Content uploaded by Wei Yang
Author content
All content in this area was uploaded by Wei Yang on Jul 22, 2020
Content may be subject to copyright.
2350 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 21, NO. 6, JUNE 2020
Automated Highway Driving Decision
Considering Driver Characteristics
Wei Yan g , Ling Zheng, Yinong Li, Yue Ren , and Zhoubing Xiong
Abstract— In the background of autonomous driving at level 3
to level 4, an automated vehicle should own smarter driving brain
to face complicated transportation situations. In order to con-
struct a safe automated driving brain under highway conditions,
this paper focused on driving motion decision in order to generate
the control target parameters in the time domain. The coordinate
transformation was proposed to convert the complicated curving
road to local straight coordinate or inverse, then a receding hori-
zon programming based on mixed logic dynamic constraints was
established to formulate a safe-guaranteed optimization model,
where the objectives were assigned by driver’s steering wheel and
speed control, as well as the lateral lane tracking performance.
Based on the motion optimization model, the links to the driver
characteristics were analyzed, and the weight for each objective
in optimization model was tuned by driver statistical features,
in which the entropy weights, variance weights, and unique
weights are compared. The simulation based on the simulating
driving scenario was developed and the optimization results
validated the safety and feasibility of motion decision and with
the help of k-nearest neighbors (KNN) classifier, the clustering
prediction results qualitatively revealed the proposed weights
tuning methods for objectives in optimization model could better
determine a human-like driving decision, furthermore, this paper
gave a basis to compromise multi-objectives in driving decision.
Index Terms—Automated vehicle, motion decision, receding
horizon programming, driver characteristics.
I. INTRODUCTION
AUTOMATED vehicle (AV) has raised great research
potential in recent years. The automatic controller in AV
would achieve faster response than human, meaning that if the
commands are correct, AV will own lower danger probability
meanwhile the optimized control strategies also bring about
better driving performance. Automated highway driving is
more practical and attractive to achieve a higher level of the
intelligent vehicle. At first, high speed makes it more fatal
accidents, and it is easier to be fatigue in the monotonous
driving environment [1], then drivers’ operation delay is prone
Manuscript received July 26, 2018; revised February 13, 2019 and
April 13, 2019; accepted May 17, 2019. Date of publication June 3, 2019; date
of current version May 29, 2020. This work was supported in part by the Key
Research Program of the Ministry of Science and Technology under Grant
2016YFB0100904, and in part by the National Natural Science Foundation
of China under Grant 51875061. The Associate Editor for this paper was B.
Fidan. (Corresponding author: Ling Zheng.)
W. Yang, L. Zheng, Y. Li, and Y. Ren are with the State Key Lab of
Mechanical Transmissions, Chongqing University, Chongqing 400044, China
(e-mail: yangwei0705@gmail.com; zling@cqu.edu.cn; ynli@cqu.edu.cn;
cqurenyue@163.com).
Z. Xiong is with the Changan Automotive Engineering Institute,
Chongqing 400044, China (e-mail: xiongzb@changan.com.cn).
Digital Object Identifier 10.1109/TITS.2019.2918117
to cause traffic jam [2]. Moreover, unidirectional transportation
flow and regular driving principles make it easy to model high-
way driving so that automated highway driving demonstrates
research value and would be a breakthrough for automotive
industrial. So far, many academic organizations and industrial
enterprises have made it possible for automated driving under
limited conditions, where the famous milestones of modern
automobile include Program on Advanced Technology for the
Highway (PATH), DARPA Grand Challenge, Tesla Autopilot,
Google Waymo and so on. Nevertheless, it is still a great
challenge to realize an intelligent highway driving attributed to
complicated driving situations. For instance, a great number of
transportation users, the limited free space and the diversity of
driving performance are hard to compromise in path planning.
In general, the automated driving system is described as the
architecture of “route-motion-control” [3]–[5], where the route
is the macroscopic path from start to finish and it doesn’t take
vehicle dynamics into account. Motion decision and control
are key components, which directly determine the real-time
driving states of the vehicle. Especially, decision making con-
tributes to the intelligence of the vehicle. A risk assessment for
the driving environment is prerequisite to plan an optimizing
motion with consideration of safety, efficiency, comfortability
and so on in the decision process. Thus, the decision system
is more related to an optimization process, which generates
the ideal trajectory to be tracked in control.
Herein, the risk assessment is the prerequisite for decision.
It is an empirical activity of driving based on perceptual
knowledge so that drivers could make a decision despite no
risk extent quantified. As artificial intelligence has not yet
developed to full maturity, it still needs a definite evaluation
index to describe the environmental risk. The artificial poten-
tial is a powerful tool to visualize the abstract risk extent and
indicate the feasible region. Yadollah Rasekhipour et al. [6]
assigned different artificial potential functions to different
types of obstacles and road structures. Hanwool et al. [7]–[9]
proposed the potential feature extraction method based on
Von Mises distribution integrating the Gaussian distribution
to formulate the artificial potential. Extensively, some other
methods are derived from artificial potential, for instance,
Song et al. [10] presented an elastic band theory to assess the
vehicle crash risk, Ren et al. [11] proposed an improved time-
to-collision (TTC) risk model integrating relative distance,
velocity and acceleration to analyze the longitudinal driving
risk within several conditions. Julian and Damerow [12],
Eggert [13], and Damerow et al. [14] calculated the involved
1524-9050 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Michigan Library. Downloaded on July 22,2020 at 23:25:16 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: AUTOMATED HIGHWAY DRIVING DECISION CONSIDERING DRIVER CHARACTERISTICS 2351
risks explicitly to model the lane-change process. All these
methods only give the tendency of the risk, this means it is
hard to make it clear where is the absolute safety and it might
be disordered as plenty of obstacles. Therefore, the artificial
potential field has a few capabilities to deal with complicated
driving situations or multi transportation participators and
obstacles. Another risk evaluation based on relative distance
is to specify the collision boundaries, then the collision con-
straints are formulated by accurate equations, which demon-
strates where is free space or obstacles. Li et al. [15] described
a two-circle equivalent model to limit the collision region.
