International Journal of Automotive Technology, Vol. 24, No. 4, pp.

969-981 (2023) Copyright © 2023 KSAE/ 134–04

DOI 10.1007/s12239–023–0079–5 pISSN 1229–9138/ eISSN 1976–3832

EMERGENCY COLLISION AVOIDANCE METHOD WITH MOVING

OBSTACLES BASED ON PIECEWISE POLYNOMIAL

Wen He1,2), Yong Chen1)*, Gang Liu3), Yitao Wu3) and Yonggang Liu3)
1)
School of Automation Engineering, University of Electronic Science and Technology of China,
Chengdu, Sichuan 611731, China
2)
Changan Automobile Intelligent Research Institute, Jiangbei, Chongqing 400021, China
3)
State Key Laboratory of Mechanical Transmission & College of Mechanical and Vehicle Engineering,
Chongqing University, Chongqing 400000, China

(Received 14 June 2022; Revised 10 December 2022; Accepted 14 January 2023)

ABSTRACT–As an important part of the advanced driver assistance system (ADAS), the active emergency collision
avoidance system can effectively reduce the incidence of accidents. Previously, static obstacle avoidance has been realized by
employing piecewise polynomial path planning to improve vehicle stability. However, regarding moving obstacles, the
collision avoidance problem becomes more intractable. To address the problem, a novel method based on the technical
framework of piecewise polynomial path planning is proposed in this paper to conduct collision avoidance with moving
obstacles. First, the Multivariate Gaussian Distribution model is adopted to estimate the velocity of the obstacle ahead, and
the longitudinal motion trajectory of the obstacle is calculated. Then, the longitudinal length of the first target path is selected
by the rule-based method. Moreover, yielding from the method for static obstacles, the collision avoidance algorithm for
moving obstacles is proposed. Finally, the co-simulations are carried out in Simulink/CarSim, of which the results prove that
the proposed method can avoid the moving obstacle effectively.

KEY WORDS : Collision avoidance, Path planning, Piecewise polynomial method, Advanced driver assistance system
(ADAS)

NOMENCLATURE automated vehicles, how to ensure safety in emergency

conditions will be highlighted. Therefore, the collision
φ : yaw angle, rad avoidance system designed to prevent or reduce accidents
y* : safe distance, m has attracted the wide attention of researchers and
K : curvature of the path, m-1 companies (Cui et al., 2019).
x, y : position coordinate, m In recent years, automatic emergency collision avoidance
lfs, d : length and width of the vehicle, m systems can be classified into automatic emergency braking
d : width of vehicle, m (AEB) and automatic emergency steering (AES), in which
Lob, dob : length and width of the vehicle, m the AEB system outputs the appropriate deceleration to stop
Q, P : coefficient vector the vehicle before a collision happens while AES devotes to
a : acceleration, m/s2 change the vehicle latitudinal motion trajectory to avoid a
v : speed, m/s collision, which has been equipped on most vehicles
ξ : state
(Cicchino, 2017; Kappes et al., 2014; Wilson et al., 2015).
Moreover, AEB can effectively assist unresponsive drivers
1. INTRODUCTION to improve the success rate of collision, thus avoiding or
reducing damage in emergencies. However, when the ego
Automated driving is considered one of the promising vehicle moving at a high speed faces short-distance
technologies to reduce traffic accidents, alleviate congestion obstacles, the collision avoidance success rate of the AEB
and energy consumption, which can further improve road system will decrease, limited by the longitudinal dynamic
traffic efficiency in the future (Fagnant and Kockelman, characteristics.
2015; Yu et al., 2020). With the commercial application of To overcome the limitations of the AEB system,
lane-changing is introduced in the AES system to avoid
collisions when the main vehicle approaches obstacles
*Corresponding author. e-mail: [email protected] ahead. The artificial potential field (APF) method (Ren et

969
970 Wen He et al.

