacturing for Vehicle Body, Hunan
University, Changsha, 410082, China (e-mail: jqysxl@hnu.edu.cn).
Xiaolin Song is also with the Joint Laboratory for Internet of Vehicles,
Ministry of Education - China Mobile Communications Corporation.
Yanjun Huang is with the Department of Mechanical and Mechatronics
Engineering, University of Waterloo, Waterloo, Ontario N2L3G1, Canada.
Jianqiang Wang is with the State Key Laboratory of Automotive Safety and
Energy, Tsinghua University, Beijing 100084, China.
of the driving task (e.g. adaptive cruise control), or warning
(e.g. a forward collision warning) a human driver before the
hazardous situation happens in recent years [3]-[6]. However,
it is not a feasible solution if a human driver is driving
inattentively and won’t take any actions to the warnings. In
addition, the fully automated driving system can be treated as
a potential solution because the human driver will not be a part
of the driving process. However, it is still under the research
phase due to a wide range of limitations [7][8]. Therefore, as
an alternative solution to the above problem, the shared control
allows both a human driver and an assistant controller to drive
the vehicle simultaneously, which takes the advantages of both
a human driver and the automatic system.
In terms of the shared control, earliest works focus on the
switch control. The measured attentiveness or driving states of
a human driver are considered into the switch control authority
between a human driver and an automated controller [9]-
[11]. However, the design guidelines of the shared control
system commonly agree that there are the dynamic control
authority allocation and interaction between a human driver
and an automation controller [12]. The haptic shared control
and the steer-by-wire system have received increasing attention
as these systems provide a possibility that a human driver and
the automation controller simultaneously control the vehicle
[12]-[14]. Besides, researchers in [14] [15] prove that the
haptic shared control and the steer-by-wire system improve
task and safety performance. Mars et al. propose that the
best cooperation of the shared control systems is realized by
relatively low-level haptic authority [16]. In reality, when and
how to intervene is an open issue in the design of shared
control system.
The cooperative shared control systems based on the game
theory are proposed by Na et al. and Flad et al., which are
effective to address path conflicts between a human driver and
the controller. Na et al. study the interaction between a human
driver and the active front steering (AFS) system via game
theory [17]. The linear quadratic (LQ) dynamic optimization
and disturbed MPC (DMPC) method are used to find an
equilibrium between two players in game theory in simulation
studies [18]. Flad et al. propose the moveme driver model that
allows the real-time experiment with a human driver using the
differential game theory for the cooperative ADAS [19][20].
However, due to the changeable environment, it is not suitable
to use the constant parameters during the whole experiment in
the current game theory researches. Besides, there exist several
methods to design the shared control according to situation
assessment. The control authority of the controller is increased
with the situation risk rising. Mars et al. argue that the human
1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2865105, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, CYBER-PHYSICAL SYSTEMS IN GREEN TRANSPORTATION 2
Fig. 1. The illustration of the shared control driving assistance system
framework.
driver deserves more intervention in poor traffic condition [16].
Anderson et al. [21] design the risk assessor module based on
the predicted front slip angle, predicted steering angle and its
changing ratio. Samyeul et al. [22] adopt three parameters to
assess the collision risk around the vehicle, time to collision
(TTC), time to brake (TTB), and the minimal safety margin
(MSM), respectively. Simultaneously, the driving mode and its
control authority are determined.
Another shared control approach involves driver’s driving
intention. The shared control systems are required to under-
stand a human driver, such as H-metaphor, and not only the
systems that shift the authority between the human driver
and the controller [12] [15]. Gao et al. [23] assume that
the planned path represents the driver’s intention and apply
it into the shared control for path planning and following. A
driver-vehicle-road (DVR) model is used to model the driver’s
behavior and intention when the shared control steering is
designed by Saleh et al. [24]. Different from Gao et al. and
Saleh et al., Erlien et al. and Song et al. directly add the
human driver’s steering command into the cost function of
MPC formulation. Erlien et al. [14] consider the front lateral
force and Song et al. [25] adopt the steering wheel angle
of a human driver. Then, the final shared input is calculated
by MPC method with the constraints of the environment and
vehicle dynamics.
Although some shared control methods have been proposed,
there are still much room to improve. In aforementioned
methods, the shared control strategy is designed based on
either the situation assessment [21] [22] or the human driver’s
factors, such as driver’s attentiveness [9]-[11], driving in-
tention [23]-[25]. However, it is reasonable to consider the
factor of a human driver, vehicle and environment simulta-
neously when the shared control strategy is applied into the
closed-loop DVR model. Therefore, the driving intention, the
situation assessment and the performance evaluation are all
taken into consideration in the proposed strategy. Besides,
differing from building a driver’s model or incorporating the
steering command into the MPC formulation to consider the
driving intention, a semi-supervised machine learning method
is adopted in this paper to predict the desired maneuvers and
the human driver’s driving intention. Moreover, inspired by
[17], the situation assessment is taken into consideration when
the desired maneuvers have conflicts with the human driver’s
driving intention.
The framework of the proposed shared control driver as-
sistance system is illustrated in Fig. 1. Assume that the
controller is well-studied in the typical scenarios so that the
controller is able to deal with emergency conditions. The
path following controller is designed by a constrained LTV-
MPC algorithm. [26][27] We also assume that the information
of environment and vehicle are provided. The road lanes
detected by front camera [28], and the localization and states
of obstacles and the host vehicle obtained by using GPS
and other vehicle sensors (e.g. radar) [29] are needed. The
obstacle avoidance path is obtained by an artificial potential
method according to the measured environment information
in real time [32]. This paper desires to deal with the situation
where a human driver controls the vehicle but with poor
driving states. Therefore, the driver’s driving intention and
the desired maneuvers are recognized through the iMLCU
algorithm [30]. An eight-degree-of-freedom (8 DOF) nonlinear
vehicle model is required to predict the vehicle path driven by
the human driver alone. And the predicted path is inputted into
iMLCU classifier to recognize the human driver’s intention.
After obtaining the recognized results, the inconsistence area
is marked by comparing the desired maneuver with the human
drivers’ intention, which is employed for the development of
the shared fuzzy controller. The situation assessment result
and the performance evaluation index are selected as input
variables and the cooperative coefficient is the output of the
shared fuzzy controllers under different intention comparison
conditions.
This paper is organized as follows. Section II describes the
path following controller based on a constrained linear-time-
varying model predictive control method. The driving intention
recognition module using iMLCU approach and the details
of driving intention collecting experiment are introduced in
Section III. The shared fuzzy controller, integrating the situ-
ation assessment, the performance evaluation are proposed in
Section IV. The simulation studies are given in Section V. The
final section presents the conclusion and future works.
II. PATH-FOLLO WI NG CO NT ROL LE R
An essential part of the shared steering system is a well-
designed path-following controller.[31] In this section, a con-
strained LTV-MPC method is introduced. [32][33]
A. Vehicle Dynamics
As the coordinate system shown in Fig. 2, a 5 DOF single-
track rigid body model that includes the vehicle longitudinal
speed u, vehicle sideslip angle β, yaw motion ϕand front/rear
wheel rolling motion ωf,ωrwill be considered for prediction
in LTV-MPC algorithm. According to the Newtown second
laws, the motions of the vehicle are governed by following
nonlinear differential equations [34]
1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2865105, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, CYBER-PHYSICAL SYSTEMS IN GREEN TRANSPORTATION 3
Fig. 2. The illustration of a five-DOF single-track vehicle model.
CoG: center of gravity.
