International Journal of Intelligent Transportation Systems Research (2022) 20:170–180

https://doi.org/10.1007/s13177-021-00281-2

Research on Path Planning Algorithm of Autonomous Vehicles Based

on Improved RRT Algorithm
Guanghao Huang1 · Qinglu Ma1

Received: 31 May 2021 / Revised: 16 September 2021 / Accepted: 29 September 2021 / Published online: 22 October 2021
© The Author(s), under exclusive licence to Intelligent Transportation Systems Japan 2021

Abstract
Recently, the path planning has become one of the key research hot issues in the field of autonomous vehicles, which has
attracted the attention of more and more related researchers. When RRT (Rapidly-exploring Random Tree) algorithm is
used for path planning in complex environment with a large number of random obstacles, the obtained path is twist and the
algorithm cannot converge quickly, which cannot meet the requirements of autonomous vehicles’ path planning. This paper
presents an improved path planning algorithm based on RRT algorithm. Firstly, random points are generated using the cir-
cular sampling strategy, which ensures the randomness of the original RRT algorithm and improves the sampling efficiency.
Secondly, an extended random point rule based on cost function is designed to filter random points. Then consider the vehicle
corner range when choosing the adjacent points, select the appropriate adjacent points. Finally, the B-spline curve is used
to simplify and smooth the path. The experimental results show that the quality of the path planned by the improved RRT
algorithm in this paper is significantly improved compared with the RRT algorithm and the B-RRT (Bidirectional RRT)
algorithm. This can be seen from the four aspects of the time required to plan the path, mean curvature, mean square deviation
of curvature and path length. Compared with the RRT algorithm, they are reduced by 55.3 %, 68.78 %, 55.41 % and 19.5 %;
compared with the B-RRT algorithm, they are reduced by 29.5 %, 64.02 %, 39.51 % and 11.25 %. The algorithm will make
the planned paths more suitable for autonomous vehicles to follow.

Keywords Autonomous vehicles · Path planning · RRT​· B-spline curve

1 Introduction force avoidance [1], graph search algorithm represented by

A* [2], convolutional neural network based on deep learning
With the development of artificial intelligence, autonomous [3], and intelligent optimization algorithm represented by
vehicles have gradually become a hot research direction of ant colony algorithm [4][5]. The above algorithms are only
new intelligent traffic, and path planning is the key technol- suitable for solving the planning of autonomous vehicles
ogy to ensure the smooth and safe operation of autonomous in simple scenarios, the iteration time and path executable
vehicles. The path planning technology of autonomous vehi- degree all have certain problems. According to the path plan-
cle is mainly to plan a passable path without collision under ning algorithm based on Rapidly-exploring Random Tree
the condition that the start point and target point are given (RRT), by detecting whether sample points in the travelling
in the open road environment and the vehicle dynamics con- space are in collision, environmental modelling is omitted,
straints are satisfied. The current path planning algorithms of which can effectively solve the path planning problem in
robots are mostly used to solve such problems, such as arti- complex environment [5][6].
ficial potential field method based on virtual potential field Although RRT algorithm can be used in unknown envi-
ronment, and it can better handle with nonholonomic con-
* Guanghao Huang straint problem of path planning algorithm at the same time.
[email protected] It has great advantages in many aspects, but there are also
* Qinglu Ma some problems:
[email protected]
1
Chongqing Jiaotong University, No. 66 Xuefu Road, Nanan
Distr, Chongqing 400074, China