al., 2019; González et al., 2015) and the discrete predict the speed of moving objects, a modified version of
optimization method (Petrov and Nashashibi, 2014; Ji et al., this method will be introduced in this paper to predict the
2016) have been successfully applied in vehicle trajectory movement of obstacles. Finally, the emergency collision
control. However, the calculation of a risk field and the avoidance method of static obstacles will be extended and
collision judgment entails a heavy computational burden, applied to moving obstacle collision avoidance. The main
which limits its real-time implementation in emergencies. contributions of this paper are as follows:
To alleviate the impact, path planning approaches based on (1) The piecewise path planning method is optimized. By
discrete optimization are proposed in (Montemerlo et al., adding the lateral position sampling at the end of the second
2008; Werling et al., 2010; Chu et al., 2012). The planner path planning, a smoother target collision avoidance path
defines the lateral offset as the perpendicular distance to a can be found.
fixed baseline, enabling a vehicle to travel a road parallel to (2) A trajectory planning method for moving obstacles is
the baseline. To select the optimal path, running over proposed. By the Multivariate Gaussian Distribution model
obstacles and deviating from the road center are penalized estimating the future speed of the obstacle, its longitudinal
in the cost function (Phung et al., 2017). On a computer moving position is determined in the future. Based on
with an Intel Core i5-450M CPU (dual-core, 2.4 GHz) and knowing the moving of obstacles, the appropriate collision
4 GB RAM, the time cost of the discrete optimization path avoidance points, the key reference of piecewise
planner with 21 path candidates is 57 ms for one planning polynomial path planning, can be selected through the
step (Hu et al., 2018). However, moving obstacles are not proposed rules.
considered in these methods. The rest of this paper is organized as follows. In Section
We propose a collision-avoidance path planning method II, the segmented path planning method, the stability speed
for lane-changing that combines the advantages of the APF planning method, the obstacle moving prediction and the
method and the discrete optimization method (Yu et al., moving obstacle avoidance framework are presented.
2021). In that study, based on the idea of the APF, the Simulation and result analysis are illustrated in Section III,
number of candidate path generating will be reduced under followed by Section IV which concludes the whole paper.
the constraint of corner collision, which effectively
improves the computational efficiency. 2. MOTION TRAJECTORY PLANNING
However, difficulties still exist in dealing with similar
problems with moving obstacles. To address it, accurate In this section, a motion planning method is proposed for
motion velocity and direction prediction should be emergency conditions. First, the path planning method that
introduced in this stage. Generally, two main approaches, ensures driving safety for the autonomous vehicle is
namely the deterministic prediction and the stochastic leveraged, and then the vehicle speed planning is conducted
prediction, are commonly employed to predict the motion to guarantee the motion stability. Different from the general
state of obstacles ahead. Deterministic prediction motion planning method, the piecewise polynomial
algorithms such as mean prediction (Li et al., 2020): linear pro-posed in the path planning makes full use of the traffic
regression (Kim et al., 2019) method can predict the motion environment space to improve the maximum allowable
state in a simpler way, however poor at nonlinear fitting speed in response to emergency conditions. Additionally, in
ability. Additionally, the time-varying impact of data the speed planning, the phase diagram of vehicle motion
sampled at different time points cannot be considered in stability is employed to replace the anti-sideslip constraint,
deterministic prediction. Therefore, this method generally which allows the full performance of the vehicle.
cannot provide satisfactory prediction accuracy. Stochastic
prediction is to establish a probability model based on 2.1. Path Planning with Piecewise Polynomial
training samples, in which the maximum posterior value is The piecewise polynomial method is adopted in emergency
utilized as the prediction state (Witt et al., 2019). Stochastic collision avoidance path planning to obtain the target path
prediction not only estimates the state of obstacles but by stages and objectives. A piecewise polynomial method
provides the probability confidence interval that that divided the path into two stages, namely the first stage
comprehensively describes and evaluates the prediction and second stage, is applied in path planning, in which the
effect. In this method, the motion of the obstacles is first stage path devotes to avoiding collision with obstacles,
supposed to conform to the Multivariate Gaussian while the path in the second stage is designed as smooth as
Distribution in the time-continuous domain (Ngo et al., possible to stabilize the movement of the vehicle. Aiming at
2013), and the correlation of the motion state at different solving the problem that the accuracy of the geometric-based
times can be determined by a covariance function. Thus, the path tracking method decreases when the path curvature
longitudinal velocity of obstacle vehicle in a certain time changes suddenly, a quartic polynomial and a quintic
domain can be predicted by learning the historical data of polynomial are applied in the first and the second stage of
obstacle vehicles. the target paths, respectively to ensure the curvature
Since the stochastic prediction method can accurately continuity of the planned path.
 EMERGENCY COLLISION AVOIDANCE METHOD WITH MOVING OBSTACLES BASED ON 971

This subsection draws on the experience of the risk Given the initial position (xi, yi) and the destination point
potential field method which finds the non-collision path by (xe, ye) of the vehicle, the coefficients of the quartic
using the risk repulsion field. However, the obstacles are polynomial can be calculated by introducing the initial
often represented as ellipses to describe the repulsion field, curvature of the path to ensure smooth steering,
which will lead to the expansion of the outline of the
obstacles. Therefore, more conservative path planning gives Q  A f 1 R f (3)
less sufficient use of the traffic environment space, which
conflicts with the goal of emergency collision avoidance. where Q is the coefficient vector of the quartic polynomial
Thus, a geometry method considering the idea of risk Q T  q1 , q2 , q3 , q4 , q5  . A and R are the quartic
f f
potential field is introduced to find the critical collision
point. polynomial transfer matrix and the constraint vector, which
As shown in Figure 1, the green solid line denotes the can be described as:
first stage of path planning, and the yellow dashed lines are
the path candidates in the second stage. It indicates that the  xi4 xi3 xi2 xi1 1 
 4 
target path of the first state is determined once the vehicle  xe xe3 xe2 xe1 1 
destination and posture in the near future are known in A f   4 xi3 3 xi2 2 xi1 1 0 (4)
advance. The destination of the first stage can be obtained  3 
 4 xe 3 xe2 2 xe1 1 0
from the critical point of collision avoidance, as:
12 x 2 6 xi1 2 0 0 
 i
 d
 xe  xc  l fs cos e  2 sin e
(1) RTf   yi , ye , tan i , tan  e , K i  (5)