˙u=1
m(Fxf +Fxr) + vr
˙
β=1
mu (Fyf +Fyr )−r
˙ϕ=r
˙r=1
Iz(laFyf −lbFyr )
˙ωf=1
Jw(Twf −Fxf r)
˙ωr=1
Jw(Twr −Fxr r)
(1)
where the vehicle has a lumped mass m, and moment of
inertia Izcorresponding to Z-axis of the vehicle. Fyf , and Fyr
represent the lateral tire force of front wheels and rear wheels,
respectively, and Fxf , and Fxr represent the longitudinal tire
force of front wheels and rear wheels, respectively. The length
from the front and rear axle to the vehicle CoG is denoted by
laand lb.Jwmeans the equivalent inertia for front and rear
tires. Twj denotes the torque acting on the wheel, which could
be a driving torque when in accelerating or a braking torque
when it brakes, for a rear-wheel drive vehicle, it could be
evaluated by,
Twf =τ Tw
0Twr =(1 −τ)Twif braking
Twif accelerating (2)
where τrefers to the braking torque distribution coefficient.
With the state variables x= [u, β, ϕ, r, ωf, ωr]Tand input
variables u= [δsw Tw]T, where δsw is the steering wheel
angle, and the dynamic system (1) could be expressed as
˙x=fv(x,u)(3)
It is well known that the main cause of this non-linearity of
the vehicle dynamical model comes from the limited available
tire force. One common way of modeling the tire force versus
the slip angle is to apply the Magic Formula under pure slip
conditions. However, the vehicle might deal with a situation
where longitudinal speed changes, thus the combined slip
should be considered, and the tire force could be evaluated
as following [35]
Fxwj (κi, αi) = σxj
σj
Fxj σj
1− |σj|(4)
Fywj (κi, αi) = σy j
σj
Fyj arctan σj
1− |σj| (5)
where Fxj , and Fyj ,j∈ {f, r}represent the front/rear
longitudinal and lateral Magic Formula tire model under pure
slip condition. κand αrefer to the longitudinal and lateral
slip, respectively. σdenotes the combined slip is defined by
σ=qσ2
x+σ2
y(6)
with σx=κ
1+|κ|,σy=tanα
1+|κ|. In addition, the tire slip angle
for front/rear wheels are approximated as follows
αf=δsw
G−β−lar
u, αr=−β+lbr
u(7)
where Gis the ratio of the steering system. We also define
the longitudinal slip as
κ=rω−u
uif rω ≤u
rω−u
rω otherwise (8)
In order to further ensure the stability performance of the
vehicle, we also add the yaw rate and rear tire slip angle into
the outputs of the system [24], since
|r| ≤µg
u(9)
|αr|=
lbr
u−β≤αr,sat (10)
where µis the road adhesion constant, gis the gravitational
constant, and αr,sat limits the rear slip angle at which lateral
tire force saturates. Therefore, the output gvof the system is
evaluated by
gv=u y |r| − µg
u|lbr
u−β|T(11)
Finally, in order to deploy a LTV-MPC, we have to reform the
dynamical system (1) as a discrete form with sampling time Ts
first. With an arbitrary working point (x0,u0), the nonlinear
vehicle dynamic model could be linearized as a linear time
variant state space form by an approximation of its first order
of its Taylor expansion, such that
x(k+ 1) = fv(x0(k),u0(k)) + An(x(k)−x0(k))
+Bn(u(k)−u0(k))
z(k) = gv(x0(k),u0(k)) + Cn(x(k)−x0(k))
(12)
where zmeans the output of the linearized system, An,Bn,
Cnare the Jacobian matrices evaluated at the working point.
B. LTV-MPC Method
Recalling the objective of the trajectory following is to
track the planned path precisely, meanwhile, keep the vehicle
moving with a desired speed as well. That means, for a given
reference signal Yref , the objective of the predictive control
system is to ensure the predicted outputs as precisely close to
the reference signal within the prediction horizon. Then the
cost quadratic function is listed as following
J= (Yref −W)TΠ(Yref −W) + UTΩU(13)
where Πis a diagonal subcomponent weight matrix for the
errors between the reference signals and predicted outputs. Ω
is the weight matrix component for the predicted control input
U.Yref is the reference signal including the vehicle’s lateral
position and its speed, which are obtained from path-planing
[32] and velocity planning [40], respectively. Wis the system
1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2865105, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, CYBER-PHYSICAL SYSTEMS IN GREEN TRANSPORTATION 4
Fig. 3. The path-following results of LTV-MPC algorithm: (a) global path,
(b) tracking error, (c) steering wheel angle and its increment, and (d) lateral
acceleration.
predicted outputs with the lateral position and longitudinal
speed.
Moreover, the input and the incremental constraints are
Umin ≤U≤Umax (14)
∆Umin ≤∆U≤∆Umax (15)
Additionally, the stability performance for each time step could
be represented by
HCnx(k)≤Q(16)
with
H=0010
0001G=0
αr,sat
Therefore, the path tracking could be solved by a convex
optimization problem expressed as follows
min
UJ(17a)
s.t.
stability constraints
HCnx(k)≤Q(17b)
input constraints
Umin ≤U≤Umax (17c)
∆Umin ≤∆U≤∆Umax (17d)
The parameters of the LTV-MPC path-following controller
are listed in Table II where all the parameters for our study are
presented. The path following performance of the LTV-MPC
controller is shown in Fig. 3. The host vehicle is controlled
by the LTV-MPC controller with the guidance of the reference
path at a constant speed of 20m/s. As observed from Fig.
3(a) and (b), the followed path of the LTV-MPC controller
is closed to the reference path, and the maximum tracking
error is lower than 0.2m. Besides, as depicted in Fig. 3(c), the
steering wheel angle and its incremental curve are smooth.
Both the maximums of two parameters are far less than its
constraint values which are listed in Table II. Moreover, the
LTV-MPC controller could achieve great stability performance
since the absolute maximum value of lateral acceleration is
lower than 0.4g. Therefore, the LTV-MPC controller is able
to be used for the path following and thus it will be utilized
for the development of the shared control.
III. INT EN TI ON RE CO GN IT IO N
Given the desired path to avoid obstacles, the human
driver tends to make wrong decisions because of the poor
driving states, such as distracted, fatigue driving. In order to
prevent the potential traffic accident, it is necessary to predict
the human driver’s lane-changing intention and the desired
obstacle avoidance maneuver. Therefore, a brief introduction
of iMLCU approach and verification of the proposed method
are presented in this section.
A. iMLCU Approach
iMLCU approach is proposed to address the semi-
supervised multiple-label learning problem [30]. For the semi-
supervised learning, the training dataset Scomprises the
labeled training dataset Sland the unlabeled training dataset
Su. Given a set of labeled training dataset Sl={(xi, Yi)},
xi∈Rdare the d-dimensional training data point, Yi=
{y1, . . . , yl}are the label space of xiwith qpossible classes
and yj∈ {−1,1},1≤j≤q. Supposing that there are n
labeled training instances and munlabeled training instances,
therefore, the training dataset Sis denoted as
S={(x1, Y1), . . . , (xn, Yn),xn+1, . . . , xn+m}(18)
The goal of iMLCU classifier is to find a family of qreal-
value functions fi(x, yi)to predict a correct label for the new
test data St, and fi(x, yi)are given by
fi(x, yi) = hwi,xi+bi,1≤i≤q(19)
where hwi,xiis the inner product of the weight vector wi
and the training data x, and biis the bias for class label yi.
