Transportation Safety and Environment, 2023, 5: tdac061

DOI: 10.1093/tse/tdac061
Advance access publication date: 21 December 2022
Research Article

Path-planning algorithms for self-driving vehicles based

on improved RRT-Connect
Jin Li1 , *, Chaowei Huang1 and Minqiang Pan2
1
School of Intelligent Engineering, Shaoguan University, Shaoguan 512005, Guangdong, China;
2
The School of Mechanical and Automotive Engineering, South China University of Technology, Guangzhou 510640, Guangdong, China.
∗
Corresponding author. E-mail: [email protected]

Downloaded from https://academic.oup.com/tse/article/5/3/tdac061/6955827 by guest on 08 January 2025

Abstract
This study aims to solve path planning of intelligent vehicles in self-driving. In this study, an improved path-planning method com-
bining constraints of the environment and vehicle is proposed. The algorithm designs a reasonable path cost function, then uses a
heuristic guided search strategy to improve the speed and quality of path planning, and finally generates smooth and continuous cur-
vature paths based on the path post-processing method focusing on the requirements of path smoothness. A simulation test shows
that compared with the basic rapidly-exploring random tree (RRT), RRT-Connect and RRT∗ algorithms, the path length of the proposed
algorithm can be reduced by 19.7%, 29.3% and 1% respectively, and the maximum planned path curvature of the proposed algorithm
is 0.0796 m-1 and 0.1512 m-1 respectively, under the condition of a small amount of planning time. The algorithm can plan the more
suitable driving path for intelligent vehicles in a complex environment.

Keywords: path planning; rapidly-exploring random tree (RRT); double tree expansion; autonomous driving; curvature constraint

1. Introduction tion index of path cost. Each new node re-selects the parent node
and re-arranges the path line of the child node, so as to realize
As a key technology in the field of autonomous driving, path plan-
planning of the shortest asymptotic path.
ning plays an important role. Applied to autonomous vehicles,
Autonomous vehicles belong to nonholonomic kinematic con-
the path-planning algorithm is based on the motion planning
straints mobile robots, but existing algorithms such as the RRT,
of robots. Representative path-planning methods include the ar-
RRT-Connect and RRT∗ algorithms do not take vehicle kine-
tificial potential field method, the graph search-based method,
matic constraints into consideration when dealing with path-
the sampling-based method, the discrete numerical value-based
planning problems, so it is difficult for autonomous vehicles to
method, optimized methods and intelligent bionic algorithms [1–
move according to the path obtained by RRT algorithms. Focusing
10]. The sampling-based method does not need to process state
on the above problems, this paper proposes an improved path-
space in advance, so the search is fast, and it can effectively
planning method based on the RRT-Connect algorithm. Firstly,
solve the problem of motion planning in a complex environ-
considering the vehicle steering motion constraint, a cost func-
ment, the most representative of which is the rapidly-exploring
tion which considers path angle and path length is proposed.
random tree (RRT) algorithm. The RRT algorithm uses a start-
Secondly, in order to speed up path planning and improve path
ing point as the root node of the tree, continuously simulates
safety, an adaptive dynamic step search method and adaptive
tree growth with a certain step size and gradually explores the
obstacle expansion method are proposed. Finally, the path post-
state space until it connects to the target point [11]. However,
processing method is used to obtain the planned path that
due to the randomness of sampling, the RRT algorithm exhibits
satisfies the steering curvature constraint of the autonomous
blindness during the process of exploration, leading to a tor-
vehicle.
tuous and jittery path. Meanwhile, due to its insensitivity to
map information, shortcomings are particularly obvious in com-
plex maze maps or slit channels. And it is easy to cause local 2. Kinematics model of the vehicle
shock. Key to a kinematic model of an autonomous vehicle is the con-
Focusing on the defects of the above basic RRT algorithm, the straint of the steering angle of the vehicle, not the lateral mo-
researchers proposed relevant improvement methods. Among the tion of the vehicle. At the same time, the speed-to-steering limita-
methods based on the essential attributes of RRT, the most in- tion and vehicle dynamics requirements will be ignored, and the
fluential are RRT-Connect and RRT∗. The RRT-Connect algorithm model should be built at the level of vehicle kinematics. A kine-
adopts bidirectional random tree growth. In each iteration, the matic model of the autonomous vehicle is established, as shown
tree growth of one side always expands towards the other side in Fig. 1, which is simplified to a two-wheel model.
alternately, and the greedy growth strategy is added to node ex- Among them, in the inertial coordinate system, (Xf , Yf ) is the
pansion to achieve rapid exploration which greatly reduces com- centre coordinate of front axle, (Xr , Yr ) is the centre coordinates
puting time. Based on RRT, the RRT∗ algorithm uses the evalua- of the rear axle, δf is the front wheel slip angle, ϕ is the head-