 y  y  l sin   d cos 
 e c fs e
2
e
where φi and φe are the initial and final yaw angle of the
vehicle; Ki is the initial curvature of the first stage path.
Where lfs and d are the length and width of the ego Here, φe remains to be defined by the optimization method
vehicle; xe and ye are the first stage final position of the ego in the following content.
vehicle; φe is the final yaw angle of the vehicle at the first Once the φe is obtained, the path in the first stage is clear.
stage; xc and yc are the pre-mentioned critical point of Then, the initial constraints of the second path can be
collision avoidance which is determined by the position of obtained which are equal to the terminal position and
the obstacle. posture of the first stage.
As for the second stage, the coefficient vector
 Lob
 xc  xob  2 P T   p1 , p2 , p3 , p4 , p5 , p6  of the quintic polynomial can
 (2) be obtained according to the starting and ending positions
 y  y  d ob  y * of the vehicle, denote by ViT   xe , ye ,  e , K e  and

c ob
2
VeT   x f , y f ,  f , K f  respectively.
where Lob and dob are the length and width of the obstacle;
y* is the safe distance to avoid a collision, which is set as
0.2 m in this paper.  xe5 xe4 xe3 xe2 x1e 1
 5 
 xf x 4f x 3
f x 2
f x1
f 1
 5x4 4 xe3 3x 2
2x 1
1 0
As   e4 e e
 (6)
 5x f 4 x3f 3x 2
f 2x 1
f 1 0
 20 x3 12 x 2 6x 1
2 0 0
 e
3
e
2 1
e

 20 x f 12 x f 6x f 2 0 0 

RsT   ye , y f , tan e , tan  f , K e , K f  (7)

P  As1 Rs (8)

where, xf, yf, φf and Kf are the final position, posture, and
Figure 1. Schematic diagram of emergency collision curvature of the vehicle at the second stage; φf and Kf are 0,
avoidance path planning. in this case, representing that the vehicle will drive along
972 Wen He et al.

with the road. The vehicle stability index which presents the stability of
The determination of unknown parameters, i.e., φe, xf and the vehicle judged by the    phase diagram is shown
yf, can be treated as an optimization problem. The object of in Figure 2. The red points correspond to the instability
the optimization is to find the smoothest path that ensures cases caused by following the target path and speed, while
vehicle stability, described as follows: the blue points represent the stable cases. To this end, a
support vector machine (SVM) is trained to identify the
min J  K e , x f , y f  vehicle stability based on the training data set.
  For nonlinear separable feature sets, the separation
 12q1 x12  6q2 x1  2q3 
 3
,  hyperplane with maximum separation is as follows:
 (1  (4q1 x13  3q2 x12  2q3 x1  p4 ) 2 ) 2 
K (e , x f , y f )  max  
  wT  ( x )  b  0 (10)
 20 p1 x2 2  12 p2 x2 2  6 p3 x2  2 p4 
 4 3 2 2 2
3

 (1  (5 p1 x2  4 p2 x2  3 p3 x2  2 p4 x2  p5 ) )  where, w and b are the parameters of the maximum
hyperplane. Φ(x) is the feature space conversion function. x
 xi  x1  xe is the distinguishing feature vector.

 xe  x2  x f The classification decision function f(x) of the separation

s.t. 0  y1  ye (9) hyperplane of the optimal design is:
 y  y  5.25
 e
f ( x)  sign  wT ( x)  b 
f

 tan 1  y ( x  x )      / 3 (11)
 e e i e .

The lateral displacement of the ego vehicle and the After supervised learning, the SVM model is trained as a
heading value at the end of the first segment path are stability classifier that can predict future vehicle stability
constrained to reduce the generation of candidate paths. In according to the road curvature, longitudinal acceleration
the constraints, the vehicle position of the target path is and speed of the vehicle,
designed to change monotonically. Additionally, the upper
and lower limits of the heading at the end of the first 0, instable
sta  ax , vx , K    (12)
segment path are determined by the initial relative position  1, stable
between the ego vehicle and the obstacle and the empirical
value. Then the stability classifier is established and used to
evaluate the transition function of the different states in the
2.2. Speed Planning for Vehicle Stability dynamic program. To this end, the problem of vehicle
After the target collision avoidance path is obtained, speed stable speed planning can be transformed into a multi-stage
planning should be leveraged and the vehicle stability phase decision-making process, and the dynamic programming
diagram is adopted to replace the anti-sideslip constraint to (DP) method is employed to solve the optimization of the
exert the full performance of the vehicle and further ensure stable speed curve with discretized vehicle state grids.
the vehicle stability. The vehicle state includes vehicle position and speed,
Hence, the D-type vehicle model in CarSim software is which is represented by ξ = [s, v] with the initial state ξ0 =
employed to obtain the vehicle responses under different [s0, v0]. The speed planning process is discretized into
inputs. The vehicle stability labels determined by the several certain stages by a fixed distance interval of Δs. In
   phase diagram (Liang et al., 2020) are recorded dynamic programming, the optimal state curves along the
with the original data set.

