Received 27 September 2023, accepted 8 October 2023, date of publication 13 October 2023, date of current version 24 October 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3324422

Optimal Lane Changing Problem of Vehicle

Handling Inverse Dynamics Based on
Mesh Refinement Method
YINGJIE LIU 1, DAWEI CUI 1, AND WEN PENG 2
1 School of Machinery and Automation, Weifang University, Weifang, Shandong 261061, China
2 State Key Laboratory of Rolling and Automation, Northeastern University, Shenyang 110819, China

Corresponding authors: Dawei Cui ([email protected]) and Yingjie Liu ([email protected])

This work was supported in part by the Science and Technology Program Foundation of Weifang under Grant 2015GX007; and in part by
the Open Research Fund of the State Key Laboratory of Rolling and Automation, Northeastern University, under Grant
2021RALKFKT008.

ABSTRACT This paper proposes a new optimal control method to solve traffic accidents such as vehicle
collisions during driver’s conscious lane changing under the dangerous state of highway driving and meet the
requirements of safety and comfort during vehicle lane changes. First, by directly analyzing the interpolation
error and slope of discrete nodes, a new mesh refinement method was used to solve the vehicle optimal lane-
changing problem. Then, a simulation of vehicle emergency left and right lane changes was established
to verify the proposed method based on a nonlinear 7-DOF vehicle model. Finally, a virtual test using the
CarSim software was conducted to verify the feasibility of the simulated results. The simulation results show
that the mesh generated by the method in this study is smaller and requires fewer grid iterations, which can
quickly and accurately solve the vehicle optimal lane-changing problem with constant components of the
optimal control variables and fewer mesh iteration times and node numbers with high accuracy. And also,
compared with the traditional Gaussian pseudospectral method(GPM), the solution accuracy of the mesh
refinement method is higher than that of the traditional Gaussian pseudospectral method, indicating the
superiority of the proposed method.

INDEX TERMS Optimal control, vehicle lane changing, minimum time, mesh refinement.

I. INTRODUCTION thousands of families. Therefore, there is an urgent need to

With the increase in car ownership and driving speed, road carry out safety reforms for automobiles.
traffic safety has become a social problem that urgently needs Vehicles driving on a road can be divided into two types:
to be solved worldwide. Road traffic congestion and park- lane-changing and lane-keeping. Lane change is a common
ing space shortages have become popular topics in society. driving behavior and condition. Drivers are vulnerable to the
The speed of vehicles has also significantly improved. Some impact of people, vehicles, roads, and the surrounding envi-
people may drive too fast to save time, which will signifi- ronment when making lane-changing decisions, which affect
cantly increase the probability of vehicle accidents worldwide driving safety and traffic operation efficiency. Compared
every year. According to big data statistics, approximately with other driving behaviors, human drivers greatly increase
500000 people are killed, and approximately 10 million are their psychological pressure when driving vehicles to change
injured in traffic accidents every year. Traffic accidents not lanes. The technical difficulty of lane-changing behavior is
only cause huge economic losses but also cause disasters to higher than that of driving behaviors such as lane keeping.
Therefore, it is more likely to cause traffic accidents owing
The associate editor coordinating the review of this manuscript and to human factors when changing lanes. Therefore, improving
approving it for publication was Jjun Cheng . the ability of vehicles to change lanes independently is of

2023 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 11, 2023 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 115617
 Y. Liu et al.: Optimal Lane Changing Problem of Vehicle Handling Inverse Dynamics