Received: June 7, 2022. Revised: September 20, 2022. Accepted: October 5, 2022
© The Author(s) 2022. Published by Oxford University Press on behalf of Central South University Press. This is an Open Access article distributed under the
terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and
reproduction in any medium, provided the original work is properly cited.
2 | Transportation Safety and Environment, 2023, Vol. 5: No. 3

Fig. 2. Schematic diagram of parent node selection of new node.

In addition, for path planning of autonomous vehicles, it is nec-

essary to safely complete various obstacle avoidance actions, and
Fig. 1. Vehicle kinematics model.
quickly plan to meet real-time requirements. The path planned

Downloaded from https://academic.oup.com/tse/article/5/3/tdac061/6955827 by guest on 08 January 2025

should be stable, economical or optimal.

ing angle of the body, R is the rear wheel turning radius, L is the
wheelbase and P is the instantaneous rotation centre of the ve- 3. Improved RRT algorithm
hicle. It is assumed that the vehicle’s centroid slip angle is con-
3.1 Improvements of random tree exploring
stant throughout the steering process, namely, the vehicle’s in-
strategies
stantaneous steering radius is consistent with the road curvature
radius.
3.1.1 Main structure of improved RRT
According to the geometric relationship of the vehicle, there RRT-Connect has advantages of bi-directional exploration and
are: greedy strategy guidance, so it has high exploration efficiency.
 In the path formed from it, the number of straight line seg-
Xf = Xr + l cos ϕ
(1) ments is directly related to the complexity of the map, having
Yf = Yr + l sin ϕ
 certain similarity with vehicle motion. Therefore, the main struc-
vf = X˙ f cos (ϕ + δf ) + Y˙ f sin (ϕ + δf ) ture of bidirectional co-exploration and greedy strategy guidance
(2)
vr = X˙ r cos ϕ + Y˙ r sin ϕ in RRT-Connect is selected. Considering that direct connect can
From formula (1), the following can be obtained: be achieved between the starting point and target point under
 barrier-free conditions, a judgement can be made in the first
X˙ f = X˙ r − l · ϕ˙ · sin ϕ growth stage of the algorithm. If the condition is satisfied, enter
(3)
Y˙ f = Y˙ r + l · ϕ˙ · cos ϕ post-processing directly to reduce unnecessary growth calcula-
When ignoring sideslip, kinematic constraints of front and rear tion.
axles are:
 3.1.2 Cost function
X˙ f sin (ϕ + δf ) − Y˙ f cos (ϕ + δf ) = 0
(4) In the iterative process, RRT and its variants select parent nodes
X˙ r sin ϕ − Y˙ r cos ϕ = 0
through the cost function. In the above algorithm, the cost func-
Combining formulas (2) and (4), the following can be obtained: tion takes the minimum Euclidean distance as the reference in-
 dex, as shown in formula (10), causing inoperability in the planned
X˙ r = vr cos ϕ
(5) path and going against requirements of vehicle kinematics.
Y˙ r = vr sin ϕ