13
International Journal of Intelligent Transportation Systems Research (2022) 20:170–180 171

(1) The random tree grows without direction since the convergent and can be used for path planning of robots with
search path strategy is uniform random sampling, nonholonomic constraints [14]. Luca et al. proposed Flat-
which affects the speed of the algorithm; RRT* algorithm that exploits pre-computed motion primi-
(2) The selection of sampling points is too redundant. The tives in the flat output representation to accelerate the con-
selection of the nearest neighbor point only considers struction of the edges, while building the tree from randomly
the distance from the node in the random tree, which extracted nodes in the configuration space. The algorithm
affects the search results and efficiency of the RRT takes into account the driving constraints of vehicle dynam-
algorithm to a certain extent; ics in the optimal trajectory design [15]. Zaid et al. proposed
(3) The RRT algorithm does not consider the constraints Potentially Guided Intelligent Bi-directional RRT* (PIB-
of vehicle dynamics when planning path, the generated RRT*) and Potentially Guided Bi-directional RRT* (PB-
path is not smooth enough, and the autonomous vehicle RRT*) which use BPG (Bi-directional Potential Gradient)
is poorly executable. heuristic to potentially guide two rapidly-exploring random
trees towards each other. Two algorithms have proven to be
Aiming to solve the problems existing in RRT algorithm, experimentally avoids getting stuck in local minima as APF
many scholars have proposed different improved algorithms. does and converges to optimal solution faster than its state of
Kuffner et al. proposed the RRT-Connect algorithm in 2000. the art counter parts like IB-RRT*, RRT* and PRRT*; fewer
A random tree was generated at the start point and the tar- memory is consumed by these algorithms as fewer itera-
get point respectively, which reduced the search space and tions and time is required by them [16]. Haoyue Liu et al.
accelerated the convergence speed of the algorithm [6]. LaV- proposed an improved RRT algorithm combining target bias
alle et al. also proposed GB-RRT (Goal-bias RRT) algo- strategy and bidirectional search method. The sixth order
rithm in 2001, taking the end point as the sampling node B-spline curve is used to connect the bidirectional random
and searching the surrounding space, and the planned path tree to smooth the path. The algorithm can not only satisfy
had target bias [5]. Jim Bruce proposed the Extend RRT the nonholonomic constraint of mobile robot but also effec-
algorithm in 2002, which introduced the set of path points to tively reduce the search time [17]. Yang Li et al. proposed a
accelerate the convergence speed and improve the stability new adaptive step size RRT algorithm. By establishing norm
of the path [7]. Karaman proposed RRT* algorithm in 2011, compatible inequality of configuration space and workspace,
which reduced redundant nodes and improved algorithm effi- the step size in workspace is constrained within the allow-
ciency by reselecting parent nodes and rewiring [8]. Jordan able range, and then effective collision detection is realized.
et al. proposed B-RRT * (Bidirectional RRT*) in 2013, that The adaptive step size RRT method can determine the step
is Bidirectional extended RRT*, and improved the heuris- size without multiple debugging, which improves the robot’s
tic function to obtain the optimal path [9]. Adiyatov et al. path planning speed [18]. Zhuohua Yang et al. proposed the
proposed the RRT*FN algorithm in 2013, which reduced LT-RRT* algorithm, which used line segment theorem to
the consumption of computer memory by giving the maxi- improve the selection of new nodes and surrounding parent
mum number of nodes in the random tree [10]. Gammell nodes, and to some extent reduced path break points and
et al. proposed the Informed RRT* algorithm in 2014, which improved the smoothness of the path [19]. Yangjie Li et al.
improves the efficiency of the algorithm by directly sampling proposed a new algorithm PQ-RRT*, which combined the
the super-ellipsoid subset of the search space [11]. Based advantages of P-RRT* (potential functions based RRT*) and
on B-RRT* algorithm, Ahmed et al. proposed intelligent Quick-RRT*. PQ-RRT* changes the sampling strategy by
bidirectional RRT* algorithm (IB-RRT*). The algorithm adopting the attraction of the target region and expands the
introduces intelligent sample insertion heuristic and heuris- search range of the selected parent node [20].
tic strategy of random tree connection is improved. In terms At present, the related research on RRT and improved
of probabilistic completeness, asymptotic optimality, conver- methods are mainly focused on the path planning of mobile
gence to optimal solution and computational complexity, the robots. In terms of path planning for autonomous vehicles,
algorithm has significant advantages over the RRT* algo- RRT-related applications are relatively rare. At the same
rithm and B-RRT* algorithm [12]. Ahmed et al. proposed an time, most current algorithms adopt random distribution
improved algorithm (P-RRT*) which incorporate APF into sampling methods when performing random tree expan-
RRT*.The inclusion of APF provides a direction to explore sion. These methods will generate a large number of random
and makes P-RRT* converge more rapidly than RRT* [13]. points that are invalid for random tree expansion, reducing
Nino et al. put forward TWIN-RRT* algorithm. The TWIN- the search efficiency of the algorithm. These RRT-related
RRT* algorithm comprises two trees growing from the same algorithms lack consideration of the constraints of the actual
initial and final configurations. The two trees grows have the vehicle model during the collision detection process. These
same inherited characteristics of RRT* algorithm. Experi- algorithms are prone to generate invalid paths that collide
mental results show that the improved algorithm is more with obstacles and the generated paths are less practical.

13
172 International Journal of Intelligent Transportation Systems Research (2022) 20:170–180