Li et al. [16] applied more circles on depicting vehicle
shape. Moon et al. [17] presented a geometric area condition,
in details, the piecewise areas of every point to two points of
obstacle should be larger than the total area of this obstacle.
Masakazu et al. [18] formulated a mixed integer problem to
restrict the crash distance by logical operators. It is found that
the more accurate collision conditions are described, the better
safety is achieved, even under a complex driving scenario.
According to risk assessment, motion planning is devel-
oped to guarantee safety. Meanwhile, more driving perfor-
mances are involved. From the view of configuration space,
the driving decision is divided into two categories, that is,
trajectory planning (generating an ideal path in space domain)
and motion planning (generating an ideal path and several
related states in the time domain). As for trajectory planning,
Heil et al. [19] proposed adaptive quintic polynomials.
Zhan et al. [20] emerged a topologic route planning model
basedonA
∗search, then quadratic programming (QP) was
designed to gain the rough long-term longitudinal motions
and short-term trajectories. Xu et al. [21] considered colli-
sion avoidance and the response of motion in the vehicle
together, proposing a set of candidate path. On the other
hand, the motion planning considering dynamic behavior
demonstrates better performance under complicated conditions
as well as motion planning could consider available target
tracking parameters in controllers. Research in aforemen-
tioned [6], [10], [18] gave the real-time motion planning,
and the corresponding vehicle states such as velocity, steer-
ing angle, yaw rate etc. in time series were also obtained.
Moreover, model predictive control (MPC) or receding horizon
programming could optimize the motion decision which takes
both safety and driving performance into account [22]–[24],
parameters tuning for MPC demonstrates more potential in
motion decision.
As the driving decision is a comprehensive process which
does not only consider proper trajectory planning, but also
the feelings of the driver are very significant. In other words,
diverse drivers have various driving habits and driving ten-
dency. Even if it is a duplicate driving scenario, different
drivers might contribute distinct results, which reveals that
driver characteristics deeply affect the driving decision and a
personalized automated driving system would bring about the
more delighted driving experience. It is an obscure problem
to model a driver quantitatively for his serious randomness.
By means of parameters in the driver model, it is possible
to distinguish the driver characteristics in decision-making.
In general, driver model can be investigated through various
methods such as Piece-Wise Auto Regressive eXogenous
model [25], [26], artificial neural networks [27], Gaussian
mixture models [28]–[30], and inverse optimal control [31],
etc. Vallon et al. [32] proposed a data-driven modeling
approach to capture the lane change decision behavior of
human drivers and apply MPC to implement personalized
autonomous driving. Stéphanie et al. [33] trained hidden
Markov model (HMM) and proposed the personalized ADAS
based on MPC. As far as driver concerned, the driving charac-
teristics demonstrated on the preference of driving objectives,
so that researchers used the driver model to match decision
and finally realize personality.
To sum up, an accurate and comfortable automated driving
is associated with precise environmental perception, decision,
and control, where the decision is the key topic in this paper,
and it is assumed that the environmental perception is clear,
and the actuator control will not be discussed. Based on the
space configuration of the decision, a personalized automated
driving motion decision algorithm to achieve a safe, comfort-
able and human-like driving is pursued. Therefore, considering
completeness of the decision, a road coordinate transformation
algorithm was built up to deal with complex road conditions
at first, then an optimization model for motion planning was
developed, herein, MPC was applied with accurate colli-
sion constraints described by mixed logical dynamic (MLD),
as well as the objectives were driving performance weighted
by driver characteristics extraction from statistical features of
human driver. When security is guaranteed by proposed MLD
constraints, the weights tuning for MPC objectives based on
the driver characteristics contribute to a personalized driving
decision and provides the basis for multi driving performances
preference. The main contribution in this paper is that the
driver characteristics are analyzed to be applied to driving
motion decision and a collision-free and personalized driving
planning are successfully achieved.
The rest of the paper is organized as follows. In section 2,
a road coordinate transformation algorithm is completed and
an optimization model for motion planning is developed.
In section 3, the decision weights related to individual driving
are analyzed and determined according to driver character-
istics. In section 4, the simulation test is demonstrated, and
it verifies the proposed method for personalized driving.
In section 5, conclusions and future work discussions in this
paper are summarized.
II. DYNAMIC MOTION PROGRAMMING
A. Coordinate Transformation
Automated vehicle must face variety of driving environ-
ments, even for highway driving. The road is not straight
all the time. Therefore, it is necessary to develop coordinate
transformation to simplify the driving condition and improve
the robustness of automated vehicle in diverse environments.
Usually, straight road is easy to deal with for these reasons:
1) it would be easier to design the collision constraints for
specific x, y and relative distance, etc. 2) it would simplify
the decision model, no path curvature is considered. Therefore,
this paper developed a method which transfers original curving
Authorized licensed use limited to: University of Michigan Library. Downloaded on July 22,2020 at 23:25:16 UTC from IEEE Xplore. Restrictions apply.
2352 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 21, NO. 6, JUNE 2020
Fig. 1. Coordinate transformation.
road to straight road, while the straight road may also be
reversed to the curving road. it is depicted in Fig.1 in details.
According to Fig. 1, the curving road coordinate and straight
road coordinate could mutually convert based on the fixed
curve distance s. Basically, the straight road coordinate is
constructed by total distance on road baseline (x direction) and
the distance between planned path points and road baseline
(y direction). The steps of transformation are described as
follows:
a) Curving road to straight road XOY →xoy
Step 1: calculating the corresponding point (X,Y)on baseline
according with X,Y
X−X1+Y−YK=0(1)
P(X,Y)=0(2)
where P(X,Y)is the acquired road baseline function and
usually they are a series of discrete sampling points.