Combining formulas (1), (3) and (5), yaw angular velocity ω can C(qi , qi+1 ) = (xi − xi+1 )2 + (yi − yi+1 )2 (10)
be obtained:
vr where, qi and qi+ 1 are the nodes on a random tree, and (xi , yi ), (xi + 1 ,
ω = ϕ˙ = tan δf (6) yi + 1 ) are the spatial coordinates of the two, respectively.
l
The following can be obtained from yaw rate ω and vehicle As shown in Fig. 2, after random point of growth qrand , if the Eu-
speed vr : clidean distance is compared, qi node is naturally selected. But in
 fact, if the degree of tortuosity of the connection path of the ran-
R = vωr dom tree is taken into account, due to θ j < θi , qj is a better choice,
(7)
δf = arctan R1 which makes the path more smooth [18].
From formulas (1) and (7), the kinematic model can be obtained Therefore, when selecting parent nodes, distance and angle
as: need to be taken into account at the same time. In Ref. [12],
⎡ ⎤ ⎡ ⎤ Du et al. proposed a metric function that considers vector an-
X˙ r cos ϕ
⎢˙ ⎥ ⎢ ⎥ gle and Euclidean distance at the same time, and normalizes
⎣Yr ⎦ = ⎣sin ϕ ⎦ vr (8)
two dimensions of distance and angle in metric function C(qi , q j ).
ϕ˙ tan δf /l
After being normalized and adding weights to be combined,
The curvature formula is: where the weights are each 0.5, finally the output position of
1 the peak of the metric function is found. In practice, an accu-
k= (9)
R rate calculation of angles in the random tree is a great bur-
when the vehicle is turning, the steering curvature should be den, which will lead to an exponential increase of the calcula-
smaller than the maximum curvature limit constrained by the tion time, and at the same time, outward exploration of new
minimum steering radius of the vehicle, so that the path follow- nodes becomes lazy, and the advantages of RRT are no longer
ing can be satisfied. obvious.
 Transportation Safety and Environment | 3

Table 1. Schematic representation of processing results of metric function.

Grip generated between qrand and parent node Processing result

Non-obtuse angle Select parent nodes based on minimum Euclidean distance (consistent
with basic RRT mode).
Both obtuse and non-obtuse angles exist Among parent nodes with obtuse angle, select parent node using
minimum Euclidean distance as standard.
All obtuse angle Select parent nodes based on minimum Euclidean distance (consistent
with basic RRT mode).

Downloaded from https://academic.oup.com/tse/article/5/3/tdac061/6955827 by guest on 08 January 2025

Fig. 3. Schematic diagram of random tree growth with improved dynamic step length.