This paper studies the path planning problem of autono- vehicle steering model is established according to Acker-
mous vehicles in complex environment. For the path plan- man’s steering geometry, and the vehicle steering motion
ning of RRT algorithm in complex environment with a large characteristics are analyzed. In the process of vehicle
number of random obstacles, the resulting path is curved movement, electronic signals control the track of vehicle
and the algorithm cannot converge quickly. An improved movement through wire control. Therefore, it is necessary
RRT algorithm is proposed, which uses circular sampling to analyze the relationship between the motion radius and
strategy to preliminarily select random points and designs the angle of the vehicle when it is moving. In the design
a cost function to extend random points. Then, the vehicle of modern vehicle steering system, in order to reduce the
corner range is considered when selecting appropriate adja- road resistance and avoid tire wear, the steering system of
cent points. Finally, the B-spline curve is adopted to simplify the vehicle is required to ensure that all the wheels make
and smooth the path. pure rolling motion when the vehicle is steering. Obvi-
ously, this is only true if all the steering axes of the wheels
intersect at the steering center. The analysis of vehicle
2 Vehicle Steering Constraint Model motion trajectory and establishment of obstacle avoid-
ance kinematics model are based on the basic principles
In the course of trajectory planning of autonomous vehi- and research criteria of Ackerman’s steering geometry, as
cles, vehicle kinematic constraints should be considered to shown in Fig. 1.
obtain the driving path satisfying the motion characteristics As shown in Fig. 1, in two dimensional coordinate
of vehicle steering mechanism. In 1817, the German vehicle space, the autonomous vehicle goes through the process
engineer Lankensperger proposed Ackerman steering geom- of obstacle avoidance. O can be as turning center of the
etry [21]. The vehicle designed according to the Ackerman vehicle, θ1 is the left front wheel steering angle, θ2 is the
steering geometry, when turning along a curve, the steering right front wheel steering angle, δ as the Ackerman steer-
angle of the inner wheel is about 2 to 4 degrees larger than ing angle, R 1, R2, R3 and R4 are turning radius for four-
that of the outer wheel by using the equal crank of the four- wheel vehicles around the turning center T, R is vehicle
bar linkage, Which makes the center of the four wheel paths turning radius for rigid body center T of mass around the
roughly intersect with the instantaneous turning center on center O, L1 and L2 are the rear suspension and front sus-
the extension line of the rear axle, so that the vehicle can pension of the vehicle respectively, W for wheel tread, L3
turn smoothly. as the wheelbase, the motion state equation of the vehicle
After reducing the obstacle avoidance process of the as Eq. (1):
autonomous vehicle to equivalent rigid body motion, the

Fig. 1  Obstacle avoidance O'

kinematic model of the autono-
mous vehicle

R'
R2' R1'
R4' R3' R'
L1 L3 L2 L1 L3 L2
A D A'' D''
T 1
W W T''
B 2 B'' C''
C
R0

R1
R2
R3 R4 R
R3 R
R4 R0
R1
R3 R2

13
International Journal of Intelligent Transportation Systems Research (2022) 20:170–180 173

⎧ ctg𝜃1 − ctg𝜃2 = B
⎪ L3
⎪ L3
⎪ 𝛿= arctan
⎪ R4 + B∕2
⎪ 2L3 × tan 𝛿 Qnear Qrand
⎪ 𝜃1 = arctan Qnew
⎨ 2L3 -B × tan 𝛿 (1) Qinit
⎪ Qgoal
⎪ 𝜃 = arctan 2L3 × tan 𝛿
⎪ 2 2L3 +B × tan 𝛿
⎪ L3 ∕2
⎪ 2sin(arctan R +B∕2 ) Fig. 2  Random tree extension of RRT algorithm
⎪K = 1= 4

⎩ R L3

The vehicle of Ackermann’s steering is affected by the Qnear and Qnew conflict with obstacles. If there are conflicts,
non-holonomic constraint. The lateral motion of the vehicle abandon the current node and reselect Qrand and repeat the
is only considered here, and the longitudinal motion is not above process of finding Qnear points and conflict judgment.
considered. It is assumed that the uniform motion velocity If there are no conflicts with the obstacle, then add the Qnew
is V. The difference between autonomous vehicle and robot node to the random extension tree. New random trees are
is that the shape of its motion trajectory can not be arbitrary. formed by continuously sampling random points and extend-
The constraint of the front wheel Ackerman steering angle ing new nodes. While avoiding obstacles, the target point is
determines that autonomous vehicle cannot reach any posi- gradually approached until the newly generated Qnew satis-
tion. The driving trajectory needs to meet some conditions, fies D (Qnew, Qrand) < step size ρ. And the nodes on the
such as minimum turning radius Rmin(R>Rmin, different line connecting Qnew and Qgoal are all in the safe area Qsafe,
values of different models) and maximum rate of curvature then Qgoal is added to the random tree. It can be regarded as
change Kmax (K<Kmax, inversely proportional to the turning having found a feasible path from the start point Qinit to the
radius) and the maximum heading angle of the vehicle µmax. target Qgoal.