Simultaneously, in order to address the multiple-label problem,
the method to predict the label of test data is adopted as follow
ˆ
Y=sign(f1(x, y1), . . . , fq(x, yq)) (20)
The iMLCU classifier is trained by using the information of
labeled data and unlabeled data. Therefore, the cost function
of iMLCU classifier comprises two parts, namely, the labeled
part and the unlabeled part. The details of the cost function
are described as follows.
1) Labeled Part: The hyperplanes for different classes
labeled instances (xi, Yi)are defined as
hwk−wn,xii+bk−bn= 0 (21)
where (yk, yn)∈Yiׯ
Yi. High classification accuracy is
achieved by finding the hyperplane that has the maximum
margin to the nearest data of any training classes. Therefore,
the labeled data are used by optimizing the following equation
min
w,Ξ
q
X
k=1
kwkk2+C
n
X
i=1
1
|Yi|| ¯
Yi|X
(yk,yn)∈Yiׯ
Yi
ξikn
s.t.hwk−wn,xii+bk−bn≥1−ξikn
ξikn ≥0,(1 ≤i≤n, (yk, yn)∈Yiׯ
Yi)
(22)
where Ξ = {ξikn|1≤i≤n, (yk, yn)∈Yiׯ
Yi}relate to the
slack variables ξikn, and Crefers to the penalty parameter.
1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2865105, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, CYBER-PHYSICAL SYSTEMS IN GREEN TRANSPORTATION 5
2) Unlabeled Part: The unlabeled instances should be
placed outside the margin and penalized when it lies within
the margin or even on the wrong side of the hyperplane in
the learning process. However, without knowing the class
label of an unlabeled instance, it can’t be determined whether
the unlabeled instance is on the right or wrong side of the
hyperplane. Therefore, the iMLCU classifier adopts the same
idea of S3VM to the unlabeled instances [36]. The prediction
label obtained from (20) is treated as the putative label sets
of an unlabeled instance x. Besides, the loss on label yiis
penalized by using the hinge loss function on x, denoted as
`i(x,ˆyi, fi(x, yi)) = max(1 − |hwi,xi+bi|,0) (23)
For sake of better classification accuracy, the total losses on
unlabeled instances should be minimized, given by
min
w
n+m
X
j=n+1
q
X
v=1
max(1 − |hwv,xji+bv|,0) (24)
However, (24) is non-convex objective function because it
consists of the sum of qnon-convex functions `ion every
unlabeled instance. In order to solve the non-convex optimiza-
tion problem presented in this paper, the ConCave Convex
Procedure (CCCP) method [37] is adopted. Therefore, (24)
should be decomposed into a convex part and a concave part.
If an unlabeled instance xjis predicted as a positive label yv
at current step, the effective loss function at next iteration is
rewritten as follow
e
L(t) =
0t≥1
1−t|t|<1
−2t t ≤ −1
(25)
with t=hwv,xji+bv. And for the case of the negative
predicted label yv, the effective loss function is defined as
e
L(t) =
2t t ≥1
1 + t|t|<1
0t≤ −1
(26)
Therefore, the objective function of iMLCU classifier for
the semi-supervised multi-label learning problem is defined as
min
w,Ξ
q
X
k=1
kwkk2+C1
n
X
i=1
1
|Yi|| ¯
Yi|X
(yk,yn)∈Yiׯ
Yi
ξikn
| {z }
Labeled part
+C2
n+m
X
j=n+1
q
X
v=1 e
L(hwv,xji+bv)
| {z }
Unlabeled part
s.t. hwk−wn,xii+bk−bn≥1−ξikn
ξikn ≥0,(1 ≤i≤n, (yk, yn)∈Yiׯ
Yi)
1
m
n+m
X
j=n+1
hwv,xji+bv=1
n
n
X
i=1
yiv
(27)
where C1and C2are nonnegative weighting constants between
the labeled part and the unlabeled part, and the value of these
two parameters are referred from [30].
Fig. 4. Driver-in-the-loop simulation platform.
B. Equipment and Scenario
All of the training and testing data were collected from
the driver-in-the-loop simulation platform (shown in Fig. 4).
Logitech G27 was used to collect the human driver’s steering
wheel angle, throttle opening, and braking pedal displacement
in the CarSim-LabVIEW real-time environment. Besides, the
driving scenario including road shape and obstacle vehicles is
designed by CarSim software (see Fig. 5) to collect the training
and testing data of lane-changing intention. As depicted in Fig.
5, the driving scenario consists of four segments, namely, two
straight road segment, a curvy road with positive curvature
segment and a curvy road with negative curvature segment.
Besides, there are four obstacles with different initial states,
and the initial lateral positions of them are at the middle point
of the right lane in this scenario. The details of each obstacle
vehicle are described as below
a) O1is a static obstacle with the initial station at 120m.
b) O2is a dynamic obstacle running alone center line of
right lane with the initial station at 200m and its initial
velocity at 10m/s.
c) O3is a dynamic obstacle running alone center line of
right lane with the initial station at 650m and its initial
velocity at 5m/s.
d) O4is a static obstacle with the initial station at 1350m.
C. Training and Testing of iMLCU
1) Task and Procedure: The primary task for each par-
ticipant is to avoid obstacles by lane changing maneuvers,
and in order to collect the lane-changing data efficiently, each
participant should turn back to the original lane after exceeding
ahead obstacle. During the data collecting procedure, each
participant should follow the rules listed as below [36]
a) All participants were in normal physical and mental
states. Fatigue or drowsy driving states were not allowed.
Besides, using smart phone, answering a call or talking
to others also were not allowed.
b) Each participant was allowed to be familiar with the
driver-in-the-loop simulation platform and the driving
scenario by testing driving twice.
c) The velocity should be lower than 120km/h.
1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2865105, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, CYBER-PHYSICAL SYSTEMS IN GREEN TRANSPORTATION 6
Fig. 5. Illustration of the designed scenario for collecting lane-changing
driving intention data.
2) Data Collecting and Processing: The driving intention
is divided into three categories in our study, namely, the left
lane-changing (LLC), the right lane-changing (RLC), and the
lane keeping (LK). To investigate the lane-changing intention,
two feature parameters in this study are selected, the relative
offset yrc and the relative lateral offset velocity vyrc to the
road center line, respectively. These two features reflect the
lane-changing velocity and the relative position of the host
vehicle to the center line. Therefore, the vehicle global lateral
position yand the information of designed road center line rc
are collected for obtaining yrc and vyrc . The expressions are
denoted as
yrc =y−rc, vyrc = ˙y−˙rc(28)
Therefore, the training data point xis given by (yrc, vyrc ).
20 licensed drivers in total participated in our study. Their
average age was 23.75 years with a standard deviation of 2.17
years, and their average driving experience was 18.6 months
with a standard deviation of 12.6 months. Each participant
was required to finish 10 complete experiments, 200 datasets
in total. However, the length of lane keeping segments is
highly larger than the length of lane changing segments in
the designed scenario. In order to keep the training data of
three categories balanced, 500 data points are filtered around
lane-changing area. There are 8 lane-changing areas in total
after a participant finished the whole procedure. Therefore, 200
datasets and 4,000 data points of each dataset were obtained
in total.
In order to validate iMLCU algorithm proposed, all fil-
tered datasets are allocated into three disjoint parts: the la-
beled dataset Sl={xi}n
i=1, the unlabeled dataset Su=
{xi}m
i=1, and the test dataset St={xi}k
i=1, where n=45,000,
m=715,000, and k=40,000.
Fig. 6. Examples of three categories driving intention.