Combining quantities of distance and angle, the following met- 3.1.3 Adaptive dynamic step length
ric function T (qi , qrand ) is put forward: Step length is the distance between a node and its growing child
nodes, which would affect calculation time and quality of planned
T (qi , qrand ) = D(qi , qrand ) + A(qi , qrand ) path. The smaller the step size, the more likely it is to jump over
⎧ obstacles to converge and the shorter the planned path may be,
⎪ D(qi , qrand ) = (xi − xrand )2 + (yi − yrand )2
⎪
⎪
⎪ 2
⎪D(qi , qk ) = (xi − xk ) + (yi − yk )
2 but the greater the cost of time [13–16]. Therefore, the adaptive
⎨ (11)
D(qk , qrand ) = (xk − xrand )2 + (yk − yrand )2 dynamic step length is proposed to balance exploration efficiency
⎪
⎪ and obstacle avoidance ability.
⎪
⎪ 0, D(qi , qrand ) + D(qi , qk ) < D(qk , qrand )
⎪
⎩A(qi , qrand ) =
1, others In RRT-Connect, the step length p of the double random tree
exploration is adjusted as follows: when one of the random tree
growth nodes explores a new node, the movement step length ex-
where, qi is a node in the matrix of the candidate parent nodes in pands a new node qnew with unit step length p. The other tree ex-
the random tree, qk is the parent node of qi , qrand is the random pands a new node q’new with a unit step size p. At this time, if
node generated before new nodes are generated in the next step, the connection between new node of the tree and its parent node
and cost function T (qi , qrand ) is determined by the square term of does not encounter obstacles and is not connected to the random
the Euclidean distance D(qi , qrand ) and angle control of the obtuse tree of the other party, the step length of the greedy expansion
angle A(qi , qrand ). The node with minimum value is selected as the in the next step will be 2p until it exits the loop when hitting the
parent node. The distance calculation function adopts the square obstacle. The improved dynamic step length method is shown in
term of the Euclidean distance, using positive correlation of the Fig. 3.
square root operation of the natural numbers, and directly remov- Establishment of this adaptive dynamic step size has the fol-
ing the square root operation, thus the calculation amount is re- lowing advantages:
duced. The angle control function uses the square relationship
of three sides of triangle, bounded by the Pythagorean theorem 1) Step length is only changed in the greedy straight line seg-
under right angles. If the square term of two adjacent sides is less ment, retaining exploration characteristics of the original
than that of the opposite side, it is an obtuse angle. Take the obtuse RRT, which will speed up the exploration speed to a certain
angle as priority, assign A(qi , qrand )=0, and assign 1 in other cases. extent and reduce the number of nodes.
In the tree node expansion, the parent node matrix (excluding 2) When approaching the obstacle with a step size of 2p, it is
starting point) of T (qi , qrand ) has the following possibilities show as equivalent to being able to perceive the obstacle at a faster
Table 1. speed. At the same time, compared with RRT of fixed step
The purpose of the re-proposed cost function is to use a sim- length, it is more likely that the node will eventually be far-
ple and effective function to guide random tree to expand for- ther away from the obstacle, which is beneficial to improve
ward in the direction of the obtuse angle. Compared with con- the smoothness of the high trajectory path.
trol of precise calculation of angle, this method of classifying the
angle of range greatly reduces the amount of calculation, retain- 3.1.4 Expansion of adaptive barrier block
ing the rapid expansion capability of RRT, and planning time does Due to the randomness of the trajectory generated by the RRT al-
not increase significantly compared with the original. Though the gorithm, there is a high probability of collision at the corner of the
planned path is still not optimal, obtuse sampling of new nodes maximum curvature of the path during the path post-processing,
has been optimized for smoother paths. as shown in Fig. 4. Thus, a supplementary expansion mechanism
4 | Transportation Safety and Environment, 2023, Vol. 5: No. 3

Fig. 4. Schematic diagram of collision.

Fig. 7. Schematic diagram of pruning with maximum curvature

constraint.

Downloaded from https://academic.oup.com/tse/article/5/3/tdac061/6955827 by guest on 08 January 2025

Fig. 5. Schematic diagram of path after supplemental expansion.

Fig. 8. Schematic diagram of key points extracted by the DP algorithm.