3.2 Optimize the Selection of Sampling Points

3 RRT Algorithm of Improved Sampling
Strategy • Preliminary sampling of random points

3.1 Basic RRT Algorithm The generation of sampling points in RRT is generally

dependent on random Gaussian sampling, which has the
LaValle et al. first proposed the RRT algorithm in 2000, and uniform sampling probability and great randomness in the
the sampling algorithm was based on the incremental motion whole motion space. This method has low search speed and
method. Taking the root node as the initial start point, add- long calculation time, which makes it difficult to meet the
ing child nodes through random sampling in the state space requirements of autonomous vehicle motion planning. This
and connecting these nodes to form a random tree. Expand- paper will more purposefully select sampling points. The
ing from one child node outward to generate several random start point and target point are the known information in
trees, it stops growing until the random trees reach the target actual map. On this basis, the algorithm in this paper sets the
area. The path expansion process is shown in Fig. 2. middle point of the line between the start point and the target
Assuming that the state space moved by the autonomous point as a virtual random point Qvir. Then it is necessary to
vehicle is Q, where the unfeasible areas with obstacles is judge whether there are obstacles at the virtual random point
Qobs, and other areas available for driving are safe areas Qvir. If collision detection shows no obstacle, then the virtual
Qsafe, then Q=Qobs+Qsafe. Initializing the start node Qinit, random point is set to a random point Qrand, otherwise need
target node Qgoal and Qobs positions. Random sampling Qrand to rechoose sampling points. Draw a circle with the line con-
node in the area of Qsafe, and find the node Qnear with the necting the start point and the target point as the diameter
shortest distance from Qrand in the extended tree. Above of circle, and the virtual sampling point Qvir as the center of
nodes satisfy the Euclidean distance D (Qnear, Qrand) = Min circle. In the sampling process, the radius of circle is used
D (q, Qrand)[q is the set of nodes in the expanded tree]. Then, as the sampling radius to generate random points on the
take Qnear as the start point and extend the step length ρ in circle. If the circle intersects with Qsafe in the state space at
the direction of Qrand to generate a new node Qnew. Deter- this time, the intersected point is taken as the random point
mine whether Qnew and the nodes on the connection between Qrand. If there is no intersection point, the sampling radius

13
174 International Journal of Intelligent Transportation Systems Research (2022) 20:170–180

is controlled to redraw the circle until the intersection point more than one step, and then determine the priority to be
is taken as a random point. extended. Random points are generated in the sector range
As shown in Fig. 3(a), in the sampling process, the target of vehicle corner, as shown in Fig. 3(b). When the vehicle
point is taken as the center and the radius of the peripheral is located at the Qnear node and is traveling along the x-axis,
circle of the planning map is the sampling radius. First, the the corner boundary is a thick solid line. The connecting
random radius R is generated by the first term of Eq. (2), angle between the random point and the nearest tree node
and then random points are generated by the second and and the connecting angle between the nearest tree node and
third terms of Eq. (2) on the circumference of the concentric its parent node is less than the maximum heading angle. The
circle of the random radius R. During the sampling process, generated random points are valid.
the concentric circles are adjusted by means of coefficient The coordinate of the random point is (x, y), the coordi-
m and random number n. This sampling strategy can greatly nate of the nearest tree node is (xn, yn), and the coordinate of
improve the planning speed while retaining the randomness the parent node is (xf, yf). The slope of the line between the
of Gaussian sampling. The specific calculation formula is nearest tree node and the parent node and the line between
as Eq. (2): the nearest tree node and the random point can be expressed
m+2
as k1 and k2, and the included angle of the two lines can be
⎧ rrand = Rs ⋅ n 4 obtained as Eq. (3):
⎪� � � � � �
⎨ Qrand (x) Qvir (x) cos 2n𝜋 (2)
⎪ Q (y) = + rrand ⎧ � � y −y y−y
⎪ ⎛ xn −xf − x−xn ⎞
⎩ Qvir (x) sin 2n𝜋 k1 − k2
= arctan ⎜ ⎟
rand n f n
⎪ 𝛽 = arctan
⎨ 1 + k ⋅ k ⎜ 1 +
yn −yf
−
y−yn ⎟
Where, n is the random number within the range of 0 to
1 2
⎝ xn −xf x−xn ⎠
⎪
1, m is the concentric circle coefficient, Rs is the sampling ⎪ 𝛽 ⩽ 𝜇max
⎩
radius, Qvir (x), Qvir (y) are respectively the horizontal and
(3)
vertical coordinate values of the virtual random point, Qrand
(x), Qrand (y) are the horizontal and vertical coordinate val-
ues of the random point. • Selection strategies for near points

• Design of random point extension rules The traditional RRT algorithm uses the shortest dis-
tance random sampling node Qrand as Qnear, and then the
Considering the stability and comfort of autonomous new node is connected with the Qnear node to obtain the
driving, based on the vehicle steering constraint condi- Qnew node. This measurement method is simple and deter-
tions, set the vehicle’s largest course angle of 40°. Due to mines the efficiency of the algorithm and the suitability of
characteristics of left and right wheel steering angle, set node selection, which is the most important part of RRT
Ackermann angle as the vehicle front wheel steering angle. algorithm optimization. However, as applied to autono-
Considering the sampling point range, extended Qrand ran- mous vehicles, the included angle relationship between
dom points must be located in the specified scope. The dis- nodes should be considered when selecting the nearest
tance between random point Qrand and the tree node need to neighbor nodes, that is, the influence of new nodes on

Fig. 3  Sampling diagram.