Fig. 7. The testing accuracy of iMLCU classifiers trained with different label
ratios.
3) Data Labeling: As illustrated in Fig. 6, some of data can
easily be labeled into LLC, LK or RLC. The data point with
larger positive relative lateral velocity (e.g. vyrc ≥1.8m/s)
could be labeled as LLC, and the data point with larger
negative relative lateral velocity (e.g. vyrc ≤ −1.8m/s) could
be labeled as RLC. Besides, if the value of vyrc nears zero and
the value of yrc nears the middle point of right or left lane,
the data point could be labeled as LK. However, the vague
data points shown in Fig. 6 are difficult to label. The semi-
supervised learning algorithm, iMLCU approach in our study
is able to determine the decision hyperplane among multiple
classes by utilizing the labeled and unlabeled data. Therefore,
a few data points that have distinct characteristic of three
lane-changing intention categories described above need to be
labeled.
4) Training and Testing Result: 15,000 labeled data point of
three categories, 45,000 in total, and 715,000 group unlabeled
data are provided for iMLCU training. However, the label ratio
between the labeled and the unlabeled data has a influence on
iMLCU classifier accuracy.[30] The iMLCU classifiers have
been trained with the label ratio of 0.01, 0.02, 0.03, 0.04,
and 0.05 to find the suitable label ratio. As for the training
datasets with different label ratios, all 715,000 unlabeled data
are used and fixed, and the training datasets with different
label ratios are constructed by tuning the number of the labeled
data. In order to make the labeled data of three lane-changing
intention categories balanced, the labeled data of equal size are
randomly picked from three categories for training. Besides,
in order to test the iMLCU classifier, 13,000 group data of
LLC, 14,000 group data of LK, and 13,000 group data of
1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2865105, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, CYBER-PHYSICAL SYSTEMS IN GREEN TRANSPORTATION 7
Fig. 8. The predicted path and recognized intention via a Carsim model:(a)
predicted path, and (b) the recognized intention; rl,rbrepresent the road
upper and bottom border respectively, and rcrefers to the road center line.
RLC are labeled according to data labeling principle. The
average accuracy of testing results are shown in Fig. 7. The
average accuracy increases with the label ratio increasing, and
the average accuracy reaches 93.18%when the label ratio is
0.05. But the training time-consuming also increases with the
label ratio increasing [30]. Therefore, the iMLCU classifier
trained with 0.05 label ratio is adopted to develop the shared
control strategy.
D. Driver’s Driving Intention and Desired Maneuver
1) Driver’s Driving Intention: As the host vehicle is con-
trolled by the shared control driver assistance system, the
final control signal inputted to the vehicle is the blend of
driver’s and the path-following controller’s. The recognized
intention based on yrc and vyrc of the host vehicle measured by
various sensors can not reflect real driver’s driving intention.
Therefore, in order to recognized the driver’s lane-changing
intention, the accurate prediction for vehicle global lateral
position driven by a human driver alone is necessary. Driver’s
steering wheel angle can be obtained by sensors, thus, an
accurate vehicle model is needed to predict the vehicle lateral
position. However, the 5 DOF single-track vehicle model
presented in (1) has been simplified for the design of the LTV-
MPC controller. Besides, an eight-degree-of-freedom (8 DOF)
four wheels vehicle model could provide a great accuracy
level.[38] Therefore, an 8 DOF vehicle model [39] [40] is
adopted in our study.
The driver’s steering wheel angle is collected from the
driver-in-the-loop platform, and then the driver’s steering
wheel angle is inputted into an 8 DOF nonlinear vehicle
model. yrc and vyrc are calculated by (28). Fig. 8 depicts
the predicted path calculated by the 8 DOF nonlinear vehicle
model. The mean error between the path calculated by an
8 DOF nonlinear vehicle model and the measured path is
0.1437m, and the value of RMSE is 0.1208m. Observed from
the Fig. 8(b), the recognized lane-changing intention based on
the data calculated by the an 8 DOF nonlinear vehicle model
has little difference with the measured data. Therefore, it is
effective to predict the driver’s driving intention by an 8 DOF
nonlinear vehicle model in this paper.
2) Desired Maneuver: Given by the information of the road
center line and the reference path, the relative lateral offset to
the road center line is denoted as
yos(k) = P(yr(k)−rc(k)) (29)
where
yr(k) =
yref (k)
yref (k+ 1)
.
.
.
yref (k+Np)
rc(k) =
rc(k)
rc(k+ 1)
.
.
.
rc(k+Np)
P=1 0 · · · 0is the matrix to take the first element,
yr(k)and rc(k)mean the reference preview path and the road
center line over the prediction horizon Nprespectively.
Thus, the relative lateral velocity to the road center line is
expressed as follow
vyos (k) = 1
Ts
(yos(k)−yos (k−1)) (30)
Then, the sample data [yos(k)vyos (k)]Tare inputed into the
iMLCU classifier to predict the desired maneuver.
IV. SHA RE D CON TROL STRATE GY
In this section, in order to follow the planned obstacle
avoidance path by the shared control driver assistance system,
the shared control strategy is investigated based on the fuzzy
control method by considering the driving intention, situation
assessment and performance evaluation. The control authority
between a human driver and the path-following controller is
also discussed in different conditions.
A. Situation Assessment
Situation assessment is a crucial part of the shared control
strategy [22][41]. The design of the strategy is based on
the safety level of current situation. The higher dangerous
situation is, the more proportion of the control authority will be
occupied by the path-following controller. Takashi summarized
that a human driver would execute lane-changing maneuver to
avoid front obstacles before TTC was lower than 6s, and 2s
is the safety margin[42]. Therefore, the shorter distance for
a human driver to execute lane-changing maneuver to avoid
obstacle remains, the higher dangerous would be. In our study,
time to collision (TTC), defined as the time for two vehicles
to collide if they continue at their current velocity and same
trajectory when a human driver keeps current lane, is selected
as the situation assessment factor. The formula of TTC is
denoted as
ttc =(xe
k
vh
k−vo
k
,xe
k
vh
k−vo
k
>0
+∞,otherwise (31)
where xe
kis the relative distance from the nearest obstacle to
the host vehicle at time k,vh
kand vo
krepresent the host vehicle
velocity and obstacle vehicle velocity at time k, respectively.
Fig. 9. The illustration of two inconsistent conditions: (a) illustration of
scenarios, and (b) the recognized intention result; xdr and xde denote a
driver’s and the desired lane-changing longitudinal position respectively.
1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2865105, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, CYBER-PHYSICAL SYSTEMS IN GREEN TRANSPORTATION 8
Fig. 10. The illustration of the shared fuzzy controller: (a) membership function of ttc, (b) membership function of the performance evaluation index η, (c)
membership function of the cooperative coefficient in the consistent intention condition kcon, (d) membership function of the cooperative coefficient in the
advanced inconsistent intention condition kad, (e) membership function of the cooperative coefficient in the lagged inconsistent intention condition kla, (f)
fuzzy surface of the consistent fuzzy controller, (g) fuzzy surface of the advanced inconsistent fuzzy controller, and (h) fuzzy surface of the lagged inconsistent
fuzzy controller.