great significance for improving road traffic efficiency and vehicles to ensure multiagent safety and ride comfort. Mutha-
vehicle driving safety [1], [2]. lagu et al. [20] proposed a compact and efficient multi-stage
Vehicle lane changes have been extensively studied. convolutional neural network architecture. Hu et al. [21]
A brief review of the relevant literature is presented below. studied the capture efficiency of wind energy capture devices
To address the influencing factors that make the and the interaction between wind energy capture devices and
lane-changing process ambiguous and uncertain, Li and vehicles in the multilane working conditions of cars and
Zhang [3] applied the membership function of a soft set trucks. Cheng et al. [22] proposed a model of mandatory
to quantify the impact of typical ambiguous factors and lane-changing behavior based on a noncooperative game in
relationships. He et al. [4] proposed a theoretical method a traditional environment and analyzed its applicability in a
to discuss the influence of implementing an autonomous connected environment. Liu and Liu [23] proposed a hybrid
vehicle-dedicated lane policy on freeway traffic efficiency. optimization scheme based on a dynamic adaptive salp swarm
To address the problem of effects on main-lane traffic after the algorithm and the Gaussian pseudospectral method because
merging position, Wang et al. [5] proposed an on-ramp merg- of its strong global search capability.
ing strategy for connected and autonomous vehicles based The above optimization methods had certain limitations:
on the collaboration of downstream main-lane vehicles and some of the methods provided in the above documents had
on-ramp vehicles. Falsone et al. [6] addressed a multivehicle large solution errors; some could not converge to the opti-
automated driving scenario in which a vehicle had to change mal solution; some had low accuracy; for some methods,
lanes and merge in a platoon in a one-way roadway with when the objective function was complex and changeable,
two lanes. Hongbo et al. [7] proposed a classified decision the number of matching points of the optimization function
control execution for a lane departure assistance system must be increased to improve the optimization accuracy to
under the premise of ensuring vehicle stability. Hu et al. a certain extent; for some methods, with an increase in the
[8] presented a novel approach for combining lane detection number of coordination points, the dimension of the state
and model predictive control to maintain the accuracy and differential matrix also increased, resulting in a large amount
stability of the trajectory tracking control for autonomous of calculation and low efficiency, and the convergence rate of
vehicles. Chen et al. [9] proposed a lane-exchanging driving some methods would be reduced owing to the discontinuity
strategy for an autonomous vehicle to cooperate with a nearby of control variables.
vehicle by integrating the vehicle trajectory prediction and In view of the above limitations, in this paper, a new
motion control. To address the failure to provide satisfactory mesh refinement method is proposed by directly analyzing
accuracy of Dead Reckoning, Fei et al. [10] proposed a posi- the interpolation error and slope of discrete nodes, which
tioning method based on traversal optimization for MIMU can solve the optimization problem of minimum-time lane
calibration parameters. Zhou et al. [11] formulated the lane- changing with high accuracy and fewer mesh iteration times
changing decision-making of multiple AVs in a mixed-traffic and node numbers.
highway environment as a multi-agent reinforcement learn-
ing (MARL) problem. Wang et al. [12] studied the process II. MATHEMATICAL DESCRIPTION OF LANE CHANGING
of automatic lane change of vehicles on low-adhesion roads OPTIMIZATION PROBLEM
driven by a front-drive automobile with an in-wheel motor. The lane-changing optimization problem can be regarded as
The form of the road and path of an autonomous vehicle the optimal control problem in control theory. Taking the
were linearized to reduce the calculation time required to Bolza-type optimal control problem as an example, it can
determine the optimal path [13]. To reduce emergency lane be described as solving the control variable z(t) ∈ Rm to
congestion, Li et al. [14] introduced the logistic regression minimize the following objective functions:
theory and proposed a vehicle detection method based on Z tf
logistic regression. Jehanfo et al. [15] presented a holistic J = ψ(x(t0 ), t0 , x(tf ), tf ) + g(x(t), z(t), t)dt (1)
t0
optimal highway design solution for connected autonomous
trucks by testing pavement failure and traffic performance The state equation is:
under different cross-sectional configurations. To improve ẋ = f (x(t), z(t), t) t ∈ [t0 , tf ] (2)
the ride comfort and safety of a traditional adaptive cruise
control system, Yao et al. [16] proposed a target vehicle selec- The boundary constrain is:
tion algorithm based on the prediction of the lane-changing ϕ(x(t0 ), t0 , x(tf ), tf ) = 0 (3)
intention of the preceding vehicle. Osman et al. [17] pro-
posed an approach that combined guidance navigation and The path constraint is:
image-based visual detection for road-following and over- C[x(t), z(t), t] ≤ 0 t ∈ [t0 , tf ] (4)
taking maneuvers. Zhang et al. [18] presented a novel
lane-changing assistance system that provides guidance for In Eqs. (1)–(4), the functions ψ, g, f , ϕ, and C are defined
drivers’ lane-changing behaviors. Lv et al. [19] presented a as follows:
novel planning approach for lane changing for autonomous ψ : Rn × R × Rn × R → R