(a) Selection of random points
(b) Limited range of random Qstart Qnear2
pointsextension
Qrand1
Qrand2
Random
radius
Qrand Sample 2
radius
Qvir
Qrand3Sector sampling range
1

Qparent Qnear1
Qgoal Sector sampling range Qgoal
Scope of the map

13
International Journal of Intelligent Transportation Systems Research (2022) 20:170–180 175

the overall path tortuous degree. Given the steering abil- ⎧ d − dmin
ity of the vehicle, the node should be chosen in favour of ⎪ N1 = dmax − dmin
flat rather than the shortest Euclidean distance. This paper ⎨ (5)
⎪ N2 = 𝜆 − 𝜆min
redesigns the measurement method, and screens the nodes ⎩ 𝜆max − 𝜆min
connected with the new node by calculating the cost func-
tion of each node in a specific area. In the experiment, the Compared with the traditional RRT algorithm, which uses
feasibility and correctness of the new measurement func- the node with the minimum distance from Qrand as Qnear,
tion can be effectively verified. the improved measurement method is more convergent. By
As shown in Fig. 3(b), the cost function values of the taking advantage of the fast convergence speed of heuristic
two nodes Qnear1 and Qnear2 are calculated, and the node function, its advantages are well integrated into the basic
with the smallest cost function is selected as the final node RRT algorithm, making it more greedy in selecting new
Qnear. The cost function as show in Eq. (4): nodes.
�
⎧ � �2 � �2 3.3 Smoothing Processing of Planned Trajectory
d
⎪ 1 = N 1 xi − xs + yi − ys
⎪ �
� �2 � �2 In this paper, a path simplification method considering the
⎪ d2 = N1 xi − xg + yi − yg
⎨ � (4) angle constraint between vehicle and the next path point is
⎪ � �2 � �2
⎪ d3 = N1 xi − xr + yi − yr proposed. The path obtained by the improved RRT algorithm
⎪ is simplified, and the simplified path is smoothed by using
⎩ F = 𝜔1 d1 + 𝜔2 d2 + 𝜔3 d3 + N2 𝜆i the B-spline curve to generate a smooth path with continu-
ous curvature and satisfying the vehicle driving constraints.
Since distance and angle are two different dimensions.
In order to unify the two different types of data under the
• Path simplification under the constraint of angle
same reference frame, it is necessary to conduct standard
normalization processing on these two variables. N1 is the
By using the random tree in the improved RRT algo-
normalized function corresponding to distance, and N2
rithm, the feasible path points and the corresponding angles
is the normalized function corresponding to angle. The
of the path points from the initial state to the target state can
standard normalized function is shown in Eq. (5). Param-
be obtained. As shown in Fig. 4, µi is the included angle
eters ω1and ω2 are corresponding weights, which can be
between the path points Qi-1 and Qi and the positive direction
set according to the actual application.
of x-axis, and µi+1 is the included angle between the path
points Qi and Qi+1 and the x-axis. In this paper, the position
angle of vehicle start point Qinit is set to 90 degrees.

Fig. 4  Diagram of path sim-

plification. (a) Angle of nodes
(b) Path simplification

13
176 International Journal of Intelligent Transportation Systems Research (2022) 20:170–180

As shown in Fig. 4, the start point Qinit is taken as the ∑

n

start point Qk for the first connection, and subsequent path Z(v) = Ni,k (v) ⋅ Pi (6)
points are successively connected. In Fig. 4, µk is the angle i=0