(a) (b)
Figure 2. Training dataset of the vehicle stability index. Figure 3. Principle of dynamic programming algorithm.
 EMERGENCY COLLISION AVOIDANCE METHOD WITH MOVING OBSTACLES BASED ON 973

distance interval are updated by selecting different control Assuming that the current time is xn, the longitudinal
inputs, namely the vehicle longitudinal acceleration, u = [a] speed of the obstacle vehicle in the time domain
 
that satisfy the vehicle driving safety constraints until the x   x1 , x2 ,  , xn  is y   y1 , y2 , , yn  . Then the
end position sk is reached, as shown in Figure 3. prediction longitudinal speed in the future time domain
Based on the vehicle kinematics model, the equation of  
state is: x*   x1* , x2* ,, xm*  is y*   y1* , y2* , , ym*  . Then the
joint probability distribution of the observation state and
k 1  T k , uk 1    sk  s, vk2  2ak 1s  (13) prediction state is
 

y    y   K K*T  
Let i and j be the indexes of speed and control  *  ~ N   *  , 
  (18)
 y     y   K* K **  
respectively, then the state of the vehicle at sk position is
expressed as  k(i , j )   sk , vki  , and the control at the state
  where K, K*, K** and K*T are multivariate covariance
 k( i , j ) is u k( i, 1j ) . Then the stability index of the vehicle in matrixes that are composed of Radial Basis Kernel
the k+1 stage can be expressed as, Function:

xx   2

L  k(i , j ) , uk(i, 1j )   sta  uk(i, 1j ) , vki , K ( s )   uk(i, 1j ) (14) 

k  x, x     2f e
2
2l
(19)
where sta  uk(i, 1j ) , vki , K ( s )  is applied to predict the where
stability of the vehicle at sk. K(s) denotes the curvature at sk.
uk( i, 1j ) is supplemented to avoid drastic vehicle speed  k  x1 , x1  k  x1 , x2   k  x1 , xn  
 
k  x2 , x1  k  x2 , x2   k  x2 , xn  
change. K 
The total stability index of the vehicle from s0 to sk is      
 
expressed as:  k  xn , x1  k  xn , x2   k  xn , xn   nn

k 1
J  sk , vk    L k(i , j ) , uk(i,1j )  (15)  k  x1* , x1  k  x1* , x2   k  x1* , xn  
k 0  
 k  x2* , x1  k  x2* , x2   k  x2* , xn  
K*   
And the target function can be written as:      
 
 k  xm , x1  k  xm , x2   k  xm* , xn  
* *
k 1
J *  sk , vk   min
u* =u1 ,u2 ,...,uk 
 L 
k 0
(i , j )
k , uk(i,1j )  (16) m n

 k  x1* , x1*  k  x1* , x2*   k  x1* , xm*  

According to Bellman's optimality principle, the  
 k  x2* , x1*  k  x2* , x2*   k  x2* , xm*  
optimization problem can be transformed into a series of K **    .
suboptimal problems in stages, and the recursive equation      
of dynamic programming is expressed as:  
 k  xm , x1  k  xm , x2   k  xm , xm  
* * * * * *
m m

J *  sk 1 , vkm1   min
(i, j )
uk 1

L  k( i , j ) , uk(i, 1j )   J *  sk , vki   (17) The conditional probability formula is expressed as,
 
  p y* , y  
where J ( sk 1 , v ) denotes the is the minimum stability
* m
k 1

index of the vehicle from the initial state ξ0 to ξk. The


p y∣*
y   
p( y )
(20)

stability index of the initial state is defined as J*(s0, v0) = 0.

Based on (20), the distribution of the prediction state
 
* can be calculated from the observation state
2.3. Motion Prediction for Obstacles y y:
To predict the longitudinal velocity of obstacle vehicles, a   
Multivariate Gaussian Distribution model is applied in this y∣*
y ~ N  K* K 1 y, K**  K* K 1 K*T  (21)
section to conduct prediction.
974 Wen He et al.

Figure 4. Schematic diagram of moving obstacle avoidance. Figure 5. Obstacle longitudinal moving trajectory.

In Figure 4, the blue scatters and the blue curve are the corresponding moment.
actual speed sampling points in the first 2 s and the speed The diagram block of motion trajectory planning for the
change curve within 5 s of the moving obstacle. The black obstacle is generated in Figure 6. The input information is
curves represent the upper and lower boundaries of the the moving trajectory of the obstacle. To obtain the
confidence interval of the front vehicle speed predicted by smoothest path, the longitudinal space of the first stage path
the Multivariate Gaussian Distribution Model. The red is selected as long as possible. Therefore, the longitudinal
curve represents the highest probability model-estimated position of the fifth-second moving obstacle is set to initial
speed. However, the lower edge of the confidence interval dp. Adding the initial relative longitudinal distance of the
is extracted as the future longitudinal speed according to the ego vehicle and front vehicle d eo0 , the longitudinal final
conservatism principle. As the Multivariate Gaussian position of the first stage path d plan1 can be constrained.
Distribution Model estimates the speed by adopting the
Gaussian function, the prediction curve is not monotonous, Then, the avoidance path of the two stages is planned by
which is inconsistent with reality. Therefore, a monotonicity the piecewise polynomial path planning method. The target
correction is made for the lower edge of the confidence motion trajectory is obtained after the planned speed
interval. The green curve is the predicted speed curve of satisfies the stability criterion and longitudinal avoidance
moving obstacles after the correction. The correcting rule criterion.
is:
tegod  tobsd (23)
plan1 plan1