Step 2: calculating the total distance from C1to C2
s=C2
C1
P(X,Y)ds (3)
Step 3: counterclockwise rotating vector
C2C3and judging
the direction between this new vector Cand the slope of
the baseline, if they are the same direction, the value after
transformation is positive, otherwise, it should be negative.
C:cos (−90◦)−sin (−90◦)
sin (−90◦)cos (−90◦)X−X,Y−Y
=Y−Y,−X+X(4)
d=sgn Y−Y1+−X+XK
X−X2+Y−Y2(5)
Step 4: the coordinates in straight road xoy is
x=s(6)
y=d(7)
k=K−K(8)
b) Straight road to curving road xoy →XOY
Step 1: calculating Xon baseline based on the interpolation
of sand X,thenYand Kare calculated by baseline function
x=s(9)
X:s→X(10)
{Y,K|P(X,Y)=0}(11)
Fig. 2. Vehicle motion model.
Step 2: calculating equation set
X−X1+Y−YK=0 (12)
X−X2+Y−Y2=d2(13)
Step 3: the coordinates in curving road XOY is
X=−sgn (K)dK2
1+K2+X(14)
Y=−1
K
X−X+Y(15)
K=K+k(16)
If the vehicle could obtain an accurate environmental infor-
mation, the coordinate could be easily transformed between
original curving road and straight road by using (1)-(16).
B. Vehicle Model and Predictive States
Highway driving is usually continuous and stable, and
the probability of extreme condition is far less than normal
driving. Furthermore, at the decision level, it is more con-
centrated on ideal motion so that a complex vehicle dynamic
model is not necessary. In other words, the vehicle model is
based on assumptions as:
a) Vehicle motion is considered as the rigid body motion
and the lateral slip is ignored.
b) emergency or extreme conditions is not addressed.
c) Planning motion parameters could be reached by control
˙x=vcos θ(17)
˙y=vsin θ(18)
˙
θ=vtan δ/l(19)
In Fig. 2, Wand Ldenote the total width and length of
the vehicle respectively, and (x,y)is the coordinates of the
midpoint of the rear axle. θindicates the heading angle as
well as δis the angle of steering tire. lmeans wheel base and
vdenotes the longitudinal velocity.
This paper adopts receding horizon dynamic program-
ming to plan the future motion under limited time horizon.
Authorized licensed use limited to: University of Michigan Library. Downloaded on July 22,2020 at 23:25:16 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: AUTOMATED HIGHWAY DRIVING DECISION CONSIDERING DRIVER CHARACTERISTICS 2353
Firstly, the state space of vehicle model is derivative into linear
model by Taylor series.
˙
X=f(x,u)=f(X0,u0)+∂f
∂XX=X0
(X−X0)
+∂f
∂uu=u0
(u−u0)+oX2(20)
At the initial state
f(X0,u0)−∂f
∂XX=X0
X0+∂f
∂uu=u0
u0=0 (21)
Then the high order term is ignored, and variables are dis-
cretized
X(k+1)=t
∂f
∂XX=X0
+INx×NxX(k)+t
∂f
∂uu=u0
u(k)=AX (k)+Bu (k)(22)
the state variables are extended, =xyθδvT, the pre-
dictive states could be written as
˙
=AB
0Nu×NxINu×Nu+B
INu×Nuu
=˜
A+˜
Bu(23)
u=δ v T(24)
where the subscripts Nxand Nuindicate the dimension of the
state vector and control vector u. Let the decision time interval
is T, thus the limited horizon is TNpwith Nppredictive
steps. Regarding all states as outputs, the output coefficient
matrix is ˜
C=IN×N(Nis the dimension of expanded
state ) and output states within Nppredictive steps at
time tis
ϒ=·+·U(25)
where
ϒ=⎡
⎢
⎢
⎢
⎢
⎢
⎣
Y(t+1|t)
Y(t+2|t)
Y(t+3|t)
.
.
.
Yt+Np|t
⎤
⎥
⎥
⎥
⎥
⎥
⎦(N·Np)×1
(26)
=⎡
⎢
⎢
⎢
⎢
⎢
⎣
˜
C˜
A
˜
C˜
A2
˜
C˜
A3
.
.
.
˜
C˜
ANp
⎤
⎥
⎥
⎥
⎥
⎥
⎦(N·Np)×N
(27)
U=⎡
⎢
⎢
⎢
⎢
⎢
⎣
u(t|t)
u(t+1|t)
u(t+2|t)
.
.
.
ut+Np−1|t
⎤
⎥
⎥
⎥
⎥
⎥
⎦(Np−1)×1
(28)
Fig. 3. Environmental perception.
=⎡
⎢
⎢
⎢
⎢
⎢
⎣
˜
C˜
B0··· 0
˜
C˜
A˜
B˜
C˜
B··· 0
˜
C˜
A2˜
B˜
C˜
A˜
B··· 0
.
.
..
.
.....
.
.
˜
C˜
ANp−1˜
B˜
C˜
ANp−2˜
B··· ˜
C˜
B
⎤
⎥
⎥
⎥
⎥
⎥
⎦(N·Np)×(Np−1)
(29)
Sometimes, the predictive horizon needs to be larger for a
danger conditions, for example, the headway distance Shto
obstacles is small. So that an alterable predictive step length
is described as
Np=min Npma x −Sh
Npmax
,Npmin (30)
where, symbolizes the rounding for floating numbers.
Finally, the predictive states are obtained through control
variable U, and the main task is to determine these predictive
control inputs to meet with the driving conditions.
C. Collision Constraints Description
In order to assess the driving risk accurately, the precise col-
lision constraints should be obtained. A novel and coincident
design rule is proposed with the help of mixed logical dynam-
ical (MLD) system [34], [35]. As aforementioned statement,
before motion decision, a coordinate transformation is set up to
ensure the road is straight within receding horizon, therefore,
the collision constraints could be abstracted as Fig. 3.