corresponding to the start point Qk, µk(i-1) is the included

angle between the connecting line of the start point Qk n!
Ni,k (v) = Cni vi (1 − v)n−i = (1 − v)n−i vi , i = 0, 1, 2 … , n
(n − i)!i!
and the path point Qi-1 and the x-axis, µi is the angle cor-
(7)
responding to the path point Qi, and µmax is the maximum
included angle between the vehicle and the next path point. Where: Z(v) - motion control point; P(i) - the position
The line between the two path points does not collide point of the ith curve; v - node vector; Ni,k(v) - Bernstein
with obstacles. When the absolute value of the difference polynomial. When n=1, the first order B-spline curve is a
between µk and µk(i-1) is less than the maximum included straight line with two positions; When n=2, the B-spline
angle µ max , and the absolute value of the difference curve of the second order is a parabola with three positions;
between µi and µk(i-1) is less than the maximum included When n≥3, the curve is a higher-order B-spline curve with
angle µmax, the path points between the start point Qk and n+1 position points.
the path point Qi-1 are eliminated. And the start point Qk On the basis of RRT algorithm and Ackermann steering
and the path point Qi-1 are directly connected. In turn, until constraints, this paper proposes an improved RRT algorithm.
the condition is not met, the parent node Q i of the path The algorithm process is shown in Fig. 5.
point Qi-1 that does not meet the condition is used as the First, random points are generated using the circular sam-
new start point Qk, and µk(i-1) is used as the angle value of pling strategy to determine whether the random points are
the new start point Qk, and the above operation is repeated within the map range. If not, random points are generated
until it is connected to the target point Qgoal. again. An extended rule of random points based on cost
function is designed to filter random points. Then, when
• Path smoothing based on B-spline curve choosing the adjacent points, take the vehicle’s corner range
into consideration and choose the appropriate adjacent
Due to the random sampling nature of the RRT algo- points. After the random tree is obtained, the angle between
rithm, the generated paths are often unnatural and dither- the vehicle and the next path point is used to simplify the
ing, which containing many unwanted folds. Especially path and the B-spline curve is used to smooth the path.
in complex environment, obstacle constraints make more
folding points generated, which will reduce the tracking
stability of autonomous vehicles. So the original trajectory 4 Simulation Experiment and Analysis
needs to be smoothed. The B-spline curve has continu-
ity and locality. In this paper, B-spline curve is selected 4.1 Simulation Experiment
for curve fitting to generate a smooth path that can be
executed by autonomous vehicles. The expression form of In this paper, Nan’an campus of the Chongqing Technol-
B-spline curve of order K are shown in Eq. (6) and Eq. (7). ogy and Business University is selected as the research area,
starting from 17 teaching dormitory buildings of Chongqing
Technology and Business University, and aiming at the inter-
section of Cuihu road and Huayuan road. The research area

Y
Algorithm initialization Set Qvir to the
Use the cost function to
random point Qrand generate Path
select the adjacent point
Y Qnear
Set the start and target points
Determine if Qrand is within
Draw a circle with The growth step generates Curve smoothing
the sector of the vehicle's
t a k e the center of the line the line between the maximum heading angle new nodes
between the starting point and the starting point and the
ending point as Qvir end point as the
diameter, and initially
N generate the random
Whether the target point
has been reached
point Qrand by N
Determine if it is within the map reducing the radius N Y

Fig. 5  Path simplification algorithm

13
International Journal of Intelligent Transportation Systems Research (2022) 20:170–180 177

is captured on Baidu map and the main roads are extracted Fig. 7 is the final result of improved RRT algorithm path
from the map. The abstract of buildings was black to repre- planning.
sent obstacles, and the abstract of roads was white to repre- It can be seen from Fig. 7 that, due to the random sam-
sent passable areas. Attempt to plan a collision free path for pling characteristics of RRT algorithm, the random trees
autonomous vehicles. The research area is shown in Fig. 6. occupy a large number of invalid search space and the cur-
In order to verify the superiority, effectiveness and real- vature of the searched path changes too much, even causing
time performance of the improved RRT algorithm, simula- right and acute angles. Although the sampling strategy of
tion comparison experiments are carried out in this paper. the B-RRT algorithm increases the bias towards the target
By comparing with RRT, B-RRT algorithm and improved points and reduces the randomness to a certain extent, there
RRT algorithm of this paper in simulation experiment, the are still many invalid branches in the random tree. The path
correctness and superiority of the algorithm are verified. curvature changes greatly, which is difficult to meet the con-
Set the vehicle in the environment as a rounded rectan- ditions of normal driving of vehicles. The improved RRT
gle frame, the whole state space size is 760 × 460 pixel. A algorithm has fewer useless branches, faster search and cal-
step in the simulation environment corresponds to a real life culation speed, more gentle path and less curvature change
autonomous vehicle traveling 5 m. Each time the vehicle than the previous two algorithms.
takes a step, it explores the unknown area forward. By gen-
erating random trees step by step, the vehicle moves to the 4.2 Comparison Results of Three Algorithms
target position. Finally, all the steps are connected to form a
planned path. The obstacle areas are represented by black, Due to the randomness of RRT algorithm, in order to objec-
the red dot represents the start position of the vehicle, and tively evaluate the merits of three algorithms, the experi-
the green dot represents the target position. The path plan- ments are shown in Fig. 7 scenario of RRT, B-RRT and
ning results are shown in Fig. 7. In Fig. 7, the red broken improved RRT algorithm, respectively. The 102 times of
lines are the random trees and branch path, and Fig. 7 is experiments were carried out on the three algorithms. This
the final result of RRT algorithm path planning. In Fig. 7, paper get rid of the best and the worst data each algorithm,
the blue broken lines are the start point random tree and its counts statistical 100 times each algorithm to calculate the
branch path, the red broken line is the end point random tree average search time, mean curvature, mean square deviation
and its branch path, and Fig. 7 is the final result of B-RRT of curvature and path length.
algorithm path planning. In Fig. 7, the red broken line is the As shown in Fig. 8(a), compared with the other two algo-
random tree and the blue broken line is branch path, and rithms, the improved RRT algorithm spends the least time

Fig. 6  Binarization of map.