 vx (k  1)  vx (k ) vx (k  1)  vx (k ) (22)
 where tegod plan1 and tobsd plan1 denote the moment of the ego
vx (k  1)  vx (k  1) vx (k  1)  vx (k )
vehicle and the obstacle reach the first target path
where vx is the speed observed value of the front vehicle; longitudinal distance.
vx is the predicted speed; k is the k-th step in the observation Once the criterion cannot be satisfied, the target terminal
sequence; It indicates that the speed will keep the lowest in speed at the end of the first stage path vek must be
the predicted domain. reduced. Moreover, if vek is reduced by 6 times and a
criterion- satisfied target speed still cannot be obtained, the
2.4. Algorithm of Obstacle Avoidance d plan1 is reduced by 5 m for a new path re-planning until
Considering that the collision avoidance disposal time the collision constraint is met.
under emergency conditions is very short, which should be
completed within only 2 ~ 3 s in general, the longitudinal 3. SIMULATION AND ANALYSIS
motion trajectory of the moving obstacle within 3 s needs to
be predicted when the first stage path planning is applied, To verify the proposed motion trajectory planning method,
as shown in Figure 5. two typical emergency conditions for static obstacle
The blue line in Figure 5 is the longitudinal position of avoidance and moving obstacle avoidance are carried out.
the moving obstacle in the future 3 s. Before the end of the The performance of the proposed motion planning method
first stage collision avoidance path, the longitudinal was assessed in CarSim under a high-adhesion surface
position of the ego vehicle must under the blue line at the condition.
 EMERGENCY COLLISION AVOIDANCE METHOD WITH MOVING OBSTACLES BASED ON 975

Figure 6. Motion planning process for moving obstacles.

3.1. Static Obstacle Avoidance Condition the red solid dash line depicted in Figure 7. The optimal
In this condition, a D-type car is employed to drive at an target path is not the closest one to the lane edge, which
initial speed of 81 km/h and is ready to avoid a static implies that the proposed piecewise path planning method
obstacle in front of it by 20 m. The proposed piecewise can find the potential optimal path compared with literature.
polynomial path planning method is compared to the The outer contour envelope of the vehicle corresponding to
benchmark which utilizes the cubic polynomial and quintic this tar-get path is represented by green thin solid lines. The
polynomial. green thin solid line does not overlap with the gray
In Figure 7, the gray solid block represents stationary rectangular block, indicating that the planned emergency
obstacles. When the emergency collision avoidance motion collision avoidance target path will not collide with
planning is triggered, the rear surface of the static obstacle obstacles. After the target path is determined, the curvature
is only 20 m away from the high-speed moving ego vehicle. of the target path needs to be obtained to plan a stable
This scene may appear at the entrance of the tunnel. As the vehicle speed profile.
visual sensor cannot recognize the obstacles in the dark at a Figure 8 shows that the continuous curvature curve
distance, the obstacle only can be found after the sensitivity reaches the maximum value of 0.011 m-1 at 41.4 m and the
increases when approaching. minimum value of -0.015 m-1 at 61.3 m. According to the
After the first collision avoidance path is planned by a vehicle kinematics model, the maximum lateral acceleration
quartic polynomial, the second path is planned by the of the vehicle reaches 7.6 m/s2 under the condition of
quintic polynomial. The thick solid lines in Figure 7 are the
remaining 7 candidate emergency collision avoidance paths
through position and curvature constraints shown in (9).
Then, based on the principle of the smoothest curve, the
target path for emergency collision avoidance is selected, as

Figure 7. Segmented collision avoidance path planning. Figure 8. Curvature of the target path.
976 Wen He et al.

constant vehicle speed. Based on the anti-sideslip constraint, almost half of the minimum curvature of the path planned
the ego vehicle needs decelerating to ensure motion by the quartic and quintic piecewise polynomial method. It
stability. indicates that the path planning by the cubic and quintic
The target speed is planned using the DP algorithm with piecewise polynomial is smoother than the path planning by
the stability classifier. In Figure 9, each blue thin solid line the quartic and quintic piecewise polynomial.
represents the executable state transition, and the red thick Figure 11 (a) depicts that both two target paths can be
solid line represents the optimal target speed determined tracked by the same ‘Driver Model’ in CarSim. Therefore,
based on the minimum comprehensive transition value. The the trajectory tracking method is the only factor affecting
planned target speed maintains the initial speed and remains the tracking performance. From the tracking results shown
constant. The tracking results of the target trajectory are in Figure 11 (b), the target path obtained by the cubic and
adopted to evaluate the rationality of the proposed planning quintic piecewise polynomial method can be tracked more
method. accurately.
Before observing the target motion trajectory tracking This is a very interesting result, which is contrary to our
results, the piecewise polynomial path planning method initial conjecture. In the section 2.1, we intend to continue
with a cubic polynomial and a quintic polynomial is the lane-change path curvature to improve path-tracking
adopted in literature, and plan the collision avoidance accuracy. However, the results show that under the action of
motion path when dealing with the same situation. the same driver model, the increase of this constraint does
In Figure 10 (a), the lane change distance of the target not lead to an increase in tracking accuracy. Because the
path planned by the cubic and quintic piecewise polynomial ‘Driver Model’ in CarSim applies the preview tracking
is 50 m which is longer than the path planned by the quartic algorithm, the small curvature mutation will not have a
and quintic piecewise polynomial. And in Figure 10 (b), the great impact on the tracking accuracy. On the contrary, the
first stage path curvature values of two different path increase of a constraint leads to the increase of the
planning methods are similar in size, however, the maximum curvature of the subsequent lane-changing
minimum curvature value of the path planned by the cubic collision avoidance path shown in Figure 10 (b), which
and quintic piecewise polynomial is -0.00823 which is leads to the maximum tracking error increase.
Figure 12 (a) shows that both two target speed profiles