Generally, safety means the ego vehicle will not collide
on obstacles. In geometry, the obstacles are abstracted as
rectangles, thus if the ego vehicle and obstacle rectangles
will not occupy the same areas within prediction horizon, ego
vehicle is safe. Note that the obstacle states are all predictable
and it is assumed that the prediction of obstacles is accurate.
Then the collision constraints are described by MLD as
|xr|≤sx↔λ1=1 (31)
|yr|≤sy↔λ2=1 (32)
λ1+λ2≤1 (33)
The symbol ↔means “if and only if”, namely, two con-
ditions on both sides of ↔are completely equivalent. The
logical operator λ1,λ
2are introduced to limit the continuous
inequality, in other words, (33) indicates three safe situations,
namely, 1) λ1=0
λ2=0,2)
λ1=1
λ2=0,3)
λ1=0
λ2=1.Thesafemar-
gin in x,ydirection is sx,sy, respectively. In general, sxis
Authorized licensed use limited to: University of Michigan Library. Downloaded on July 22,2020 at 23:25:16 UTC from IEEE Xplore. Restrictions apply.
2354 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 21, NO. 6, JUNE 2020
TAB L E I
COLLISION CONSTRAINTS DESCRIP TION
related to the longitudinal velocity since the safety distance is
positive correlation to time to collision (TTC), on the contrary,
lateral velocity is usually too small than longitudinal velocity
and lateral accommodating space of highway is limited, thus
the safety margins are defined as
sx=L+ξ(vx)·t(34)
sy=W
2+(35)
where is the minimum admissible gap to obstacles, and the
positive correlation function is selected as ξ(vx)=vx.
Another more logical operators are introduced to decompose
absolute operation and Table I gives the relationship.
According to the research in [34], the logical operations
could be converted to inequalities based on big-M method
like
[f(x)≤0] ↔[σ=1]
is equivalent to f(x)≤M(1−σ)
f(x)≥ε+(m−ε)σ(36)
Here, Mand mare defined as
M=max
x∈χf(x)(37)
m=min
x∈χf(x)(38)
εis a small tolerance and χis the definition domain of
the given function. Then, Operators transformation obeys
following inequalities.
λ=σ1σ2is equivalent to ⎧
⎪
⎪
⎨
⎪
⎪
⎩
−σ1+λ≤0
−σ2+λ≤0
σ1+σ2−λ≤1
(39)
Finally, the collision constraints with decision variable U
and auxiliary binary variable bare totally written as linear
matrix form for every obstacle.
INp⊗M1·xy ·+θxy ·U+INp⊗M2·b
≤(⊗M3)(40)
Herein,
M1
=−111−100000··· 0
0000−111−10··· 0T
15×2
(41)
M2
=
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
Mx00000
mx−ε00000
0Mx0000
0mx−ε0000
00My000
00my−ε000
000My00
000my−ε00
−100010
0−10 010
1100−10
00−1001
000−101
00110−1
000011
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦15×6
(42)
M3
=
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
−xr+sx+Mx
xr−sx−ε
xr+sx+Mx
−xr−sx−ε
−yr+sy+My
yr−sy−ε
yr+sy+My
−yr−sy−ε
0
0
1
0
0
1
1
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦15×1
(43)
b
=σt
1σt
2σt
3σt
4λt
1λt
2···T
(6·Np)×1∈{0,1}(44)
=11··· 1T
Np×1(45)
xy and xy are calculated based on (27), (29) with the output
matrix ˜
C=1
0
0
1
0
0
0
0
0
0.⊗means Kronecker product.
D. Bounds and Objectives
As for dynamic process, the predictive states and collision
constraints in every step have been developed, then, a multi-
objective optimization model involved different weights can
be considered. Although this would bring about subopti-
mal or dominated solutions, the weighted sum multi-objective
formula is still an efficient trick to obtain the feasible solutions
which ensure the collision avoidance during driving. There-
fore, the four objectives are considered.
1) The control cost is described as
J1=1
Npt+T·Np
i=tuT
iRui(46)
Authorized licensed use limited to: University of Michigan Library. Downloaded on July 22,2020 at 23:25:16 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: AUTOMATED HIGHWAY DRIVING DECISION CONSIDERING DRIVER CHARACTERISTICS 2355
2) In highway driving, less steering is required to keep
stability and reduce control cost.
J2=1
Npt+T·Np
i=tδT
iR2δi(47)
3) speed fluctuation should be small, which means ego
vehicle maintains car-following speed or free-driving
speed.
J3=1
Npt+T·Np
i=tvi−vref TR3vi−vref (48)
4) Less lane departure is required which helps the vehicle
track the lane under any driving conditions.
J4=yt+T·Np−ymidTR1yt+T·Np−ymid (49)
Equation (46) – (49) can be written as a general form
J=Ji=R1(Kϒ−ymid)2+1
Npϒ−ϒref T
×Qϒ−ϒref +uTRu (50)
where
K=0··· 0
!"# $
N·(Np−1)
01000
!"# $
NN·Np×1
(51)
ϒref =⊗⎡
⎢
⎢
⎢
⎢
⎣
0
0
0
0
vref
⎤
⎥
⎥
⎥
⎥
⎦
(52)
Q=INp×Np⊗diag 000R2R3(53)
R=I(Np−1)×1⊗diag R2R3(54)
is the same coefficient matrix as (45). R1,R2,R3are weight
coefficients.
At last, the optimization problem could be formulated as:
Optimization objective: (50)-(54)
Subjects to:
a) collision constraints: (40)-(45) for every obstacle.
b) bounds of state: min ≤≤max
c) bounds of control rate: umin ≤u≤umax
According to proposed decision model, the intelligent vehi-
cle could realize motion planning involved multi transportation
participators, Meanwhile, intelligent vehicle could also avoid
dynamic obstacles adaptively by means of dynamic optimiza-
tion in planning.