(a) Map of Nan'an campus
(b) Map after extracting main
roads

13
178 International Journal of Intelligent Transportation Systems Research (2022) 20:170–180

Fig. 7  Comparison among three

algorithms. (a) Planning pro-
cess of RRT (b) Result of RRT
(c) Planning process of B-RRT
(d) Result of B-RRT (e) Process
of improved RRT (f) Result of
improved RRT​

in path planning and has a faster path planning speed. As speed and the planning path quality, the improved RRT algo-
shown in Fig. 8(b) and (c), compared with the other two rithm based on B-spline curve smoothing is proposed. The
algorithms, the mean curvature and mean square devia- algorithm uses circular sampling to overcome problems of
tion of curvature of the path planned by this algorithm are random sampling and calculation speed slow in RRT algo-
smaller, indicating that the path quality planned by this algo- rithm, and incorporates the vehicle position and the next
rithm is better. As shown in Fig. 8(d), the length of the path path point angle constraint into the process of path planning
planned by this algorithm is shorter than that of the other at the same time. Finally, using B-spline curve to simplify
two calculation methods. The average of the 100 statistics the path and smoothing process reduces the path length
in Fig. is shown in Table 1. while reducing the mean curvature and mean square devia-
Compared with the other two algorithms, the improved tion of curvature of the path. The improved RRT method
RRT algorithm proposed in this paper reduces the average proposed in this paper provides a better algorithm model
time of path planning by 55.3 % and 29.5 %, and the aver- for the path planning system of autonomous vehicles. By
age length of path planning by 19.5 % and 11.25 %. The constructing a priori search space within the road or lane
mean curvature of the planned path decreased by 68.78 % boundary and processing exhaustive sampling and greedy
and 64.02 %, and the mean square deviation of curvature search in the state space, it can improve the efficiency of
decreased by 55.41 % and 39.51 %. vehicle path search technology while avoiding obstacles
accurately. In the end, the path generated by the algorithm
is smooth enough to be executed by the vehicle’s motion
5 Conclusions control system in real time.
Due to the randomness of RRT, there is no guarantee
In this paper, the RRT algorithm is applied to path planning that the path planned for each time in the same scenario
of autonomous vehicles in complex environment. Aiming at is the same. It is hoped that through further research, the
the shortcomings of the original algorithm in the calculation planning results can be stabilized and approach the global

13
International Journal of Intelligent Transportation Systems Research (2022) 20:170–180 179

Fig. 8  Comparison of experi-

mental data. (a) Comparison
figure of processing time (b)
Comparison figure of mean
curvature (c) Comparison figure
of mean square deviation of
curvature (d) Comparison figure
of path length

optimal solution in real-time calculations. The improved Acknowledgements This research was funded by National Social
RRT algorithm proposed in this paper does not consider Science Fund National Emergency Management System Construction
Research special project: Research on the path of improving the qual-
the impact of vehicle speed changes on the planned path. ity and upgrading of urban comprehensive emergency management
Therefore, a systematic study of the relationship between capability under major emergencies, grant number 20VYJ023 and
turning angle constraints and vehicle speed is of great sig- This research was funded by Key project of science and Technology
nificance to ensure maximum stability and road capacity. Research Plan of Chongqing Education Commission: Theory and Prac-
tice of self-organization optimization of traffic block chain in interlaced
However, both the RRT algorithm and the improved RRT area under the environment of intelligent network connection, grant
algorithm are model-driven and have certain limitations, number KJZD-K202000704.
which require further research. Combining data-driven
approach to path planning and obstacle avoidance of
autonomous vehicles will ultimately make the technology
of autonomous vehicles more mature. References
1. Khatib, O.: Real-time obstacle avoidance for mobile robots manip-
ulators and mobile robots. J. Int. J. Robot. Res. 5(1), 90–98 (1986)
2. Hart, P.E., Nilsson, N.J., Raphael, B.: A formal basis for the heu-
Table 1  Data of experimental ristic determination of minimum cost paths. IEEE Trans. Syst. Sci.
Cybern. 4(2), 100–107 (1968)
Time/s Mean Mean square Path length/step 3. Lu Y., Yi S., Liu Y., et al.: A novel path planning method for
curvature/ deviation of biomimetic robot based on deep learning. Assem. Autom. 36(2),
m−1 curvature 186–191 (2016)
4. Jiang, M., Wang, F., Sun, L.: Research on path planning of mobile
RRT​ 5.827 0.189 0.278 1207.895
robot based on improved ant colony algorithm. Chin. J. Sci.
B-RRT​ 3.692 0.164 0.205 1095.619 Instrum. 4(2), 113–121 (2019)
Improved 2.603 0.059 0.124 972.353 5. Lavalle, S.M., Kuffner, J.R.J.J.: Randomized kinodynamic plan-
RRT​ ning. Int. J. Robot. Res. 20(5), 378–400 (2001)