(a) (b)

Figure 11. Path tracking results in different motion

trajectories.
Figure 9. Speed planning result.

(a) (b)
(a) (b)
Figure 12. Speed tracking results in different motion
Figure 10. Curvature of the target path. trajectories.
 EMERGENCY COLLISION AVOIDANCE METHOD WITH MOVING OBSTACLES BASED ON 977

can be accurately tracked. The speed tracking error in As shown in Figure 14, the blue vehicle tracks the target
Figure 12 (a) is caused by the vehicle longitudinal and motion trajectory planned by the quartic and quintic
lateral dynamic coupling. The longitudinal tracking error is piecewise polynomial, and the red vehicle tracks the target
always caused by the lateral dynamic fierce change. It motion trajectory planned by the cubic and quintic
shows that the target trajectory based on quartic and quintic piecewise polynomial. In the animation, both models can
piecewise polynomial programming is not satisfying. avoid static obstacles, which shows the correctness of the
In Figure 13, kappa presents the tire slip rate of the two planning methods.
vehicle; L and R present the left and right side of the
vehicle, respectively; 1 and 2 present the front and rear 3.2. Moving Obstacle Avoidance Condition
axles of the vehicle respectively. In this condition, the ego vehicle needs to avoid the moving
Figure 13 shows that the 4 maximum slip rates of the obstacle. This moving obstacle comes from a tracked
emergency collision avoidance trajectory obtained from the vehicle that suddenly decelerates ahead. The condition is
cubic and quintic piecewise polynomials are smaller than that the ego vehicle and the front vehicle drive in a straight
those of the emergency collision avoidance trajectory line along the lane with a relative distance of 20 m at a
obtained from the quartic and quintic piecewise speed of 90 km/h. At 1.2 s, the front vehicle encounters
polynomials. The results of the vehicle dynamic response of sudden obstacles and decelerates in an emergency. After 2 s,
different target paths indicate that the proposed method the front vehicle decelerates to 46 km/h and keeps the
using the quartic polynomial planning of the first stage path current speed constant.
leads the vehicle to a more dangerous moving state. To plan the collision avoidance trajectory of moving
The results from Figures 10 ~ 13 indicate that the obstacles, the Multivariate Gaussian Distribution Model is
high-order polynomial path planning method to ensure path carried out first to predict the speed of the front vehicle and
continuity increases the peak value of the planned path due calculate its longitudinal trajectory in the next 3 s, as shown
to the increase of constraints. This characteristic can not in Figure 15.
improve the trackability of the target trajectory. On the In Figure 15 (a), after adjusting the prediction speed
contrary, because the curvature of the target path increases, curve, the red line is treated as the future speed if the
the longitudinal and lateral dynamic coupling of vehicle moving obstacle is lower than its real speed. It proves the
tires increases, and the accuracy of lateral position and effectiveness of speed prediction. Based on the adjusted
speed tracking decreases. speed prediction, the longitudinal moving trajectory of the
obstacle is calculated as shown in Figure 15 (b). Then,
under the blue line in Figure 15 (b), the biggest value at 5 s
is chosen as the final longitudinal position of the path in the
first planning stage.
When the end longitudinal position of the path in the first
stage is determined to be 85 m, the path planning results
can be obtained by the piecewise polynomial method based
on two sampling schemes.
In Figure 16, the red line is the optimal path selected
from the candidate path clusters obtained by the piecewise
polynomial path planning method based on sampling φe and
xf two-dimensional variables. And the blue line is chosen

Figure 13. Kappa of four wheels.

(a) (b)

Figure 15. Speed and moving trajectory prediction of the

Figure 14. Actual collision avoidance animation. moving obstacle.
978 Wen He et al.