III. DETERMINATION OF THE DECISION WEIGHTS
A. Decision Making Scheme
In the proposed optimization model, the weight for every
objective is key parameters which provide diverse candi-
date trajectories. Moreover, these weights are also closely
related to the driver’s characteristics. Generally, vehicle keeps
two states on the highway, that is, lane changing and lane
Fig. 4. Driving simulator.
keeping [36], [37], correspondingly. As far as driver con-
cerned, the steering wheel angle and the speed are controlled
directly by drivers, therefore, these two parameters basically
reflect driver characteristics. Meanwhile, the heading angle and
lateral position are another two key factors associated with
driver characteristics. However, the heading angle could be
simplified as the integration of small steering wheel input,
therefore, heading angle is regarded as the derived attribute
from steering wheel input. Similarly, longitudinal acceleration,
steering frequency, and many other measurable attributes are
derived values from three independent indexes (steering wheel
angle, speed and lateral position), too. Thus, three measur-
able indexes, namely, steering wheel angle, speed and lateral
position to midline are selected for driver characteristics.
In addition, it is assumed that the capability of the driver
(or driver characteristic) is constant in identification horizon,
their statistic features should obey the constant regularities
of distribution, this implies that the principles of statistics
represent the driver characteristic for a steady driving state.
Furthermore, since the driver features in lane changing and
lane keeping usually different, the respective weights are inte-
grated to indicate the total characteristics for the driver. Finally,
the weights of driving objectives are derived as
W=rlc rlk wlclp wlctwlcs
wlklp wlktwlks(55)
Herein, rlc and rlk mean the ratio of lane changing and lane
keeping, wdenotes the respective weights in each driving
process. Furthermore, Wis configured as the corresponding
weights for each driving objectives in (50) based on the feature
statistics.
B. Driving Simulation Experiment
In order to discover the driver features, a data collection
test based on driving simulator is demonstrated. Fig. 4 depicts
the driving simulator in experiment. The driving scenario is
designed with a straight three-lane highway. Because the test
focuses on lane changing features, the curving road driving
Authorized licensed use limited to: University of Michigan Library. Downloaded on July 22,2020 at 23:25:16 UTC from IEEE Xplore. Restrictions apply.
2356 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 21, NO. 6, JUNE 2020
Fig. 5. Driving trajectory in simulator global coordinate.
TAB L E II
LANE CHANGING AND LANE KEEPING DURATION INDRIVING PROCES S
is ignored. Finally, the description for test and date collection
is shown as tables in Appendix in details.
C. Weights Determination
In simulating driving, the global positions for driver 1 were
recorded as Fig. 5.
The lane changing sections are segments through continuous
wave peak and trough across road marker lines, so that the
corresponding lane changing durations could be measured.
rlc =tlc
ttotal
(56)
rlk =tlk
ttotal
(57)
For four driver samples, the validating fragments of their lane
changing and lane keeping maneuver duration time are as
Table II.
The maneuver durations reflect the driver decision which
are used to standardize the maneuver preference of lane
changing and lane keeping so that the corresponding weights
are calculated as (56) and (57).
Lane changing and lane keep are two independent processes,
therefore they are separated via maneuver time duration,
meanwhile, the corresponding measurable indexes for driver
characteristics are derived out respectively. Fig. 6 shows
measurable indexes for driver 1 within a specific segment.
Statistical index reflects the features for a period, therefore,
the weights for each driving performance objective could
be manipulated by data statistics. Herein, two assessment
methods are adopted and compared, namely, standard devi-
ation and entropy weights, which can demonstrate some sim-
ilar properties for drivers, especially for continuous random
variables. Therefore, weights for driving objectives as (50)
are configured by standard deviation and entropy weights,
furthermore, a unique weight are also compared to verify the
improvements of proposed weights tuning methods.
Fig. 6. Driving parameters in lane changing and lane keeping.
Three indexes X1,X2,X3are considered and every index
is described Xi=%x1,x2,x3,··· ,xj&with jsamples.
Step 1: Standardizing of driving data
Yij =Xij −min (Xi)
max (Xi)−min (Xi)(58)
Step2: Weights calculation
a) standard deviation (variance weights)
wi='1
nn
j=1Yij −¯
Yi2(59)
b) entropy weights
Calculation of information entropy
Ej=−ln 1
nn
i=1pij ln pij (60)
Herein,
pij =Yij
(n
i=1Yij
(61)
It is defined that
lim
pij→0pij ln pij =0 (62)
the derivation of weights
wi=1−Ei
3−(Ei
,i=1,2,3 (63)
c) unique weights
W=[1,1,1] (64)
Finally, every weight for steering wheel angle, speed and
lateral position in lane changing and lane keeping process are
obtained. The larger statistical weights indicate the greater
variation of the index, which means the smaller objective
importance in optimization model (50). Thus, the total weights
of each optimization objectives are calculated after standard-
izing, that is
R1R2R3=γ2
W−12
(65)
Herein, γis the dimensionless coefficient.
γ=1
ymax
1
δmax
1
vmax−vre f (66)
Authorized licensed use limited to: University of Michigan Library. Downloaded on July 22,2020 at 23:25:16 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: AUTOMATED HIGHWAY DRIVING DECISION CONSIDERING DRIVER CHARACTERISTICS 2357
TABLE III
SIMULATION SETTINGS
Fig. 7. Relative distance to ego vehicle.
Fig. 8. Real driving trajectory and motion planning by decision algorithm.
IV. SIMULATIONS
To verify the proposed algorithm, the motion decision
simulation under the same scenario as simulating driving
experiment is demonstrated and the main settings are shown
in Table III.