13
180 International Journal of Intelligent Transportation Systems Research (2022) 20:170–180

6. Kuffner, J.R.J.J., Lavalle, S.M.: RRT-connect: An efficient 20. Li, Y., Wei, Wu, Gao, Y., et al.: PQ-RRT*: An improved path
approach to single-query path planning. Proceedings of the IEEE planning algorithm for mobile robots. Expert systems with appli-
International Conference on Robotics and Automation (ICRA), cations, vol. 152 (2020)
vol. 2, pp. 995-1001. IEEE, San Francisco (2000) 21. Veneri, M., Massaro, M.: The effect of Ackermann steering on the
7. Bruce, J., Veloso, M.M.: Real-time randomized path planning for performance of race cars. Veh. Syst. Dyn. 59(6), 907–927 (2021)
robot navigation. In: Kaminka, G.A., Lima, P.U., Rojas, R. (eds.)
RoboCup 2002. Robot Soccer World Cup VI, RoboCup (2002) Publisher’s Note Springer Nature remains neutral with regard to
8. Karaman, S., Frazzoli, E.: Sampling-based algorithms for optimal jurisdictional claims in published maps and institutional affiliations.
motion planning. Int. J. Robot. Res. 30(7), 846–894 (2011)
9. Jordan M, Perez A.: Optimal Bidirectional Rapidly-Exploring
Random Trees,”[EB/OL]. [2021-10-07]. http://​dspace.​mit.​edu/​
bitst​ream/​handle/​1721.1/​79884/​MIT-​CSAIL-​TR-​2013-​021.​pdf?​ GUANGHAO HUANG Guanghao
seque​nce=2. Huang received the bachelor’s
10. Adiyatov, O., Varol, H.A.: Rapidly-exploring random tree based degree in traffic engineering
memory efficient motion planning. IEEE International Conference from Yancheng Institute of Tech-
on Mechatronics and Automation, Takamatsu, pp. 354-359 (2013) nology, Jiangsu, China, in 2019.
11. Gammell, J.D., Srinvasa, S.S., Barfoot, T.D.: Informed RRT*: He is currently pursuing the
Optimal sampling-based path planning focused via direct sam- master’s degree with the Chong-
pling of an admissible ellipsoidal heuristic. International Confer- qing Jiaotong University. His
ence on Intelligent Robots and Systems. pp. 2997-3004. IEEE, current research interests include
Chicago (2014) intelligent transportation and
12. Qureshi, A.H., Ayaz, Y.: Intelligent bidirectional rapidly-exploring traffic big data processing.
random trees for optimal motion planning in complex cluttered
environments. Robot. Auton. Syst. 68, 1–11 (2015)
13. Qureshi, A.H., Mumtaz, S., Ayaz, Y., et al.: Triangular
geometrized sampling heuristics for fast optimal motion planning.
Int. J. Adv. Rob. Syst. 12(2), 86–93 (2015)
14. Pereira, N., Ribeiro, A.F., Lopes, G., et al.: Path planning towards QINGLU MA Qinglu Ma received
non-compulsory multiple targets using TWIN-RRT*. Ind Robot the Ph.D. degree in computer
43(4), 370–379 (2016) technology and science from
15. Bascetta, L., Iñigo Arrieta, M., Prandini, M.: Flat-RRT*: A Chongqing University, Chong-
sampling-based optimal trajectory planner for differentially flat qing, China, in 2012. He is cur-
vehicles with constrained dynamics. IFAC-Papers Online, vol. rently a Professor with the
50, no. 1, pp. 6965-6970 (2017) Chongqing Jiaotong University.
16. Tahir, Z., Qureshi, A.H., Ayaz, Y., et al.: Potentially guided bidi- He is the co-inventor of 10
rectionalized RRT* for fast optimal path planning in cluttered issued Chinese patents and coau-
environments. Robotics 108, 13-27 (2018) thored 30 papers. His current
17. Liu, H., Zhang, X., Wen, J., et al.: Goal-biased Bidirectional RRT research interests include intel-
based on Curve-smoothing. IFAC-Papers Online 52(24), 255–260 ligent transportation and safety,
(2019) big data processing technology,
18. Li, Y., Xu, D., Zhou, C.: Cooperation path planning of dual-robot and cloud computing and inter-
based on self-adaptive stepsize RRT. Robot 42(5), 606–616 net of things.
(2020)
19. Yang, Z., Wang, Y., Qi, A.: Improved RRT* algorithm based
global obstacle avoidance planning for unmanned surface vehi-
cles. ShipSci. Technol. 41(23), 167–172 (2019)

13