B. Performance Evaluation
In the shared control driver assistance system, the objective
of a human driver and the path-following controller is to follow
the planned path as closely as possible. The lateral deviation
to the planned path is selected to evaluate the performance for
each participant. Obviously, the larger lateral deviation caused
by a human driver alone is, the higher proportion assistance
the human driver needs. Besides, the bigger lateral deviation
of the shared system is, the higher control proportion the path-
following controller occupies to correct the path error. Thus,
the lower lateral deviations of the human driver and the shared
control system are, the better the performance is and the lower
the performance evaluation index will be. The performance
evaluation of shared strategy is designed as follow
η=ε
yl, ε ≤yl
1, otherwise (32)
with
ε=|∆y+λ∆yp|
2(33)
where yldenotes a upper threshold of lateral deviation that
can be tuned according to different environment, ∆ymeans
the lateral deviation between the measured vehicle path to the
planned path, ∆yprefers to the lateral deviation between the
driver’s predicted path calculated by the 8 DOF nonlinear
vehicle model and the planned path, and λis a constant
compensation coefficient for the prediction error.
C. The Shared Fuzzy Controller
The form of the shared control strategy in our work is
inspired by Anderson et al. [21]. The final steering command
is a blend of a human driver’s input and the controller’s input,
denoted as
Us=kUc+ (1 −k)Udr , k ∈[0,1] (34)
where Us,Uc, and Udr refer to the shared steering input and
the controller’s input and a human driver’s input, respectively;
kmeans the cooperative coefficient determined by threat
assessment in [21].
However, the difference and the novelty in our study is that
the shared control strategy is designed based on the driving
intention, situation assessment, and performance evaluation
index. The fuzzy control method is chosen as the control
process could be more smooth and effective [43], and the
safety level from situation assessment and performance eval-
uation index are more suitable to describe by fuzzy language.
The basic idea of the fuzzy control strategy is that the
assistance proportion of a human driver is determined by
the safety level, the performance evaluation index, and the
consistency between the desired maneuver intention and the
human driver’s lane-changing intention. When comparing the
intention consistency between two participants, there are two
inconsistent conditions illustrated in Fig. 9. The first condition
is that a human driver takes a lane-changing maneuver before
the desired lane-changing maneuver position xde, where xde
is obtained by detecting the rising or falling edge of desired
maneuver signal, and the inconsistent area is marked as 1
.
The advanced lane-changing maneuver makes the host vehicle
more safely than the desired maneuver. Thus, the controller
intervention proportion should be weakened during the first
area. In contrast to the first condition, the second condition
is that a human driver turns left to avoid the obstacle vehicle
behind the desired lane-changing maneuver position xde. The
area marked as 2
appears more dangerous, because it is closer
to the obstacle vehicle. The controller intervention should be
1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2865105, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, CYBER-PHYSICAL SYSTEMS IN GREEN TRANSPORTATION 9
TABLE I
THE SHARED FU ZZY CO NT ROLL ER RU LE BAS E
.
ttc Performance evaluation index η
S M MH H
HD M MH H H
D M M MH H
S S M MH MH
strengthened. Therefore, in order to meet the requirements
of the shared control strategy described above, three fuzzy
controllers are designed based on the situation assessment
and the performance evaluation index, namely, the consistent
fuzzy controller, the advanced inconsistent fuzzy controller,
and the lagged inconsistent fuzzy controller. The cooperative
coefficient kis calculated by three fuzzy controllers under
different intention comparison conditions in this strategy. Be-
sides, three fuzzy controllers have same fuzzy input variables
and its membership function, and fuzzy rules. The output
membership functions are different, as shown in Fig. 10. The
details of the shared fuzzy controller are given below:
1) Fuzzy Variables: The shared fuzzy controller obtains
two input variables which are ttc and ηcalculated from (31)
and (32), respectively, and three output variables, named kcon,
kad and kla. Besides, kcon,kad and kla are the cooperative
coefficient of the consistent, the advanced inconsistent, and
the lagged inconsistent fuzzy controller, respectively.
ttc has three associated linguistic values: high dangerous
(HD), dangerous (D), and safe (S). Takashi suggests that 2s is
the safety margin [42], and Janssen et al. [44] and Wang et al.
[41] recommend that 4s is a warning time criterion. Thus, we
define ttch as the safety threshold that is set to 4s and ttcl is
defined as the dangerous threshold that is set to 2s. Therefore,
its membership function shape is shown in Fig. 10(a).
ηis greater or equal to zero, therefore, it has four positive
associated linguistic values: small (S), medium (M), medium
high (MH), and high (H). Its membership function shape is
shown in Fig. 10(b).
Finally, kcon,kad and kla have four associated positive
linguistic values: S, M, MH, H. The value of kla should
be bigger than the value of kcon, and the value of kcon
should be bigger than the value of kad, which are calculated
by same input variables. In order to tune the cooperative
coefficient values in different intention consistency conditions,
the membership function shape of kcon,kad and kla are
shown in Fig. 10(c)- (e).
2) Fuzzy Rules and Inference: The fuzzy rules is the nu-
cleus part of the fuzzy controller and these rules are produced
based on the expert knowledge. In our study, the fuzzy rules
is generated from the simulation experience, and followed
the principle that the more dangerous obtained from situation
assessment is and the bigger the performance evaluation index
ηis, the higher cooperative coefficient kwould be. Therefore,
all the rules are listed in Table I, and the fuzzy surfaces are
shown in Fig. 10(f)- (h).
Besides, the Mamdani inference method [45] is adopted to
solve fuzzy implication in this paper, and the inference form
is denoted as
IF ttc is Aand ηis BTHEN kis C(35)
where A,B, and Care the fuzzy values defined in inputs and
outputs fuzzification.
3) Defuzzification: Defuzzification is the process to trans-
form the fuzzy output values generated by an inference method
to the output values that could be inputted into the system. The
centre of area (CoA) defuzzification method [39] is used in
this paper.
V. CASE ST UD IE S
In order to study the shared control driver assistance system,
two cases are provided with a low traffic in this section. Espe-
cially, a driver with distracted driving state [46] is studied in
the first case and its aim is to discuss the shared control system
performance of dealing with poor driving states. Besides, in
order to focus on the shared steering control, we assume that
the longitudinal speed of the host vehicle is constant. There
are listed as follows:
a) There are three static obstacles on a straight road , and
a human driver will receive a phone call after exceeding
the second obstacle.
b) A human driver drives with a normal state on a curvy
road, where there are a static obstacle and a dynamic
obstacle.
Besides, the lateral deviation and lateral acceleration are
chosen as the performance index to evaluate the shared con-
trol strategy in this paper. Moreover, in order to show the
advantages of the shared control driver assistance system by
comparing with the performances of the vehicle driven by a
human driver alone, the steering inputs of a human driver
are recorded from the driver-in-the-loop platform during case
studies, then the data are imported to the 8 DOF nonlinear
TABLE II
PARA MET ER S FOR SI MU LATI ON STU DIE S
Symbol Description Value(Units)
mThe mass of the host vehicle 1800( kg )
laDistance from CoG to the front axle 1.402(m)
lbDistance from CoG to the rear axle 1.650(m)
IzVehicle yaw moment of inertia 3234(kg·m2)
JwWheel inertial 2.9(-)
GRatio of the steering system 18(-)
gGravitational constant 9.8(m/s2)
rWheel effective radius 0.36(m)
αr,sat
Rear slip angle at which lateral 7.5(deg)
force saturates
Umin/Umax Constraints of the steering input -150/150(deg)
∆Umin/∆Umax Constraints of steering increment -15/15(deg)
Np/NuPrediction/Control horizon length 100/100(-)
TsLTV-MPC sample time 0.02(s)
C1/C2iMLCU classifier parameters 20/0.01(-)
ylUpper threshold of lateral deviation 1.2(m)
λCompensation coefficient of 0.95(-)
predicted path
1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2865105, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, CYBER-PHYSICAL SYSTEMS IN GREEN TRANSPORTATION 10
Fig. 11. Scenario I: two static obstacles on a straight road; (a) the illustration
of scenario I and the vehicle paths of the shared system and a driver, (b)
the driving intention recognized results of a human driver and the desired
maneuvers, (c) the steering wheel angle, (d) the cooperative coefficient, and (e)
the lateral acceleration. ttch1,ttch2, and ttch3represent the safety threshold
of the obstacle O1,O2and O3respectively; ttcl1,ttcl2, and ttcl3represent
the dangerous threshold of the obstacle O1,O2and O3respectively.
vehicle model to obtain the performance parameters of the
vehicle driven by the human driver alone (e.g. the vehicle path
and lateral acceleration). All the parameters in the simulation
are fixed, and listed in Table II.