Four simulations for proposed automated driving decision
algorithm are demonstrated based on four driver samples
Fig. 7 depicts the relative distance between obstacles and
ego vehicle with different optimization weight planning, and it
reveals that the proposed MPC decision model with MLD con-
strains are valid for collision avoidance. Because the collision
constraints strictly restrict the collision, the objective weights
will not influence on the safety. Therefore, the algorithm
guarantees the collision-free basically.
Fig. 8 demonstrates the planning trajectories based on three
kind of objective weights. As statements, the entropy weights
almost have the same consequences as variance weights within
the range of errors because they all indicate the fluctuation
of states and the relative ratio among states are similar,
furthermore, the sensitivity of MPC optimization model allows
a limited difference between entropy weights and variance
weights, which would not be detailly discussed in this paper.
Obviously, the results by entropy weights and variance weights
have large discrepancy with unique weights. Although for
TAB L E IV
KNN CLUSTER MODEL ACCURACY
Fig. 9. KNN clustering verification.
driver 3, the bias of three planning trajectories are small, it
reveals that the basic target for planning is collision free,
and under this pre-requirement, the final planning trajectories
maybe similar. Hence, in order to quantitatively evaluate three
weights configuration, driver models are trained by classifier
with inputs of steering wheel angle, speed and lateral position,
and it will be used to verify which kind of objective weights
contribute to a human-like decision. It is found that KNN
classifier has better performance on driver model, and the
accuracy of classifiers are listed in Table IV.
Therefore, KNN driver model is applied and the 12 sets
(4 drivers and 3 kinds of weights) of data are verified as Fig. 9.
Fig. 9 is the confusion matrix of prediction for four drivers
and the diagonals indicate the correct prediction for each
driver. It is evident that the planning by variance weights
and entropy weights are more accurate than unique weights,
which means the motion behavior by variance weights and
entropy weights are more like the corresponding drivers, and
despite the unique weights could also generate the safe motion
planning. However, it cannot reflect the real human drivers’
features.
V. C ONCLUSIONS AND FUTURE WORK
This paper proposed a novel motion decision framework
for highway automated driving. A coordinate transformation
was developed to deal with various curving road conditions
and simplified the decision process under straight road coor-
dinate, then a receding horizon optimization model based on
mixed logic dynamic constraints was set up to formulate
a basic collision-free decision model and plan the vehi-
cle target motion states. In order to achieve a personalized
driving motion decision, the driver’s statistical features were
considered. By tuning of objective weights in basic motion
optimization model, the link between driver characteristics
Authorized licensed use limited to: University of Michigan Library. Downloaded on July 22,2020 at 23:25:16 UTC from IEEE Xplore. Restrictions apply.
2358 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 21, NO. 6, JUNE 2020
TAB L E V
SIMULATING DRIVING SCENARIO DESI GN
TAB L E VI
SIMULATING DRIVING TES TER
TAB L E VII
SIMULATING DRIVING PROCEDURE
and automated driving was developed. The simulation based
on the simulating driving scenario was developed and the
optimization results validated the safety and feasibility of
motion decision and with the help of k-nearest neighbors
classifier, the clustering results qualitatively revealed the pro-
posed weights tuning methods for objectives in optimization
model could better determine a human-like driving decision.
The algorithm in this paper aims to build up a safe, comfort
and better lane tracking motion decision so that the optimized
motion is usually different from driver’s maneuver, but by
application of driver characteristics, the automated driving
motion would meet more features with driver model, that
means, the proposed methods achieved an automated driving
with better performance than human and better personality
than crude planning.
However, driver characteristics are time-varying and cou-
pled with too many states, so that a key potential for this paper
is that some more derived parameters need to be discussed in
future work. In addition, this paper is an independent part
without considering upper-level environment and lower level
control implementation, therefore, it is just an abstract achieve-
ment for the automated vehicle, which is hard to independently
test in the real world. Finally, as for complicated transportation
participators and environment, the proposed methods need to
model every potential obstacle for constraints respectively,
therefore, the calculation cost needs to be optimized in the
real application.
APPENDIX
See Table V–VII.
REFERENCES
[1] P. Thiffault and J. Bergeron, “Monotony of road environment and driver
fatigue: A simulator study,” Accident Anal. Prevention, vol. 35, no. 3,
pp. 381–391, May 2003.
[2] T. Nagatani and K. Nakanishi, “Delay effect on phase transitions in
traffic dynamics,” Phys. Rev. E, Stat. Phys. Plasmas Fluids Relat.
Interdiscip. Top., vol. 57, no. 6, p. 6415, Jun. 1998.
[3] D. Bevly et al., “Lane change and merge maneuvers for connected and
automated vehicles: A survey,” IEEE Trans. Intell. Vehicles, vol. 1, no. 1,
pp. 105–120, Mar. 2016.
[4] D. González, J. Pérez, V. Milanés, and F. Nashashibi, “A review of
motion planning techniques for automated vehicles,” IEEE Trans. Intell.
Transp. Syst., vol. 17, no. 4, pp. 1135–1145, Apr. 2016.
[5] W. Schwarting, J. Alonso-Mora, and D. Rus, “Planning and decision-
making for autonomous vehicles,” Annu. Rev. Control, Robot., Auton.
Syst., vol. 1, pp. 187–210, May 2018.
[6] Y. Rasekhipour, A. Khajepour, S. Chen, and B. Litkouhi, “A potential
field-based model predictive path-planning controller for autonomous
road vehicles,” IEEE Trans. Intell. Transp. Syst., vol. 18, no. 5,
pp. 1255–1267, May 2017.
[7] W. Hanwool et al., “Dynamic potential-model-based feature for lane
change prediction,” in Proc. IEEE Int. Conf. Syst., Man, Cybern. (SMC).
Budapest, Hungary, Oct. 2016, pp. 838–843.
[8] W. Hanwool et al., “Lane-change detection based on vehicle-trajectory
prediction,” IEEE Robot. Autom. Lett., vol. 2, no. 2, pp. 1109–1116,
Apr. 2017.