3.2 Post-processing of planned paths

3.2.1 Extraction of way-points
Due to the random growth of the tree shape of the RRT algorithm,
despite the guidance of the greedy strategy, there exists unnec-
essary commutation in the final planned path, causing the for-
mation of redundant and frequently changing paths. Way-points
need to be further extracted to meet the economic requirements
of the lower path cost and the comfort requirements of fewer di-
Fig. 6. Schematic diagram of distance judgement between obstacle
blocks. rection changes. The maximum curvature limit of vehicle steering
should be considered when extracting nodes. Therefore, a post-
processing operation of tree nodes that considers the above is re-
quired.
of adaptive obstacle blocks is put forward. Under the premise of Du et al. [12] proposed a pruning function based on the maxi-
safety, the planning of a better path should be done as much as mum curvature constraint, shown in Figs. 7 and 8. The first step is
possible, as shown in Fig. 5. to prune the path, reaching the goal of the simplest path; the sec-
The supplementary expansion method of obstacle blocks is ond step is to eliminate the sharp angles which is done by taking
shown in Fig. 6: take the centre point of the obstacle blocks as the a point one unit ahead of the current path point.
centre of the circle, make the envelope circle of obstacle blocks, Based on the maximum curvature constraint, pruning function
define the minimum distance between the envelope circles be- processing method simplifies the path to the greatest extent un-
tween the obstacle blocks as L2 and define the minimum allow- der the requirement of obstacle avoidance, and at the the same
able passing distance between the obstacle blocks as Lmin . If the time repairs the sharp included angle of the path in the form of
minimum distance L2 between the obstacle blocks is greater than supplementing adjacent nodes, which reduces the probability of
Lmin , the boundary expansion of the rectangle can be carried out, forming excessive path curvature to a certain extent. However,
that is, the length and width are expanded by a at the same time, path points formed under random tree are in the point set near
otherwise keep the original size. Under the reasonable Lmin limit, the sharp angle, and the sharp angle may not be repaired. For ex-
the expanded obstacle blocks repel the random tree, so as to avoid ample, when the more suitable points are not within a unit or the
collision at the corner of the maximum curvature, and guide and optional point set is an empty one.
plan a better path as far as possible on the premise of safety. In addition, Jiang et al. [17] proposed to use DP (Douglas-
This adaptive barrier block supplements the expansion. The Peucker) algorithm for extraction of key landmarks for planned
judgement circle formed by rough classification is used to make paths, which keeps looking for the farthest chord distance be-
the judgement, which is a simplified calculation of maximum tween two points, and stops when distance to the farthest point
safety bias. Obstacle block expansion is performed when it strictly is not greater than the threshold. The process of DP algorithm is
complies with restriction equation. Otherwise, to preserve original shown in Fig. 8. Extracting key road signs using DP can retain main
solvable convergence, no changes will be made. Within the rea- heading of the original path to a certain extent, and at the same
sonable Lmin limit, the expanded obstacle block repels the random time remove some redundant points. Its key lies in the selection of
tree to avoid collisions at the corners with maximum curvature, threshold related to the complexity of environment, only optimal
and guide planning of better paths as much as possible under the extraction path can be obtained when threshold can be changed
premise of safety. temporally and dynamically.
 Transportation Safety and Environment | 5

where, i is the serial number of the judgement point, i-1 is the

serial number of the previous point from the judgement point, i+1
is the serial number of the next point from the judgement point.
Detect from one after the starting point to one before the target
point, if function Angle(i ) outputs 1, the above-mentioned second
step for this point is performed.
The re-proposed heading angle constraint and path key point
Fig. 9. Schematic diagram of supplementary nodes. extraction method with low path cost orientation continues the
idea of the pruning function based on maximum curvature con-
straint which maximizes the simplification of path and removal of
redundant points. Secondly, nodes are supplemented from a wider
range and better standard to remove sharp angles more reason-
ably. Therefore, balance is achieved between curvature constraint
on heading angle and low path cost to a certain extent.

Downloaded from https://academic.oup.com/tse/article/5/3/tdac061/6955827 by guest on 08 January 2025

3.2.2 Curve fitting of way-points
As one of the mainstream methods for descriptive shape mathe-
Fig. 10. Schematic diagram of final planning result of new pruning matics, the B-spline function has functions of describing and de-
function. signing continuous free-form curves and surfaces [18, 19], which
can be used to fit a curve with continuous curvature to meet the
Both the above have their unique advantages for path process- trajectory requirements of vehicle motion.
ing. Combining advantages of both, an innovative way-point ex- The B-spline equation is expressed as:
traction method with heading angle constraints and low path cost n

orientation is proposed. The first step is the same as that of the p(u ) = pi Ni,k (u ) (13)
i=0
pruning function based on the maximum curvature constraint,
and the pruning function is used to remove redundant points to where, pi is the coordinate of the control point, which determines
the greatest extent. In the valid path obtained after random tree the control polygon of the B-spline; i is the node number; Ni,k is the
finishes growing, from the starting node, define subsequent nodes normative B-spline basis function of number k, with the highest
as path nodes from target node to current node, and then connect number k. The basis function is a kth-order segmented polyno-
the current node to subsequent nodes in turn. Nodes with later mial consisting of the parameters u (u1 ≤ u2 ≤ ... ≤ un+k+2 ) in a se-
serial numbers are connected preferentially until there is no col- quence of non-decreasing vectors U. From recursion formula Cox-
lision between connection line and obstacle node, and so on, until de Boor, the following can be obtained:
⎧
it is connected to the target node. ⎪
⎨ 1, ui ≤ u ≤ ui+1
⎪
,k = 0
In the second step, path nodes are supplemented to eliminate 0,
Ni,k (u )= others (14)
the sharp included angle of the path, as shown in Fig. 9, and node ⎪
⎪ ui+k+1 −u
⎩ uu−u i
Ni,k −1 ( u ) + Ni +1 ,k −1 ( u ) , others
i+k −ui u −ui+1
selection is carried out on the basis of point-line distance from the i+k+1