from the candidate path clusters obtained by a three- In Figure 18, each bar represents the time duration of the
dimensional sampling of φe, xf and yf. It indicates that the ego vehicle to reach the end longitudinal position with
increased sampling variable is helpful to find the smoothest different ending positions and speeds of the first stage. The
path. Moreover, it can be found from Figures 16 (a) and 16 three dashed lines are the threshold values for judging
(b) that the smoothest avoidance path is not necessarily at longitudinal collision avoidance. Only the value of the red
the edge of the lane after changing lanes. bar with the end speed of 22 m/s exceeds its corresponding
In the second step of motion planning for moving threshold. Therefore, the moving obstacle avoidance
obstacle avoidance, the vehicle speed for the paths based on motion trajectory is no longer planned. The end speed of
the DP algorithm is obtained to guarantee stable movement the first stage is 22 m/s and the longitudinal position of the
and avoid a longitudinal collision. end is 75 m, which is utilized as the constraint condition for
6 speed curves are exhibited in Figure 17 (a), which the moving obstacle avoidance trajectory planning.
reveals that longitudinal collision cannot be avoided by the In order to verify the effectiveness of collision avoidance
first five planned speed curves although they meet the trajectory, the vehicle model in CarSim is deployed to track
stable movement criterion. And the vehicle adopts the sixth the target motion trajectory. In addition, the traceability of
speed curve to finish the first stage path in 4.48 s. However, target paths obtained by different sampling methods is also
as shown in Figure 17 (b), the moving obstacle reaches the compared.
same longitudinal position in 4.96 s. This indicates that As shown in Figure 19 (a), both the two target paths are
before finishing the first stage collision avoidance path, the accurately tracked. Moreover, it can be noticed that the
ego vehicle will catch up with moving obstacles, which biggest lateral tracking error is 0.018 m, which implies that
means the collision will not avoid. the planned path is reachable. And the tracking error of the
The target line in Figure 17 (b) represents the time point, target path obtained by 3-dimensional sampling is lower
at which the moving obstacle passes the same longitudinal than that of 2-dimensional sampling. This manifests the
position. The time duration of the ego vehicle to reach the significance of adding sampling variables in this study.
end longitudinal position of the first stage path s less than
that of the moving obstacle, which cannot realize the
longitudinal collision avoidance.
To address the problem, the terminal longitudinal
position should be reduced by 5 m, and the path and speed
should be replanned correspondingly.

Figure 18. Judgment of the longitudinal collision

(a) (b) avoidance.

Figure 16. Path planning results.

(a) (b)
(a) (b) Figure 19. Path tracking results in moving obstacle
Figure 17. Path planning result. avoidance.
 EMERGENCY COLLISION AVOIDANCE METHOD WITH MOVING OBSTACLES BASED ON 979

(a) (b)

Figure 20. Speed tracking results in moving obstacle

avoidance.

Figure 22. Motion planning process for moving obstacles.

Figure 21. Longitudinal moving trajectory.

Figure 20 illustrates that the ego vehicle begins to Figure 22 (b), the brake light of the front vehicle lights up,
decelerate at 2 s from 90 to 79.2 km/h in 1.278 s. Since the and the front vehicle starts to decelerate at 1.81 s, and the
speed of the target vehicle does not change dramatically, the two vehicles begin to become closer. At 2.85 s in Figure 22
speed can be accurately tracked just by the PID controller. (c), the target path of the main vehicle diverges, and the
The ego and obstacle longitudinal moving trajectory are main vehicle is in a braking state, which means that the
carried out in Figure 21. main vehicle is tracking the target collision avoidance
The blue and yellow lines shown in Figure 21 are the trajectory at this time. The most critical point is at 3.96 s in
longitudinal movement on the first stage and second stage Figure 22 (d), which is the time to complete the first target
path of the ego vehicle, respectively. It can be noticed that path calculated in Figure 18. The ego vehicle successfully
the blue line is below the red line, indicating that the ego conducts the collision avoidance policy with an obstacle at
vehicle did not catch up with the moving obstacles in the that moment. The animation shown in Figure 22 indicates
first stage path. The accurate tracking of lateral position and that the proposed method can realize moving obstacle
speed shown in Figure 20 proves the stability of the ego avoidance.
vehicle, and the results in Figure 21 manifest the ability of Compared with the tracking error of static obstacle
collision avoidance of the ego vehicle. Thus, the avoidance in Section 3.2, Figures 19 and 20 indicate that
effectiveness of the proposed method in this study can be the speed and path tracking error in dynamic obstacle
demonstrated. avoidance is even similar with static obstacle avoidance.
In order to intuitively exhibit the process of the dynamic Under this condition, the avoidance process is more
obstacle avoidance of the ego vehicle, the animation in complex, but the tracking accuracy and safety can be
CarSim is intercepted in Figure 22. ensured at the same time, which manifests the advantage of
In Figure 22 (a), both the ego vehicle and the front the proposed method.
vehicle drive along the lane at 90 km/h. Consequently in
980 Wen He et al.