[9] W. Hanwool et al., “Automatic detection method of lane-changing
intentions based on relationship with adjacent vehicles using artificial
potential fields,” Int. J. Automot. Eng., vol. 7, no. 4, pp. 127–134, 2016.
[10] X. Song, H. Cao, and J. Huang, “Vehicle path planning in various
driving situations based on the elastic band theory for highway collision
avoidance,” Proc. Inst. Mech. Eng., D, J. Automobile Eng., vol. 227,
no. 12, pp. 22–1706, Dec. 2013.
[11] Y. Ren et al., “The research on the vehicle collision avoidance control
based on vehicle motion estimation,” in Proc. IET Int. Conf. Intell.
Connected Vehicles (ICV), Chongqing, China, Sep. 2016, pp. 1–6.
[12] E. Julian and F. Damerow, “Complex lane change behavior in the
foresighted driver model,” in Proc. IEEE 18th Int. Conf. Intell. Transp.
Syst. (ITSC), Las Palmas, Spain, Sep. 2015, pp. 1747–1754.
[13] J. Eggert, “Predictive risk estimation for intelligent ADAS functions,”
in Proc. IEEE Conf. Intell. Transp. Syst. (ITSC), Qingdao, China,
Oct. 2014, pp. 711–718.
[14] F. Damerow, B. Flade, and J. Eggert, “Extensions for the foresighted
driver model: Tactical lane change, overtaking and continuous lateral
control,” in Proc. IEEE Intell. Vehicles Symp. (IV). Gothenburg, Sweden,
Jun. 2016, pp. 186–193.
[15] B. Li, Y. Zhang, Y. Ge, Z. Shao, and P. Li, “Optimal control-based
online motion planning for cooperative lane changes of connected
and automated vehicles,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots
Syst. (IROS), Vancouver, BC, Canada, Sep. 2017, pp. 3689–3694.
[16] X. Li, Z. Sun, D. Cao, D. Liu, and H. He, “Development of a new
integrated local trajectory planning and tracking control framework for
autonomous ground vehicles,” Mech. Syst. Signal Process., vol. 87,
pp. 118–137, Mar. 2017.
[17] J. Moon, I. Bae, and S. Kim, “Real-time near-optimal path and
maneuver planning in automatic parking using a simultaneous dynamic
optimization approach,” in Proc. IEEE Intell. Vehicles Symp. (IV),
Redondo Beach, CA, USA, Jun. 2017, pp. 193–196.
[18] M. Masakazu, M. Junichi, K. Taketoshi, N. Hikaru, T. Yoshitaka,
and D. Yoshitaka, “Optimal path generation for automotive collision
avoidance using mixed integer programming,” SICE J. Control, Meas.,
Syst. Integr., vol. 1, no. 3, pp. 222–226, May 2008.
[19] T. Heil, A. Lange, and S. Cramer, “Adaptive and efficient lane
change path planning for automated vehicles,” in Proc. IEEE 19th Int.
Conf. Intell. Transp. Syst. (ITSC), Rio de Janeiro, Brazil, Nov. 2016,
pp. 479–484.
[20] W. Zhan, J. Chen, C. Chan, C. Liu, and M. Tomizuka, “Spatially-
partitioned environmental representation and planning architecture for
on-road autonomous driving,” in Proc. IEEE Intell. Vehicles Symp. (IV),
Redondo Beach, CA, USA, Jun. 2017, pp. 632–639.
[21] G. Xu, L. Liu, Y. Ou, and Z. Song, “Dynamic modeling of driver
control strategy of lane-change behavior and trajectory planning for
collision prediction,” IEEE Trans. Intell. Transp. Syst., vol. 13, no. 3,
pp. 1138–1155, Sep. 2012.
Authorized licensed use limited to: University of Michigan Library. Downloaded on July 22,2020 at 23:25:16 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: AUTOMATED HIGHWAY DRIVING DECISION CONSIDERING DRIVER CHARACTERISTICS 2359
[22] H. Guo, J. Liu, D. Cao, H. Chen, R. Yu, and C. Lv, “Dual-envelop-
oriented moving horizon path tracking control for fully automated
vehicles,” Mechatronics, vol. 50, pp. 422–433, Apr. 2018.
[23] M. Elbanhawi, M. Simic, and R. Jazar, “Receding horizon lateral vehicle
control for pure pursuit path tracking,” J. Vib. Control, vol. 24, no. 3,
pp. 619–642, Feb. 2018.
[24] Q. Cui, R. Ding, B. Zhou, and X. Wu, “Path-tracking of an autonomous
vehicle via model predictive control and nonlinear filtering,” Proc. Inst.
Mech. Eng., D, J. Automobile Eng., vol. 232, no. 9, pp. 1237–1252,
2018.
[25] H. Okuda, N. Ikami, T. Suzuki, Y. Tazaki, and K. Takeda, “Modeling
and analysis of driving behavior based on a probability-weighted ARX
model,” IEEE Trans. Intell. Transp. Syst., vol. 14, no. 1, pp. 98–112,
Mar. 2013.
[26] T. Lin, E. Tseng, and F. Borrelli, “Modeling driver behavior during
complex maneuvers,” in Proc. Amer. Control Conf. (ACC), Jun. 2013,
pp. 6448–6453.
[27] D. A. Pomerleau, “Alvinn: An autonomous land vehicle in a neural
network,” in Proc. Adv. Neural Inf. Process. Syst., 1989, pp. 305–313.
[28] V. A. Butakov and P. Ioannou, “Personalized driver/vehicle lane change
models for ADAS,” IEEE Trans. Veh. Technol., vol. 64, no. 10,
pp. 4422–4431, Oct. 2015.
[29] S. Ramyar, A. Homaifar, S. M. Salaken, S. Nahavandi, and A. Kurt,
“A personalized highway driving assistance system,” in Proc. IEEE
Intell. Vehicles Symp. (IV), Redondo Beach, CA, USA, Jun. 2017,
pp. 1596–1601.