115618 VOLUME 11, 2023

Y. Liu et al.: Optimal Lane Changing Problem of Vehicle Handling Inverse Dynamics

g : Rn × Rm × R → R ≤ i ≤ Nτ , 1 ≤ l ≤ q};
f : Rn × Rm × R → Rn Ū = {uil : τil ∈ Ḡ}.
ϕ : Rn × R × Rn × R → Rq
C : Rn × Rm × R → Rc IV. MESH REFINEMENT METHOD
The problems described by Eq. (1)–(4) are called the Bolza Because the state equation determines the change rule of
optimal control problems. the state variable with the control variable, the nonsmooth
property of the state variable is usually reflected in the control
III. DISCRETING OF COLLOCATION METHOD BASED ON variable (corresponding to the drastic change or even dis-
RUNGE-KUTTA SCHEME continuity of the control variable), and the control variable
First, the optimal control problem(Eq. (1)–(4)) is transformed is refined in this study. If necessary, the state and control
into a time interval τ ∈ [0, 1] via integral transformation τ = variables can also be refined simultaneously. For a group
(t − t0 )/(tf − t0 ). It is assumed that the nodes of N discrete of discrete nodes G = {τi : i = 0, 1, · · · , N } and the
intervals in the unit interval [0, 1] are: corresponding discrete optimal control variable φ = {ui :
i = 0, 1, · · · , N }, mesh refinement involves inserting new
G = {τi : τi ∈ [0, 1], i = 0, 1, · · · , N ; τ0 = 0, τN = τf discrete nodes or deleting redundant discrete nodes in some
= 1; τi < τi+1 , i = 0, 1 (5) areas according to the non smooth characteristics of the lane
change trajectory; thus, the optimization problem can be
where τi is called a node or mesh, and τi can be uniformly or
solved with high precision by using fewer discrete nodes.
nonuniformly distributed on [0,1].
It is set that xi = x(τi ), ui = u(τi ). Eq.(6) can be
obtained by adopting the discrete format of q-order Runge- A. NODE INSERTION METHOD
Kutta method in Eq.(2).
Without losing generality, taking τi as an example, the control
q
X variable at τi is solved with interpolation calculation utilizing
xi+1 = xi + 1t · hi βj fij , i = 0, 1, · · · , N − 1 (6) the peripheral nodes of τi . If the difference (i.e., interpolation
j=1 error) between the interpolation result û(τi ) and optimization
where 1t = tf − t0 , hi = τi+1 − τi , fij = results u(τi ) is large, which exceeds the predetermined error
f (xij , uij , τij ; t0 , tf ), xij , uij , τij are the intermediate variables. limit, it indicates that it is necessary to insert nodes around
q
τi . In contrast, if the interpolation error is small, it indicates
where xij = xi + 1t · hi αjl fil , 1 ≤ j ≤ q.
P
l=1 that the control variable at τi can be interpolated from the
where τij = τi + hi ρi , uij = u(τij ). And ρj , βj , αij are all surrounding nodes, the surrounding area of τi is smooth,
constants and satisfying 0 ≤ ρ1 ≤ · · · ≤ ρq ≤ 1. and no further refinement is required. For each node, the
When αij = 0(l ≥ j), Eq.(6) is explicit; otherwise, Eq. mesh was refined using the above process. If the optimization
Equation (6) is an implicit scheme. problem has multiple control variables, the grid around the
The objective function can be discretized using a similar node must be refined as long as the interpolation error of any
method. control variable component at a node exceeds the predeter-
The discrete nonlinear programming problem solves the mined error. Compared with multiresolution technology, the
variables X , U , Ū , t0 , and tf to minimize the following objec- above node insertion method has the following advantages.
tive function: First, this method is simple and easy to operate, and there is
N
X −1 q
X no limit to the number of nodes in the initial grid. Second,
J = ψ(x0 , t0 , xf , tf ) + 1t hi βi Lij (7) this method can continuously refine the grid in the area with
i=0 j=0 a large arbitrary interpolation error of the trajectory until the
accuracy requirements are met or the predetermined highest
and meet the following constraints:
resolution is reached, avoiding the omission or failure of
q
X refinement and a more complex inspection. Finally, when
ξi = xi+1 − xi − 1t · hi βj fij = 0 (8) calculating the interpolation error of a node, this method
j=1
selects the optimal interpolation node from all nodes around
Ci = C(xi , ui , τi ; t0 , t0 ) ≤ 0 (9) it. It can be seen that the method in this study makes use of
C̄ij = C(xij , uij , τij ; t0 , tf ) ≤ 0τij ∈ Ḡ (10) more information in the current grid, so the number of nodes
ϕ(x0 , t0 , xf , tf ) = 0 (11) to be inserted may be less.

where
B. NODE DELETION METHOD
Lij = L(xij , uij , τij ), i = 0, 1, · · · , Nτ − 1;
The node insertion method mentioned above can add redun-
X = {x0 , x1 , · · · , xNτ }; dant discrete nodes. If they can be deleted, it is obviously con-
U = {u0 , u1 , · · · , uNτ }; Ḡ = {τil ∈ [0, 1] : τil ∈
/ G, 0 ducive to reducing the number of optimization calculations.