A. Scenario I
As depicted in Fig. 11(a), the host vehicle moves with
the constant speed as 20m/s, and the first and the last static
obstacle, O1and O3, are at the middle point of right-hand lane.
The initial longitudinal position of two obstacles are 120m
and 300m far from the initial position of the host vehicle.
Besides, the second static obstacle O2is at the middle point of
the left-hand lane and its initial longitudinal position is 210m
far from the initial position of the host vehicle. The obstacle
avoidance path is planned according to the information of the
road map and the obstacles’ position in real time. Both a naive
young driver and the shared control assistance driver system
follow the planned path. As we can observe, the mean lateral
deviation to the planned path of a human driver is 0.7281m,
and the value of RMSE is 0.5927m, while the mean lateral
deviation of the shared control system is 0.0389m and its
RMSE value is 0.0236m. Therefore, the shared control system
achieves better path following performance than a human
driver alone. Besides, a human driver’s lane-changing intention
results and the desired maneuvers are shown in Fig.11(b).
There are several inconsistent areas, and two areas are marked
as 1
and 2
for discussion, where they are also depicted in
Fig. 11(a). The human driver tends to change the current lane
to avoid the ahead obstacles O1and O2behind the desired
lane-changing maneuver in the situation where ttc is lower
than the safety threshold and bigger than the dangerous thresh-
old of two obstacles in both two areas. Therefore, the shared
control system is under the second inconsistent condition, and
the cooperative coefficient kis calculated by the lagged fuzzy
controller, in which the controller intervention is strengthened.
As illustrated in Fig. 11(c), the cooperative coefficient kis
higher than the original value. And the original cooperative
coefficient value is calculated by the fuzzy controller without
considering driving intention. Moreover, the human driver
received a phone call after exceeding O2, and the ringing
phone was taken out from the pocket in the third marked
area 3
. The process of dealing with the phone call made the
human driver distracted. The symptoms of distracted driving
are shown in Fig. 11(a) and (c). The human driver forgot to
keep lane and continued keeping changing lane, and the wide
and violent steering action occurred after the driver realized
the dangerous situation. The distracted driving state presented
in this case is well solved by the shared control strategy as it
helps the human driver keep the host vehicle in the current lane
and avoid the obstacle O3safely. Fig. 11(d) shows the shared
control system’s and a human driver’s steering wheel angle
curve. The former curve has a smaller range than a human
driver and is also smooth. The last subplot depicts the lateral
acceleration. The maximum lateral acceleration of a human
driver reaches 0.4494gwhile the maximum lateral acceleration
of the shared control system is 0.2471g. It proves that the
shared control system makes the host vehicle stabler and safer
than a human driver alone, and even with poor driving state.
B. Scenario II
The curvy road is designed by the clothoid curve, and the
expression of the curve road center lane is given by [47]
yc(x) = y0+δ0x+1
2c0x2+1
6c1x3(36)
where y0and δ0represent the initial lateral offset and heading
direction, respectively, c0refers to the initial curvature and c1
means the curvature changing ratio.
The designed curvy road parameters are shown in the top
left square of Fig. 12(a), and the setting of scenario II is also
in this subplot. The static obstacle is at the middle of right-
hand lane with its longitudinal position 120m far from the
initial position of the host vehicle. The host vehicle moves
with a constant velocity as 20m/s and the dynamic obstacle
vehicle runs along the center line of the left-hand lane with
a constant velocity at 10m/s. The initial longitudinal position
of the dynamic obstacle vehicle is 150m far from the initial
position of the host vehicle. Besides, the path of a human
driver and the shared control system are also shown in this
subplot. The mean error of the human driver path reaches
1.074m and the value of RMSE reaches 0.6070m, while the
mean error of the shared control system path is 0.0475m and
the value of RMSE is 0.0245m. The shared control system
could significantly decrease the lateral deviation to the planned
path. Besides, a human driver’s lane-changing intention results
and the desired maneuvers are compared in Fig. 12(b). In this
scenario, there are also two inconsistent areas discussed and
marked as 1
and 2
. As observed from Fig. 12(a), a human
driver changes the current lane with violent steering wheel
angle at the very close point to the obstacle to avoid the static
1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2865105, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, CYBER-PHYSICAL SYSTEMS IN GREEN TRANSPORTATION 11
Fig. 12. Scenario II: A static obstacle and a dynamic obstacle on a curvy road;
(a) the illustration of scenario II and the vehicle paths of the shared system
and a driver, (b) the driving intention recognized results of a human driver
and the desired maneuvers, (c) the cooperative coefficient, (d) the steering
wheel angle, and (e) the lateral acceleration. ttch1and ttcl1represent the
safety and dangerous threshold of the static O1;ttch2and ttcl2represent
the safety and dangerous threshold of the dynamic O2.
obstacle. And it makes the host vehicle unstable as the maxi-
mum value of lateral acceleration in 1
area reaches 0.3926g
shown in Fig. 12(e). Therefore, the shared control system is
under the second inconsistent condition, and kis calculated
by the lagged fuzzy controller which makes the values of
shared cooperative coefficient larger than the original values.
Besides, the maximum lateral value is only 0.2667gin 1
area.
However, the human driver tries to avoid the dynamic obstacle
O2ahead of the desired lane-changing maneuver, illustrated
in the partially enlarged drawing in Fig. 12(a). It would make
the host vehicle in the dangerous situation as the host vehicle
was close to back obstacle O1. And the controller intervention
is weakened in the advanced inconsistent condition, therefore,
the cooperative coefficient is calculated by the advanced fuzzy
controller which is shown in Fig. 12(c). The maximum lateral
acceleration of a human driver is 0.2842gwhile the shared
control system reduces the maximum lateral acceleration to
0.119gin 2
area. Besides, the steering wheel angle of the
shared control system is smooth on the curvy road, depicted
in Fig. 12(d). From above discussion, the shared control driver
assistance system can help a human driver to avoid obstacles
and obtain a better stability performance on the curvy road in
this scenario.
VI. DI SC US SI ON A ND CO NC LU SI ON
This paper proposed a framework of the shared control
driver assistance system based on the driving intention and
situation assessment. It integrated four modules for path plan-
ning, path following, intention recognition and shared control
strategy. The lane-changing intention of a human driver and
the desired maneuvers of the planned path are recognized
by iMLCU classifier which is a semi-supervised machine
learning method. Then the shared fuzzy controller is designed
according to the consistency between a human driver’s driving
intention and the desired maneuver, the safety level obtained
from the situation assessment, and the performance evaluation
index in our study. Besides, the cooperative coefficient is
calculated by the consistent, the advance inconsistent, and the
lagged inconsistent fuzzy controller in different conditions,
respectively. Moreover, the simulation studies validate that
the system proposed in this paper is effective for a human
driver to avoid obstacles, and obtains a great vehicle stability
performance.