DP algorithm. The algorithm is as follows: detect whether there where the denominator term ui+k − ui =0 or ui+k+1 − ui+1 =0, then
is an angle smaller than the set maximum limit angle αmin in define denominator value to be 1.
the connection path formed by the pruning function processing Generation rules of node vector sequence U are as follows:
in the previous step. If the path exists, define the furthest path ⎧
⎪
⎨0 , i = 1 : (1 + k )
point from the path segment qi -qi-1 as qt , then connect qi -qt and qt -
ui = uk+i−1 + n−k1+1 , i = others (15)
qi-1 in sequence, and check whether connection lines collide and ⎪
⎩1 , i = (n + 2 ) : (n + 2 + k )
whether angles  qi+1 qi qt and  qi qt qi−1 meet the maximum angle
requirements. If it is satisfied, supplementary path node qt enters Among them, there are a total of n + 1 control points, and a
the key point of the path. If there is no path that satisfies the con- total of n + k + 2 node vectors are generated. Nodes at both ends
dition, the algorithm tries to define the node closer to the line seg- of the node vector have a degree of repetition k + 1, namely: u1 =
ment qi -qi -1 as qt . If none of the node between qi -qi-1 satisfies the u2 = ... = uk+1 , un+2 = un+3 = ... = un+k+2 . Nodes in the middle are
condition, it tries to extract the node qt between qi -qi+1 . The final incremented point by point with tolerance n−k1+1 as arithmetic
planning result of the same schematic diagram is shown in Fig. 10, progression. The example of way-points curve fitting is shown
q1 is added between q3 and q0 , so curvature of the final fitted curve inFig. 11.
changes less. However, in special cases, in the process of supple- Take number k of fitting curve as 2, because nodes at both ends
menting nodes, the final set of selectable points is an empty one. have a lower repetition rate by one unit, and there are more node
Since new nodes regenerated are not supplemented, the second vectors increasing in the middle, which makes it closer to the
step is exited and the result of original first step is retained. control point and can improve planning, and maximum steering
Definition of maximum limit angle αmin =90◦ herein is based on curvature requirements are given to optimize control points de-
the square term of Euclidean distance, as shown in the following scribed later.
formula.
3.2.3 Control way-point supplement
⎧
⎪a = (xi − xi+1 )2 + (yi − yi+1 )2 Processed using the pruning function, path points become sparse,
⎪
⎪
⎪ 2 2
⎨b = (xi − xi−1 ) + (yi − yi−1 )
⎪ especially in segments with large turning. Fitting collision will oc-
xi−1 ) + (yi+1 − yi−1 )2
c = (xi+1 −  2
(12) cur due to the lack of control points, as shown in Fig. 12. If node
⎪
⎪
⎪
⎪ 0, a + b < c coordinates are (0,30), (0,0) and (30,0), the fitting result of the B-
⎪
⎩Angle(i ) =
1, others spline function described in Section 3.2.2 is the blue line, lead-
6 | Transportation Safety and Environment, 2023, Vol. 5: No. 3

tance multiple N and total number j of control points.

⎧
⎪
⎪di = (xi+1 − xi )2 + (yi+1 − yi )2
⎪
⎨j = 1 − 1
N     (16)
⎪
⎪ p(k ) = (1 − N) · xi + N · xi+1
⎪
⎩
yi yi+1

After several experiments, in the environment of grid density

100 × 100, considering maximum curvature constraint, smooth-
ness of trajectory and safe passing rate of fitted trajectory, dis-
tance between points 4 × 4 is used as the boundary, threshold δ is
defined as 5.66 and distance multiple N is 0.5.