VOLUME 11, 2023 115619

 Y. Liu et al.: Optimal Lane Changing Problem of Vehicle Handling Inverse Dynamics

The left and right slopes were defined as follows: (c) Increase i by 1. If i ≤ Nold − 1 returns to step 3 (a),
 u(τi ) − u(τi−1 ) otherwise proceed to the next step.
 u′l =
 4) It is set that i = 3, and then step 4(a)- 4(c) is executed.
τi − τi−1
u(τi+1 ) − u(τi ) (12) (a) The surroundings of nodes τi−1 ∈ Gold and τi+1 ∈ Gold
were refined. Then, the nodes τl0 = τl +τ τi +τr
′
2 and τ0r = 2 are
 ur =
 i
τi+1 − τi
added to Gint , where τl and τr are the left and right adjacent
Obviously, u′l = u′r means that the control variable near nodes of the node τi ∈ Gold in Gint .
node τi is constant. Therefore, τi is redundant, and can be (b) Increase i by 1. If i ≤ Nold − 2 returns to step 4 (a),
deleted. The aforementioned node deletion method can only otherwise proceed to the next step.
delete nodes whose left and right slopes are both zero. In fact, 5) Grid Gint is the new grid obtained by the node insertion
the slopes of redundant nodes are not necessarily zero, and the algorithm, and the corresponding function value is 8int . If the
above method is relatively conservative. function value of the new node is unknown, then the value
is calculated by p-order essentially non-oscillatory interpola-
C. MESH REFINEMENT METHOD tion utilizing Gold and 8old .
Based on the above node insertion and deletion algorithms, Node deletion algorithm:
an optimization mesh refinement algorithm can be designed. 1) Finding the nodes in the Gint that have reached the
Before mesh refinement, it is necessary to select the initial highest resolution, that is, the adjacent nodes with a distance
number of discrete intervals N , error limit of grid refinement less than 1/(221 N ), and recording them as Gmin .
ε and, highest resolution Jmax as well as maximum number of 2) Using Gcheck = {τki : τki ∈ Gint \(Gbase ∪ Gmin ), i =
iterations Imax (Imax ≥ Jmax /2 + 1). The RK discrete method 1, · · · , Ncheck } to represent the set of nodes belonging to Gint
was used to transform the lane-change optimization problem but not G0 or Gmin .
into an NLP. It is set that I = 1, Gold = G0 . In addition, the 3) It is set that i = 2, and then step 3(a)-3(c) is executed.
initial value of the NLP optimization variable was set as Xold . (a) The left slope u′l and right slope u′r of τki in Gint were
where G0 is N + 1 discrete nodes uniformly distributed in calculated.
the interval [0,1]. The mesh refinement process is as follows: (b) If u′l ≤ ε1 and u′r ≤ ε1 , τki ∈ Gcheck is deleted from
Step 1: NLP is solved with Xold as the initial value on the Gint ; otherwise, go to step 3(c), where ε1 is a small number
grid Gold . If I ≥ Imax occurs, the refinement is stopped; to measure whether the slope is zero.
otherwise, proceed to the next step. (c) By increasing i by 1, if i ≤ Ncheck , return to step 3(a);
Step 2: Node insertion algorithm. otherwise, the node deletion algorithm is stopped.
Step 3: Node deletion algorithm.
Step 4: Nonuniform grid Gnew = Gint obtained by this V. MATHEMATICAL MODEL OF VEHICLE LANE
refinement and the corresponding function value φnew = φint . CHANGING PROBLEM
Step 5: If the refined new mesh Gnew is the same as the A. MATHEMATICAL MODEL OF VEHICLE LANE
old mesh Gold , the refinement is ended. Otherwise, it is set CHANGING PROBLEM
I = I + 1, the NLP optimal solution solved in Step 1 is To describe the motion characteristics of the vehicle accu-
interpolated to Gnew , which is set as a new initial value Xold . rately, it is assumed that the vehicle is traveling on a
And then it is set Gnew = Gold , and return to Step 1, and horizontal road surface, and the vehicle center of mass serves
continue refining. as the coordinate origin of the vehicle coordinate system. The
Step 6: The node insertion algorithm and node deletion suspension was simplified as a rigid body during the vertical
algorithm are the keys to the process of the above mesh movement of the vehicle. The vehicle had no pitch or roll
refinement technology, which are given below. motion. The influence of longitudinal rolling resistance on
Node insertion algorithm: the parameters was neglected.
1) It is set 8old = {uk : τk ∈ Gold , k = 1, · · · , Nold } The nonlinear 7-DOF vehicle model shown in Fig. 1 is used
which is then rewritten as 8old = {φl (τk ) : τk ∈ Gold , to describe the vehicle lane-change problem.
l = 1, · · · , m}. In state space form it is [24]:

2) Initializing the grid Gint = Gold and function values  m(v̇x − rvy ) = (Fxfl + Fxfr ) cos δf + Fxrl + Fxrr
φint = φold .

−(Fyfl + Fyfr ) sin δf )




3) It is set that i = 2, and then step 3(a)- 3(c) is executed. 
m(v̇y + rvx ) = (Fxfl + Fxfr ) sin δf + Fyrl + Fyrr


(a) Ḡold = {τk : τk ∈ Gold , k ̸ = i} and 8̄old are



 +(Fyfl + Fyfr ) cos δf )

corresponding function values. The function value 8̄(τi ) at


τi ∈ Gold is calculated by p-order essentially non-oscillatory Iz ṙ = (Fxfl + Fxfr )lf sin δf + [(Fxfr − Fxfl ) cos δf (13)
interpolation utilizing Ḡold and 8̄old . B

w



 +(Fxrr − Fxrl )]
(b) If each component of the interpolation error di = 
 2
δf + (Fyfl − Fyfr )

8(τi ) − 8̄(τi ) at τi is less than the corresponding component +(F + F )l cos

yfl yfr f



of the error limit, Step 3(c) is executed. Otherwise, the four
 B w
sin δf − (Fyrl + Fyrr )lf



nodes shown in the equation are added to Gint . 2

115620 VOLUME 11, 2023

Y. Liu et al.: Optimal Lane Changing Problem of Vehicle Handling Inverse Dynamics

TABLE 1. Simulation parameters.