4. CONCLUSION multiconstraints. IEEE Trans. Vehicular Technology 66,

2, 952–964.
In this paper, a trajectory planning method for emergency Kappes, S., Schweter, F. and Jung, M. (2014). Method for
collision avoidance of moving obstacles is proposed. Firstly, Controlling Operation of a Braking System Independent
the piecewise polynomial path planning developed for static of the State of an Electronic Stability Control System.
obstacles is optimized. By adding sampling variables, the Google Patents.
smoother target collision avoidance path is effectively Kim, C. Y., Kim, Y. H. and Ra, W.-S. (2019). Modified 1D
obtained. Then, the optimized path planning method is also virtual force field approach to moving obstacle avoidance
extended to the emergency collision avoidance task for for autonomous ground vehicles. J. Electrical
moving obstacles. Derived from the estimation and Engineering & Technology 14, 3, 1367–1374.
correction of the speed of the front vehicle, the longitudinal Li, S., Wang, S. and Li, Z. (2020). Dynamic trajectory
trajectory of the front vehicle is estimated. Finally, by planning considering the moving trend of obstacle in the
estimating the longitudinal displacement of the front predictive horizon. IFAC-PapersOnLine 53, 5, 285–290.
obstacle, the terminal position and speed of the collision Liang, Y., Li, Y., Yu, Y. and Zheng, L. (2020). Integrated
avoidance path, namely the first stage polynomial path, are lateral control for 4WID/4WIS vehicle in high-speed
determined to ensure the application of the static emergency condition considering the magnitude of steering. Vehicle
collision avoidance trajectory method. The simulation System Dynamics 58, 11, 1711–1735.
results show the effectiveness of the proposed dynamic Montemerlo, M., Becker, J., Bhat, S., Dahlkamp, H.,
obstacle emergency collision avoidance method. Dolgov, D., Ettinger, S., Haehnel, D., Hilden, T.,
Hoffmann, G. and Huhnke, B. (2008). Junior: The
ACKNOWLEDGEMENTThis research is supported by stanford entry in the urban challenge. J. Field Robotics
Chongqing Technological Innovation and Application Development
Special Project-Chongqing University Theme Special Project
25, 9, 569–597.
(cstc2019jscx-zdztzxX0032). Ngo, P., Al-Sabban, W., Thomas, J., Anderson, W., Das, J.
and Smith, R. N. (2013). An analysis of regression
REFERENCES models for predicting the speed of a wave glider
autonomous surface vehicle. Australasian Conf. Robotics
Chu, K., Lee, M. and Sunwoo, M. (2012). Local path and Automation (ACRA), Sydney, Australia.
planning for off-road autonomous driving with avoidance Petrov, P. and Nashashibi, F. (2014). Modeling and
of static obstacles. IEEE Trans. Intelligent Transportation nonlinear adaptive control for autonomous vehicle
13, 4, 1599–1616. overtaking. IEEE Trans. Intelligent Transportation 15, 4,
Cicchino, J. B. (2017). Effectiveness of forward collision 1643–1656.
warning and autonomous emergency braking systems in Phung, M. D., Quach, C. H., Dinh, T. H. and Ha, Q. (2017).
reducing front-to-rear crash rates. Accident Analysis & Enhanced discrete particle swarm optimization path
Prevention, 99, 142–152. planning for UAV vision-based surface inspection.
Cui, Q., Ding, R., Wu, X. and Zhou, B. (2019). A new Automation in Construction, 81, 25–33.
strategy for rear-end collision avoidance via autonomous Ren, Y., Zheng, L., Yang, W. and Li, Y. (2019). Potential
steering and differential braking in highway driving. field–based hierarchical adaptive cruise control for
Vehicle System Dynamics 58, 6, 1602732. semi-autonomous electric vehicle. Proc. Institution of
Fagnant, D. J. and Kockelman, K. (2015). Preparing a nation Mechanical Engineers, Part D: J. Automobile
for autonomous vehicles: Opportunities, barriers and Engineering 233, 10, 2479–2491.
policy recommendations. Transportation Research Part Werling, M., Ziegler, J., Kammel, S. and Thrun, S. (2010).
A: Policy and Practice, 77, 167–181. Optimal trajectory generation for dynamic street
González, D., Pérez, J., Milanés, V. and Nashashibi, F. scenarios in a frenet frame. IEEE Int. Conf. Robotics and
(2015). A review of motion planning techniques for Automation (ICRA), Anchorage, Alaska, USA.
automated vehicles. IEEE Trans. Intelligent Transportation Wilson, M., Aylor, D., Zuby, D. and Nolan, J. (2015). Brake
17, 4, 1135–1145. burnishing effect on aeb performance. SAE Int. J.
Hu, X., Chen, L., Tang, B., Cao, D. and He, H. (2018). Transportation Safety 3, 1, 39–46.
Dynamic path planning for autonomous driving on Witt, M., Kompaß, K., Wang, L., Kates, R., Mai, M. and
various roads with avoidance of static and moving Prokop, G. (2019). Driver profiling–data-based
obstacles. Mechanical Systems and Signal Processing, identification of driver behavior dimensions and affecting
100, 482–500. driver characteristics for multi-agent traffic simulation.
Ji, J., Khajepour, A., Melek, W. W. and Huang, Y. (2016). Transportation Research Part F: Traffic Psychology and
Path planning and tracking for vehicle collision Behaviour, 64, 361–376.
avoidance based on model predictive control with Yu, Y., Cui, T., Li, Y., Zheng, L. and Zhang, Z. (2021).
Motion planning based on steering obstacle avoidance
 EMERGENCY COLLISION AVOIDANCE METHOD WITH MOVING OBSTACLES BASED ON 981

under emergency conditions. 6th IEEE Int. Conf. Dynamic decoupling and trajectory tracking for
Advanced Robotics and Mechatronics (ICARM), automated vehicles based on the inverse system. Applied
Chongqing, China. Sciences 10, 21, 7394.
Yu, Y., Li, Y., Liang, Y., Zheng, L. and Yang, W. (2020).
Publisher’s Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.