The shared control driver assistance system based on the
driving intention and situation assessment is studied in the
simulated environment where sensors are optimal. It assumes
that the data for simulation studies are accurate and without
noises. However, the lane detection and the vehicle localization
accuracy have great influences on the path following controller
and the intention recognition module, which is training based
on the relative offset and velocity to the road center lane.
When the shared control system proposed in this paper is
applied to the real situations, the steerable filter [28] and
fully convolutional neural networks method [48] provide a
great lane detection performance in complex shadowing or
lighting changing conditions. Besides, the vehicle localization
is obtained with high accuracy by the extended kalman filter
(EKF), Bayesian filter (BF) or interacting multiple model
(IMM) method, which are based on GPS and internal vehicle
sensors [29]. Therefore, the shared control system proposed in
this paper could be feasible in real situations.
In future work, the shared control driver assistance system
will be implemented into an advanced driving simulator for
taking the driver’s reaction and acceptance to the shared sys-
tem into consideration. Besides, the reference path for shared
control system will be updated by considering the driver’s
driving intention or driver’s characteristics in the path planning
module, such as driving style. Furthermore, in order to obtain
better safety and stability performance, the longitudinal control
will be studied for developing the shared control system.
REF ER EN CE S
[1] National Highway Traffic Safety Administration, “Distracted Driving
2015.” Traffic safety facts research note 2017., Washington, DC, USA,
20590, March, 2017.
[2] National Highway Traffic Safety Administration, “Drowsy Driving 2015.”
Traffic safety facts research note 2017., Washington, DC, USA, 20590,
Octerber, 2017.
[3] Katzourakis et al., “Driver steering override for lane-keeping aid using
computer-aided engineering,” IEEE/ASME Trans. Mech., vol. 20, no.4,
pp.1543-1552, 2015.
[4] Dang R. et al., “Coordinated adaptive cruise control system with
lane-change assistance,” IEEE Trans. Intell. Transp. Syst.,vol.16, no.5,
pp.2373-2383, 2015.
[5] Wang J. et al., “A forward collision warning algorithm with adaptation to
driver behaviors,” IEEE Trans. Intell. Transp. Syst., vol. 17, no.4, pp.1157-
1167, 2016.
[6] Katzourakis et al., “Driver steering override for lane-keeping aid using
computer-aided engineering,” IEEE/ASME Trans. Mech., vol.20, no. 4,
pp.1543-1552, 2015.
[7] Bonnefon J.F., Shariff A., and Rahwan I., “The social dilemma of
autonomous vehicles,” Sci., vol.352, no.6293, pp.1573-1576, 2016.
[8] Autonomous driving: Context and state-of-the-art, Handbook of Intelli-
gent Vehicles, Springer, London, UK, 2012, pp. 1271-1310.
[9] Enache N.M. et al., “Driver steering assistance for lane-departure avoid-
ance based on hybrid automata and composite Lyapunov function,” IEEE
Trans. Intell. Transp. Syst., vol.11, no.1, pp.28-39, 2010.
1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2865105, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, CYBER-PHYSICAL SYSTEMS IN GREEN TRANSPORTATION 12
[10] Andrew G. et al., “Predictive control for agile semi-autonomous ground
vehicles using motion primitives,” in Proc. Amer. Control Conf. (ACC),
Jun. 2012, pp.4239-4244.
[11] Sentouh C. et al., “Human-Machine Interaction in Automated Vehicle:
The ABV Project,” IFAC Proc. Vol., vol.47. no.3, pp. 6344-6349, 2014.
[12] D.A. Abbink, M. Mulder, and E.R. Boer, “Haptic Shared Control:
Smoothly Shifting Control Authority,” Cogn. Tech. &Work, vol. 14, pp.
19-28, 2012.
[13] M. Mulder, D.A. Abbink, and E.R. Boer, “Sharing Control with Haptics:
Seamless Driver Support from Manual to Automatic Control,” Hum.
Factors, vol. 54, no. 5, pp. 786-98, 2012.
[14] Erlien S. M., Fujita S., and Gerdes J. C., “Shared steering control using
safe envelopes for obstacle avoidance and vehicle stability,” IEEE Trans.
Intell. Transp. Syst., vol.17, no.2, pp. 441-451, 2016.
[15] F. Flemisch et al., “Towards a Dynamic Balance between Humans and
Automation: Authority, Ability, Responsibility and Control in Shared and
Cooperative Control Situations,” Cogn. Tech. &Work, vol. 14, pp. 3-18,
2012.
[16] Mars F., Mathieu D., and Jean-Michel H., “Analysis of human-machine
cooperation when driving with different degrees of haptic shared control,”
IEEE Trans. Haptic, vol.7, no.3, pp. 324-333, 2014.
[17] Na X. and Cole D.J., “Linear quadratic game and non-cooperative
predictive methods for potential application to modelling driver-AFS
interactive steering control,” Veh. Syst. Dyn., vol.51, no.2, pp.165-198,
2013.
[18] Na X. and Cole D.J., “Game-theoretic modeling of the steering interac-
tion between a human driver and a vehicle collision avoidance controller,”
IEEE Trans. Hum.-Mach. Syst., vol. 45, no.1, pp.25-38, 2015.
[19] Flad M., Fr¨
ohlich L. and Hohmann S., “Cooperative Shared Control
Driver Assistance Systems Based on Motion Primitives and Differential
Games,” IEEE Trans. Hum.-Mach. Syst., vol. 47, no.5, pp.711-722, 2017.
[20] Flad M. et al., “Individual driver modeling via optimal selection of
steering primitives.” in Proc. World Congr., Cape Town, South Afirca,
Aug. 2429, 2014, pp.6276-6282.
[21] Anderson S.J. et al., “An optimal-control-based framework for trajectory
planning, threat assessment, and semi-autonomous control of passenger
vehicles in hazard avoidance scenarios,” Int. J. Veh. Auton. Syst., vol.8,
no.2-4, pp. 190-216, 2010.
[22] Noh S., Kyounghwan A., and Wooyong H., “Situation assessment
and behavior decision for vehicle/driver cooperative driving in highway
environments,” in Autom. Sci. Eng., 2015 IEEE Int. Conf. on, Gothenburg,
Sweden, Aug. 24-28, 2015, pp.626-633.
[23] Gao, Y. et al., “Predictive control of autonomous ground vehicles with
obstacle avoidance on slippery roads,” in Proc. ASME Dyn. Syst. Control
Conf, Cambridge, Massachusetts, USA, Sep. 1215, 2010, pp.265-272.
[24] Saleh L. et al., “Shared steering control between a driver and an
automation: Stability in the presence of driver behavior uncertainty,” IEEE
Trans. Intell. Transp. Syst., vol.14, no.2, pp.974-983, 2013.
[25] Song L. et al. “Model predictive control oriented shared steering control
for intelligent vehicles.” In Control Decis. Conf., 2017 29th Chinese,
Chongqing, China, May 28, 2017, pp.7568-7573.
[26] J. Rawlings and D. Mayne, Model Predictive Control, Theory and
Design.Madison, WI: Nob Hill Publ., 2009
[27] Hu C. et al., “Differential steering based yaw stabilization using ISMC
for independently actuated electric vehicles,” IEEE Trans. Intell. Transp.
Syst., vol.19, no.2, pp.627-638, 2018.