FIGURE 1. Nonlinear 7-DOF vehicle model.

( B. OPTIMAL CONTROL OBJECT OF LANE CHANGE

v̇x = vy r + ax MOTION PLANNING PROBLEM
(14)
v̇y = −vx r + ay The minimum-time problem of lane-change motion planning
can be regarded as the optimal control problem in control
where m is the vehicle mass, Iz is the moment of inertia around theory. Therefore, the minimum time performance indicator
the z-axis, vx and vy are the lateral and longitudinal speeds, is
respectively, r is the yaw rate of the vehicle, Fyi is the lateral Z tf
force of the front and rear tires (i = fl, fr, rl, rr), Fxi is the J= dt (18)
longitudinal force of the front and rear wheels, and δf is the t0
front steering angle. where t0 is the initial time, tf is the final time.
The lateral and longitudinal forces of the front and rear tires
can be expressed as [25]: C. CONTRAINS

Fyfl Fyfr
 
µcfl αf µcfr αf
 The initial and terminal states are described as:
= (15)
Fyrl Fyrr µcrl αr µcrr αr x(t0 ) = [vx (t0 ), 0, 0, 0, 0]T (19)
T
The angular acceleration of the rotational motion of four x(tf ) = [vx (tf ), 0, 0, x(tf ), y(tf )] (20)
wheels is: When a braking maneuver is applied to decelerate the
vehicle, the constraints onFxf and Fxr can be rewritten as

1

 ω̇fl = [−Re · Fxfl − Tbfl + Tdfl ] follows:
Iω



1 µmg(b + µhg ) 

 
 ω̇rl = [−Re · Fxrl − Tbrl + Tdrl ]


Iω Fxf ≥ − 
a+b

1 (16) (21)
 ω̇fl = [−Re · Fxfl − Tbfl + Tdfl ]
 a − µhg
Fxr = Fxf

Iω
 
b + µhg

 
1


 ω̇rr = [−Re · Fxrr − Tbrr + Tdrr ]


Iω VI. NUMERICAL SIMULATIONS AND EXPERIMENTAL
The state equation can be described as: VERIFICATION
A. NUMERICAL SIMULATIONS
ẋ = f [x(t), z(t)] (17) In this study, the solver SNOPT, developed based on the SQP
algorithm, was used to solve the NLP problem. When SNOPT
where x(t) and z(t) are the state and the input which are is used to solve the NLP, it is provided with the first-order par-
denoted respectively as tial derivative of the NLP, that is, the Jacobian matrix (defined
as the first-order partial derivative matrix of the objective
T
x(t) = vx (t), vy (t), r(t), x(t), y(t) , z(t) = [δ(t)]T .

function and all constraints on all independent variables),
which can significantly improve the computational efficiency
where vx and vy as well and δf are the states that can be and obtain the discrete optimal solution of the lane-changing
measured using sensors. yand r are different to be measured optimization simulation parameters are shown in Table 1.
because of expensive sensors.
Simultaneously, for the optimal lane-changing problem, 1) LEFT LANE CHANGING CONDITION
the state of lateral distance y is the desired state in which the Fig.2 shows the simulation results under the left-lane-
controller attempts to drag the system state into. changing condition.

VOLUME 11, 2023 115621

 Y. Liu et al.: Optimal Lane Changing Problem of Vehicle Handling Inverse Dynamics

lane. The steering angle of minimum time lane-changing

can provide guidance for drivers’ lane-changing behavior,
including under emergency avoidance conditions.
From Fig.2(c), it can be seen that there are peak values of
lateral acceleration at longitudinal distance of 16m and 38m.
And the maximal of the lateral acceleration is about 0.3 m/s2
which appears at the time of changing current lane to the adja-
cent left lane indicating that when the vehicle changes current
lane to the adjacent left lane and then keep driving in a straight
line in the adjacent lane there are no dangerous conditions
such as rollover controlled by the proposed method. However,
the maximum value of lateral acceleration does not reach the
threshold of the rollover, which indicates that the vehicle has
no risk of rollover during the lane-changing process.
From Fig.2(d), it can be seen that there are peak values
of the yaw rate at longitudinal distance of 16m and 38m.
This indicates that when the vehicle changes the current
lane to the adjacent left lane and then continues driving in
a straight line in the adjacent lane, there is a larger yaw rate.
Therefore, considering the control input and response of the
vehicle, starting the lane-changing process at a lower speed
is beneficial for the safe driving of the vehicle.