4. Simulation verification and analysis

Downloaded from https://academic.oup.com/tse/article/5/3/tdac061/6955827 by guest on 08 January 2025

4.1 Setup of simulation experiments
Fig. 11. Schematic diagram of curve fitting. Parameters of simulated vehicle are shown in Table 2.
When defined, the basic step size of the basic RRT and RRT-
Connect is the same as that of the improved RRT algorithm here,
which is 5 m, while the RRT∗ is not comparable with the other
three algorithms in terms of time cost, being the optimal orienta-
tion of path asymptotics. So the basic step length of RRT∗ is 2 m,
and the role is to reconnect the parent node and child node with
the exploration radius of 15 m. In addition, other variables that are
not related to growth strategy, such as environmental variables,
are controlled consistently.
The simulation is as follows: in environment maps under four
road conditions of straight overtaking and a complex parking
lot, the basic RRT, RRT-Connect, RRT∗ and improved RRT al-
gorithms here are calculated respectively, and then images are
output for planned path comparison, and curvature informa-
tion of the improved RRT algorithm is recorded. And operate 20
times in complex parking lot conditions, record operation time,
number of random tree nodes and final calculated path length
respectively.
Fig. 12. Schematic diagram of collision after fitting.
4.2 Results of simulation experiments
1) Simulation of road conditions for overtaking
Table 2. Simulation test vehicle parameters.
In Fig. 13, (a) is the planning results of the basic RRT (blue
Parameter Value line), (b) is the planning results of the RRT-Connect (blue-red
line), (c) is the planning results of the RRT∗ (blue line), (d) is the
Length/m 4.30
Width/m 1.80 schematic diagram of the growth of the improved RRT-Connect
Wheelbase/m 2.51 in the sampling phase, (e) is the schematic diagram of the path
Minimum urning radius/m 6.0 processed by the post-processing stage of the improved RRT, (f)
Maximum curvature/m−1 0.167 is the final planning result after fitting of improved RRT-Connect.
Maximum steering angle/◦ 30 As shown in Fig. 14, the peak curvature of the planned path is
0.0796 m−1 .

2) Simulation of complex parking lot road conditions

ing to a collision. However, if nodes (0,15) and (15,0) are added,
the fitting result is the red line and the obstacle will be avoided. In Fig. 15, (a) is the planning results of the basic RRT (blue line),
Therefore, it is necessary to supplement control points between (b) is the planning results of the RRT-Connect (blue-red line), (c)
straight line segments to control the curvature, but if number of is the planning results of the RRT∗ (blue line), (d) is a schematic
control points is increased in individual short path segments, the diagram of the growth of the improved RRT-Connect in the sam-
curvature will be large. pling phase, (e) is a schematic diagram of the path processed
To achieve this, a control way-point supplement of parame- by the post-processing stage of the improved RRT, (f) is the fi-
terized conditional linear interpolation is proposed, using path nal planning result after fitting of the improved RRT-Connect.
length as judgement and linear interpolation for supplemented As shown in Fig. 16, the peak curvature of the planned path is
control points appropriately. The steps are as follows: 0.1512 m−1 . The simulation results are shown in Figs. 17–20 and
Define distance threshold δ, and calculate square di of the Eu- Table 3.
clidean distance between the nodes processed using the pruning
function in the previous step. If it is larger than or equal to the 4.3 Discuss of simulation results
square of the threshold, insert control point p(k ), serial number In the sampling stage of the random tree growth, the cost func-
k ⊂ {1 : j} in the middle of the two key points of the path with dis- tion re-proposed of exploring parent nodes makes sampling suc-
 Transportation Safety and Environment | 7

Downloaded from https://academic.oup.com/tse/article/5/3/tdac061/6955827 by guest on 08 January 2025

Fig. 13. Planning results of four algorithms under road condition map for straight overtaking: (a) RRT; (b) RRT-Connect; (c) RRT∗; (d) improved
RRT-Connect; (e) paths after the extraction of way-points; (f) optimized path.