[28] McCall Joel C. and Trivedi M.M. , “Video-based lane estimation and
tracking for driver assistance: survey, system, and evaluation,” IEEE
Trans. Intell. Transp. Syst., vol.7, no.1, pp.20-37, 2006.
[29] Rezaei S. and Raja S., “Kalman filter-based integration of DGPS and
vehicle sensors for localization,” IEEE Trans. Control Syst. Technol.,
vol.15, no.6, pp.1080-1088, 2007.
[30] Le M. and Zhang M.,“Multi-Label classification with unlabeled data:
An inductive approach,” In Pro. of the 5th Asian Conf. on Mach. Learn.,
Canberra, Australia, Oct. 21, 2013, pp.197-212.
[31] Hu C. et al., “Robust Composite Nonlinear Feedback Path-Following
Control for Independently Actuated Autonomous Vehicles With Differ-
ential Steering,” IEEE Trans. Transp. Electrif., vol.2, no.3, pp.312-321,
2016.
[32] Cao H. et al., “Simulation research on emergency path planning of an
active collision avoidance system combined with longitudinal control for
an autonomous vehicle,” Proc IMechE Part D: J Automobile Eng, vol.230,
no.12, pp.1624-1653, 2016.
[33] Huang Y. et al., “Modelling and optimal energy-saving control of
automotive air-conditioning and refrigeration systems,” Proc IMechE Part
D: J Automobile Eng, vol.231, no.3, pp.291-309, 2017.
[34] R. Rajamani, Vehicle Dynamics and Control. New York: SpringerVerlag,
2005.
[35] H. Pacejka, Tyre and Vehicle Dynamics. Warrendale, PA: SAE, 2006.
[36] Wang W. et al., “Driving style classification using a semisupervised
support vector machine,” IEEE Trans. Hum.-Mach. Syst., vol.47, no.5,
pp.650-660, 2017.
[37] Chapelle O., Vikas S., and Sathiya S.K., “Optimization techniques for
semi-supervised support vector machines,” J. Mach. Learn. Res., vol.9,
no.Feb, pp.203-233, 2008.
[38] Attia R., Orjuela R., and Basset M., “Combined longitudinal and lateral
control for automated vehicle guidance,” Veh. Syst. Dyn., vol.52, no.2,
pp.261-279, 2014.
[39] Boada B.L., Boada M.J.L., and Diaz V. “Fuzzy-logic applied to yaw
moment control for vehicle stability,” Veh. Syst. Dyn., vol.43, no.10,
pp.753-770, 2005.
[40] Cao H. et al., “An optimal model-based trajectory following architecture
synthesising the lateral adaptive preview strategy and longitudinal velocity
planning for highly automated vehicle,” Veh. Syst. Dyn., vol.55, no.8,
pp.1143-1188, 2017.
[41] Wang J.et al., “Driving safety field theory modeling and its application
in pre-collision warning system,” Transp. Res. Part C: Emerging Technol.,
vol.72, pp. 306-324, 2016.
[42] Wakasugi T., “A study on warning timing for lane change decision aid
systems based on drivers lane change maneuver,” In Proc. 19th Int. Tech.
Conf. on Enhanced Saf. of Veh., no.05-0290, 2005.
[43] Chen J. et al., “Adaptive fuzzy logic control of fuel-cell-battery hybrid
systems for electric vehicles,” IEEE Trans. Ind. Inf., vol.14, no.1, pp.292-
300, 2018.
[44] Janssen W. and Lena N., An experimental evaluation of in-vehicle
collision avoidance systems. Statens V¨
ag-och Trafikinstitut. VTI s¨
artryck
181, 1992.
[45] Mamdani E.H. and Assilian S., “An experiment in linguistic synthesis
with a fuzzy logic controller,” Int. J. of Man-Mach. Stud., vol.7, no.1, pp.
1-13, 1975.
[46] Lee J.D., Young K.L., and Regan, M.A., “Defining driver distraction,”
in Driver Distraction: Theory, Effects, and Mitigation, Boca Raton, FL,
USA: CRC Press, 2008, pp. 31-40.
[47] Cao H., Song X., and Huang J., “Vehicle path planning in various
driving situations based on the elastic band theory for highway collision
avoidance,” Proc IMechE Part D: J Automobile Eng, vol.227, no.12,
pp.1706-1722, 2013.
[48] Caltagirone L. et al., “Fast LIDAR-based road detection using fully
convolutional neural networks,” In IEEE Intell. Veh. Symp. (IV), Los
Angeles, CA, USA, Jun. 11, 2017, pp.1019-1024.
Mingjun Li is currently a PhD candidate from
College of Mechanical and Vehicle Engineering at
Hunan University since from 2017. And he received
his B.S. degree in the college of the same university
in 2016.
His research interests include the shared control
strategy, vehicle dynamics and control, driver assis-
tance systems, and human driver behaviors analysis
for intelligent vehicle.
Haotian Cao is currently a Ph.D. candidate in
College of Mechanical and Vehicle Engineering at
Hunan University. He received his B.E. degree at
the same school in 2011.
His interests include trajectory planning and fol-
lowing control for autonomous vehicles, technology
related to vehicle dynamical systems, driver behavior
modeling and naturalistic driving data analysis.
1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2865105, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, CYBER-PHYSICAL SYSTEMS IN GREEN TRANSPORTATION 13
Xiaoling Song received the B.E., M.E. and PhD
Degrees at Institute of Mechanical and Vehicle
Engineering, Hunan University in 1988, 1991 and
2007 respectively. From 2008 to the present, she is
consecutively a professor at Hunan University and a
senior research fellow at University of Waterloo.
Her research interests include vehicle active
safety, vehicle dynamics control, driver modeling
and human factors in driving safety.
Yanjun Huang is a Postdoctoral Fellow at the De-
partment of Mechanical and Mechatronics Engineer-
ing at University of Waterloo, where he received his
PhD degree in 2016. His research interest is mainly
on the vehicle holistic control in terms of safety,
energy-saving, and intelligence, including vehicle
dynamics and control, HEV/EV optimization and
control, motion planning and control of connected
and autonomous vehicles, human-machine coopera-
tive driving.
He has published over 50 papers in journals and
conference. He is serving as associate editors and editorial board member of
IET Intelligent Transport System, SAE Int. J. of Commercial vehicles, Int. J.
of Vehicle Information and Communications, Automotive Innovation, AIME,
etc.
Jianqiang Wang received the B.Tech. and M.S.
degrees from Jilin University of Technology, Jilin,
China, in 1994, 1997 respectively, and the Ph.D.
from Jilin University, Jilin, China, in 2002. He is
currently a professor with the Department of Au-
tomotive Engineering, Tsinghua University, Beijing,
China. He is an Editor-in-Chief of Journal of Intel-
ligent and Connected vehicles, an associate editor
of IET Intelligent Transport Systems, an associate
editor of International Journal of Vehicular Telemat-
ics and Infotainment Systems and an editorial board
member of Traffic Information and Safety.
His research interests include intelligent vehicles, driving assistance sys-
tems, V2V/V2I and driver modeling. He has published more than 150 journal
and conference papers. He is the co-inventor of more than 80 patents. Dr.
Wang has engaged in over twenty sponsored projects and has received more
than ten awards. He is the winner of National Science Fund for Distinguished
Young Scholars.
Zhi Huang received the B.E. degree in mechatronic
engineering, M.S. and Ph.D. degree in mechani-
cal engineering from Hunan University, Changsha,
China. He is currently an associate professor with
Hunan University.
His research interests include chassis control, au-
tonomous vehicle, intelligent control and advanced
driving assistance technology.