2) RIGHT LANE CHANGING CONDITION

Fig.3 shows the simulation results under the left-lane-
changing condition.
From Fig.3(a), it can be seen that the vehicle can complete
the maneuver of emergency lane changing with minimum
time successfully controlled by the proposed method, which
can provide safety references during lane changing for
drivers.
From Fig.3(b), it can be seen that there are peak values
of the steering angle at longitudinal distance of 13m and
36m. This is because the vehicle must address larger steering
angles to change the current lane to the adjacent right lane
and then continue driving in a straight line in the adjacent
lane. The steering angle of minimum time lane-changing
can provide guidance for drivers’ lane-changing behavior,
including under emergency avoidance conditions.
From Fig.3(c), it can be seen that there are peak values of
lateral acceleration at longitudinal distance of 16m and 38m.
FIGURE 2. Simulation results under left lane changing condition. And the maximal of the lateral acceleration is about 0.3 m/s2
which appears at the time of changing current lane to the
adjacent right lane indicating that when the vehicle changes
current lane to the adjacent right lane and then keep driving
From Fig.2(a), it can be seen that the vehicle can com- in a straight line in the adjacent lane there are no danger-
plete the maneuver of emergency lane-changing with mini- ous conditions such as rollover controlled by the proposed
mum time successfully controlled by the proposed method. method. To ensure the safety of the lane-change process, the
To ensure that the vehicle moves from the initial position to driver should control the initial speed to control the lateral
the terminal position in the minimum time, the driver needs acceleration.
to start lane-changing immediately. From Fig.3(d), it can be seen that there are peak values
From Fig.2(b), it can be seen that there are peak values of the yaw rate at longitudinal distance of 16m and 38m.
of the steering angle at longitudinal distance of 13m and This indicates that when the vehicle changes its current lane
36m. This is because the vehicle must address larger steering to the adjacent right lane and then continues to drive in a
angles to change the current lane to the adjacent left lane straight line in the adjacent lane, there is a larger yaw rate.
and then continue driving in a straight line in the adjacent This indicates that when changing the current lane to the

115622 VOLUME 11, 2023

Y. Liu et al.: Optimal Lane Changing Problem of Vehicle Handling Inverse Dynamics

TABLE 2. Calculation time.

both the ends. However, once the number of nodes is given,

the distribution of the nodes is also determined. Therefore,
the discrete nodes of the GPM method cannot be randomly
arranged, and they are not as flexible as those of the local
collocation method in mesh refinement. At the same time,
the mesh refinement method of GPM usually optimizes the
problem by adding nodes and uses the characteristics of
dense discrete nodes near nodes to achieve mesh refine-
ment. Therefore, the solution accuracy of MRM was higher
than that of GPM, indicating the superiority of the proposed
method.

C. EFFICIENCY VERIFICATION
Table 2 presents the comparison results of the calculation
time. The calculation time required by the algorithm in this
study is less than that of the GPM, which has greater advan-
tages in terms of calculation efficiency than the GPM.

D. EXPERIMENTAL VERIFICATION
A virtual test using the CarSim software was conducted to
verify the feasibility of the simulated results.
CarSim has become standard software in the automo-
bile industry. The Carsim algorithm has good prediction
performance and high reliability for the relevant dynamic
performance of vehicles and different working conditions.
CarSim is widely used in the global market because of its
high simulation accuracy, fast operation speed, convenient
use, effective stability and reliability, standardization, and
FIGURE 3. Simulation results under right lane changing condition. scalability.
A vehicle is a complex, nonlinear system. At present, most
studies widely use simplified or local modeling methods to
build vehicle models, but it is difficult to accurately describe
adjacent lane, the vehicle will generate a larger yaw rate, the nonlinear characteristics of the vehicle and to reflect the
which indicates that there is a risk of the vehicle slipping, real running state of the vehicle. Therefore, to make the simu-
and the driver should control the vehicle well at this point. lation closer to the real running state of the vehicle, the mature
commercial software CarSim was used to build the dynamic
B. ACCURACY VERIFICATION model of the vehicle. Using Carsim to establish the dynamic
Comparison is established in order to verify the accuracy of model of the entire vehicle can ensure the accuracy of the
the proposed. controlled object and can more accurately verify the effec-
From Fig. 4, it can be seen that the steering angle error of tiveness of the horizontal and vertical control strategies of
the MRM is smaller than that of the GPM. This is because the unmanned vehicle. CarSim is a special simulation soft-
the discrete nodes of the GPM method are the roots of the ware for the dynamics of small- and medium-sized vehicles.
orthogonal polynomials. The distribution characteristics in In Carsim, researchers can select different models and driving
the discrete interval were sparse in the middle and dense at conditions and select and adjust the parameters of various

VOLUME 11, 2023 115623

 Y. Liu et al.: Optimal Lane Changing Problem of Vehicle Handling Inverse Dynamics

FIGURE 4. Comparison results between MRM and GPM.

parts of the vehicle. CarSim is widely used in the research simulation of Carsim and Simulink, making the design of
and development of automobile control systems because of lateral and longitudinal controllers easier.
its flexible and convenient simulation characteristics. Using Fig. 5 shows the experimental results for the lateral accel-
Carsim to model the dynamics of the entire vehicle can not eration, steering angle, and yaw rate.
only avoid the complicated and tedious dynamic modeling From Fig. 5, it can be seen that for the lateral acceleration,
process but also make the simulation closer to the actual steering angle, and yaw rate, there are errors between the
running state of the vehicle. At the same time, to facilitate simulation and actual test values. This is because the model
the design of control strategies for researchers, Carsim can in the virtual test ignored the nonlinearity of the steering
use ‘Send to Simulink’ to call Simulink to realize the joint and suspension systems. However, the trends of the curves

115624 VOLUME 11, 2023

Y. Liu et al.: Optimal Lane Changing Problem of Vehicle Handling Inverse Dynamics

in this study will be effectively expanded. This study can

help drivers easily identify safe lane-changing trajectories
and areas under collision avoidance conditions. Additionally,
a driver assistance system that can provide practical guidance
for lane-changing handling inputs can be constructed.