[30] J. Xie, J. Gong, S. Wu, G. Xiong, and C. Lu, “A personalized
curve driving model for intelligent vehicle,” in Proc. IEEE Int. Conf.
Unmanned Syst. (ICUS), Beijing, China, Oct. 2017, pp. 301–306.
[31] S. Levine and V. Koltun, “Continuous inverse optimal control with
locally optimal examples,” 2012, arXiv: 1206.4617. [Online]. Available:
https://arxiv.org/abs/1206.4617
[32] C. Vallon, Z. Ercan, A. Carvalho, and F. Borrelli, “A machine
learning approach for personalized autonomous lane change initiation
and control,” in Proc. IEEE Intell. Vehicles Symp. (IV), Jun. 2017,
pp. 1590–1595.
[33] S. Lefèvre, A. Carvalho, Y. Gao, H. E. Tseng, and F. Borrelli, “Driver
models for personalised driving assistance,” Vehicle Syst. Dyn., vol. 53,
no. 12, pp. 1705–1720, 2015.
[34] A. Bemporad and M. Morari, “Control of systems integrating logic,
dynamics, and constraints,” Automatica, vol. 35, no. 3, pp. 407–427,
Mar. 1999.
[35] Y. Du, Y. Wang, and C.-Y. Chan, “Autonomous lane-change controller,”
in Proc. IEEE Intell. Vehicles Symp. (IV), Seoul, Korea, Jun./Jul. 2015,
pp. 386–393.
[36] X. Wang, J. Zhang, X. Ban, and D. Tan, “Driver’s behavior and decision-
making optimization model in mixed traffic environment,” Adv. Mech.
Eng., vol. 7, no. 2, pp. 1–23, Nov. 2014.
[37] G. Cesari, G. Schildbach, A. Carvalho, and F. Borrelli, “Scenario model
predictive control for lane change assistance and autonomous driving
on highways,” IEEE Intell. Transp. Syst. Mag., vol. 9, no. 3, pp. 23–35,
Jul. 2017.
Wei Yang received the B.S. degree in mechanical
engineering from Chongqing University, Chongqing,
China, in 2015, where he is currently pursuing
the Ph.D. degree in automotive engineering. His
research interests include intelligent vehicle on deci-
sion and control, personalized driving, and ADAS.
Ling Zheng was born in Beijing, China, in 1963.
She received the B.S. and M.S. degrees in mechan-
ical engineering from Chongqing University, China,
in 1984 and 1989, respectively, and the Ph.D. degree
in automobile engineering from Chongqing Univer-
sity in 2005. From 2005 to 2006, she was a Visiting
Scholar with the Smart Material and Structure Lab-
oratory, University of Maryland, USA.
From 1984 to 1986, she was a Research Assis-
tant with the Mechanical Engineering Department,
Chongqing Jiaotong University, where she was an
Assistant Professor from 1989 to 1997. Since 1999, she has been an Associate
Professor and a Professor with the Automobile Engineering Department,
Chongqing University. She has authored four books, more than 150 articles,
and more than 20 inventions. She is also the Chairman of structural sessions
in ICSV19 (International Congress of Sound and Vibration), Egypt, 2010,
ICSV22, Italy, 2015, ICSV23, Greece, 2016. Her research interests include
magneto-rheological (MR) fluid and its application in automobile engineering,
such as MR damper, MR engine mount, semi-active suspension system, semi-
active and active vibration and noise control, smart vehicle and dynamic
control in chassis, and information fusion and advanced driving assistant
system (ADAS). She is an Associate Editor of journal Advances in Mechanical
Engineering.
Dr. Zheng was a recipient of the China Automotive Industry Science and
Technology Progress Award for First Prize in 2016.
Yinong Li received the B.S. degree in automotive
engineering from Changan University, Xi’an, China,
in 1983, the M.S. degree in automotive engineer-
ing from Chongqing University, Chongqing, China,
in 1993, and the Ph.D. degree in mechanical engi-
neering from Northeastern University, Shenyang,
China, in 1999.
From 1999 to 2002, he was a Research Assis-
tant Professor with the State Key Laboratory of
Mechanical Transmissions, Chongqing University,
where he has been a Professor since 2002. He has
authored three books, more than 180 articles, and more than 10 inventions.
His research interests include vehicle system dynamics and control, vehicle
chassis integrated control, intelligent vehicles, active, and semi-active control
of vibration and noise for vehicles.
Dr. Li is the member of the Chinese Vibration Engineering Society, and
of the Vibration and Noise Control Committee of the Chinese vibration
engineering society. He was a recipient of the State Science and Technology
Progress Award for Second Prizes, China, in 2008. He is an Editor of the
Journal of Vibration Engineering.
Yue R e n received the B.S. degree in automo-
tive engineering from Chongqing University, China,
in 2013, where he is currently pursuing the
Ph.D. degree (skipping master’s degree) in auto-
motive engineering. His current research direction
is autonomous vehicles, including vehicle detection,
path planning and tracking, vehicle dynamics, and
stability control.
Zhoubing Xiong was born in Fengdu, Chongqing,
China, in 1984. He received the B.S. degree in
communication engineering from the Harbin Insti-
tute of Technology, Heilongjiang, China, in 2008,
and the M.S. and Ph.D. degrees in communication
engineering from the Politecnico di Torino, Turin,
Italy, in 2010 and 2014, respectively.
Since 2015, he has been an Engineer with the
Changan Automotive Engineering Institute, Changan
Automobile, Chongqing. His research interests focus
on environment perception for autonomous driving,
including the development of sensor fusion algorithms, high-definition map-
ping generation, and hybrid and cooperative localization.
Authorized licensed use limited to: University of Michigan Library. Downloaded on July 22,2020 at 23:25:16 UTC from IEEE Xplore. Restrictions apply.