ACKNOWLEDGMENT
Yingjie Liu gratefully acknowledges the support agency.

REFERENCES
[1] Y. Liu and D. Cui, ‘‘Vehicle state and parameter estimation based on
double cubature Kalman filter algorithm,’’ J. Vibroeng., vol. 24, no. 5,
pp. 936–951, Aug. 2022.
[2] Y. Liu and D. Cui, ‘‘Vehicle state estimation based on adaptive fading
unscented Kalman filter,’’ Math. Problems Eng., vol. 2022, pp. 1–11,
Apr. 2022.
[3] H. Li and S. Zhang, ‘‘Lane change behavior with uncertainty and
fuzziness for human driving vehicles and its simulation in mixed
traffic,’’ Phys. A, Stat. Mech. Appl., vol. 606, pp. 128130–128141,
Nov. 2022.
[4] S. He, F. Ding, C. Lu, and Y. Qi, ‘‘Impact of connected and autonomous
vehicle dedicated lane on the freeway traffic efficiency,’’ Eur. Transp. Res.
Rev., vol. 14, no. 1, pp. 12–26, Dec. 2022.
[5] S. H. Wang, M. Zhao, D. H. Sun, and X. Liu, ‘‘A merging strategy based on
optimal control of main-lane downstream and on-ramp vehicles,’’ KSCE J.
Civil Eng., vol. 26, no. 11, pp. 4777–4792, Sep. 2022.
[6] A. Falsone, B. Melani, and M. Prandini, ‘‘Lane change in automated
driving: An explicit coordination strategy,’’ IEEE Control Syst. Lett., vol. 7,
pp. 205–210, 2023.
[7] H. Wang, W. Cui, and B. Ye, ‘‘Lane departure assistance coordinated con-
trol system for in-wheel motor drive vehicles based on dynamic extension
boundary decision,’’ J. Syst. Sci. Complex., vol. 33, no. 4, pp. 1040–1063,
Aug. 2020.
[8] J. Hu, S. Xiong, J. Zha, and C. Fu, ‘‘Lane detection and trajectory tracking
control of autonomous vehicle based on model predictive control,’’ Int. J.
Automot. Technol., vol. 21, no. 2, pp. 285–295, Apr. 2020.
[9] Y. Chen, H. Yu, J. Zhang, and D. Cao, ‘‘Lane-exchanging driving strategy
FIGURE 5. Experimental results of the lateral acceleration and the for autonomous vehicle via trajectory prediction and model predictive
steering angle as well as the yaw rate. control,’’ Chin. J. Mech. Eng., vol. 35, pp. 71–83, Dec. 2022.
[10] Z. Fei, S. Jia, and Q. Li, ‘‘Low-cost and high-precision lane-level vehi-
cle positioning method in an urban area based on traversal optimization
for MIMU calibration parameters,’’ Arabian J. Sci. Eng., vol. 47, no. 8,
in Figs. 5(a)–(c) are sufficient to illustrate the correctness of pp. 9877–9886, Jan. 2022.
the proposed method. [11] W. Zhou, D. Chen, J. Yan, Z. Li, H. Yin, and W. Ge, ‘‘Multi-agent
reinforcement learning for cooperative lane changing of connected and
autonomous vehicles in mixed traffic,’’ Auto. Intell. Syst., vol. 2, no. 1,
VII. CONCLUSION pp. 5–25, Dec. 2022.
This study proposes an adaptive mesh refinement method for [12] Z. Wang, D. Tan, G. Ge, and S. Liu, ‘‘Optimal trajectory planning and
control for automatic lane change of in wheel motor driving vehicles
an optimal lane-changing problem based on a 7-DOF vehicle on snow and ice roads,’’ Autom. Control Comput. Sci., vol. 54, no. 5,
model. This method uses the local collocation method to pp. 432–445, Sep. 2020.
transform the vehicle’s optimal lane-changing problem into [13] C. Park, N. T. Jeong, D. Yu, and S. H. Hwang, ‘‘Path generation algorithm
an NLP problem. A simulation of vehicle emergency left- based on crash point prediction for lane changing of autonomous vehicles,’’
Int. J. Automot. Technol., vol. 20, no. 3, pp. 507–519, 2019.
and right-lane changes was established to verify the proposed [14] G. Li, Q. Wang, and C. Zuo, ‘‘Emergency lane vehicle detection and
method. The simulation results showed that the grid gener- classification method based on logistic regression and a deep convolu-
ated by the method in this study was smaller and required tional network,’’ Neural Comput. Appl., vol. 34, no. 15, pp. 12517–12526,
Aug. 2022.
fewer grid iterations. It can quickly and accurately solve a [15] H. Jehanfo, S. Hu, I. Kaparias, J. Preston, F. Zhou, and A. Stevens,
vehicle’s optimal lane-changing problem with constant com- ‘‘Redesigning highway infrastructure systems for connected autonomous
ponents of the optimal control variables. truck lanes,’’ J. Transp. Eng., A, Syst., vol. 148, no. 12, Dec. 2022,
Art. no. 04022104.
For a problem in which the optimal solution does not con-
[16] J. Yao, G. Chen, and Z. Gao, ‘‘Target vehicle selection algorithm
tain constant components, the inability to delete redundant for adaptive cruise control based on lane-changing intention of pre-
nodes may result in a large number of nodes. Of course, ceding vehicle,’’ Chin. J. Mech. Eng., vol. 34, no. 1, pp. 135–153,
this defect is caused by the limitations of node-deletion algo- Dec. 2021.
[17] K. Osman, J. Ghommam, and M. Saad, ‘‘Guidance based lane-changing
rithms. If efficient node deletion algorithms can be designed control in high-speed vehicle for the overtaking maneuver,’’ J. Intell.
in the future, the applicability of the node insertion algorithm Robotic Syst., vol. 98, nos. 3–4, pp. 643–665, Jun. 2020.

VOLUME 11, 2023 115625

 Y. Liu et al.: Optimal Lane Changing Problem of Vehicle Handling Inverse Dynamics

[18] X. Zhang, Y. Zhao, W. Zhang, F. Lin, and H. Li, ‘‘Minimum time lane DAWEI CUI received the Ph.D. degree from
changing problem of vehicle handling inverse dynamics considering the the Material Science and Engineering Institute,
driver’s intention,’’ Int. J. Automot. Technol., vol. 20, no. 1, pp. 109–118, University of Science and Technology Beijing,
Feb. 2019. Beijing, China, in 2008. He is currently with the
[19] X. Lv, W. Li, and J. Wang, ‘‘Safety-field-based path planning algorithm School of Machinery and Automation, Weifang
of lane changing for autonomous vehicles,’’ Int. J. Control, Autom. Syst., University, Weifang, China. His current research
vol. 20, no. 2, pp. 564–576, Feb. 2022. interest includes control and vehicle system
[20] R. Muthalagu, A. Bolimera, and V. Kalaichelvi, ‘‘Vehicle lane markings
dynamics.
segmentation and keypoint determination using deep convolutional neural
networks,’’ Multimedia Tools Appl., vol. 80, no. 7, pp. 11201–11215,
Jan. 2021.
[21] W. Hu, E. Jiaqiang, E. Leng, F. Zhang, J. Chen, and Y. Ma, ‘‘Investigation
on harvesting characteristics of convective wind energy from vehicle driv-
ing on multi-lane highway,’’ Energy, vol. 263, Jan. 2023, Art. no. 126062.
[22] G. Cheng, Q. Sun, and Y. Bie, ‘‘Mandatory lane-changing modelling based
on a game theoretic approach in traditional and connected environments,’’
Transp. Saf. Environ., vol. 5, no. 1, pp. 1–7, Jan. 2023.
[23] Y. Liu and Y. Liu, ‘‘Research on lane change motion planning steering
input based on optimal control theory,’’ Math. Problems Eng., vol. 2022,
pp. 1–12, Jun. 2022.
[24] Y. Li, J. Fan, Y. Liu, and X. Wang, ‘‘Path planning and path tracking
for autonomous vehicle based on MPC with adaptive dual-horizon-
parameters,’’ Int. J. Automot. Technol., vol. 23, no. 5, pp. 1239–1253,
Oct. 2022.
[25] X. He, Y. Liu, C. Lv, X. Ji, and Y. Liu, ‘‘Emergency steering control
of autonomous vehicle for collision avoidance and stabilisation,’’ Vehicle
Syst. Dyn., vol. 57, no. 8, pp. 1163–1187, Nov. 2018.

YINGJIE LIU received Ph.D. degree from the WEN PENG received the Ph.D. degree from the
College of Energy and Power Engineering, Nan- State Key Laboratory of Rolling and Automa-
jing University of Aeronautics and Astronautics, tion, Northeastern University, Shenyang, China,
Nanjing, China, in 2014. He is currently with the in 2014. He is currently with the State Key Lab-
School of Machinery and Automation, Weifang oratory of Rolling and Automation, Northeastern
University, Weifang, China. His current research University. His current research interests include
interests include vehicle system dynamics and the intelligent control of strip rolling and the devel-
control theory to ground vehicles. opment of continuous rolling automation control
systems.

115626 VOLUME 11, 2023