Fig. 14. Improved RRT planning path curvature change (normalized path length).

cessfully inclined to the obtuse angle and the change of angle can overcomes blindness of exploration of the RRT algorithm. Mean-
be clearly seen compared with the original random tree of RRT- while, the number of random tree nodes is close to RRT-Connect.
Connect. Secondly, the adaptive dynamic step length and adap- Due to the adaptive dynamic step length strategy, the number
tive barrier block supplement expansion strategy can make the of nodes is further reduced. However, due to the addition of the
random tree far away from the barrier to a certain extent, approx- post-processing operation of road condition planning, the plan-
imating a safer path. In the post-processing stage of the planned ning time is increased by 0.03 s compared with RRT-Connect and
path, the way-point extraction strategy that considers both the 0.16 s compared with basic RRT. The increased time has little im-
heading angle constraint and the low path cost orientation re- pact in practical application, but the pruning function process-
duces unnecessary tortuous paths. Redundant way-points are re- ing of the post-processing operation of the improved RRT-Connect
moved, while the obtuse angle of the path is preserved. In this way, algorithm greatly reduces the path cost. The simulation results
sudden increase of curvature is limited to a certain extent, and show that the overall performance of the improved RRT-Connect
finally the curve fitting of the B-spline enables a continuous cur- algorithm is better than that of RRT, RRT-Connect and RRT∗ algo-
vature path to be realized. Compared with RRT, RRT-Connect and rithms. The obtained path meets the steering curvature constraint
RRT∗ algorithm, paths planned by improved RRT-Connect can be of autonomous vehicles, and the planned path is easier to be ex-
directly executed and meet requirements of smooth driving and ecuted by vehicles.
safe obstacle avoidance.
In case of straight overtaking road conditions and complex
parking lot conditions, since the interval of the convergence re- 5. Conclusions and future works
gion is wide enough, the curvature of the planned path satisfies
5.1 Conclusions
the limit curvature of the simulated vehicle, and the planned path
Focusing on the problem of path planning for autonomous ve-
is almost in line with the ideal path. From the simulation graph
hicles, this study proposes an improved path-planning algorithm
under complex parking lot conditions, it can be seen that the
based on RRT-Connect. The algorithm takes environmental con-
improved RRT-Connect algorithm inherits characteristics of at-
straints and dynamic constraints of intelligent vehicles into ac-
traction growth between two random trees of RRT-Connect, and
count at the same time. Firstly, a reasonable path cost function is
8 | Transportation Safety and Environment, 2023, Vol. 5: No. 3

Downloaded from https://academic.oup.com/tse/article/5/3/tdac061/6955827 by guest on 08 January 2025

Fig. 15. Planning results of four algorithms under complex parking lot road map: (a) RRT; (b) RRT-Connect; (c) RRT∗; (d) improved RRT-Connect ; (e)
paths after the extraction of way-points; (f) optimized path.

Fig. 16. Curvature change of improved RRT planning path (normalized

path length).

Fig. 18. RRT-Connect simulation iteration results.

Fig. 17. Basic RRT simulation iteration result graph.

Fig. 19. RRT∗ simulation iteration results.
 Transportation Safety and Environment | 9

tion (Grant No. 2018KQNCX207), Shaoguan science and technol-

ogy plan (Grant No. 2018sn043) and Shaoguan university scientific
research (Grant No. sz2018KJ06).

Conflict of interest statement

The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to
influence the work reported in this paper.

References
Fig. 20. Improved RRT-Connect simulation iteration results. 1. Yu Zhuo-ping, Li Yi-shan, Xion G Lu. A review of motion plan-

Downloaded from https://academic.oup.com/tse/article/5/3/tdac061/6955827 by guest on 08 January 2025

ning problem of autonomous vehicle [J]. Journal of Tongji Univer-
Table 3. Four RRT simulation results. sity(Natural Science) 2017;45:1150–9.
2. Lee J, Nam Y Y, Hong S J. Random force based algorithm for
Average Average
Average path local minima escape of potential filed method[C]. 11th Interna-
Method number of calculation
length/m tional Conference on Control Automation Robotics & Vision. Singapore:
nodes time/s
Nanyang Technological University, 2010;827–32.
Basic RRT 101.6000 125.3047 0.0180 3. An Lin-fang, Chen Tao, Cheng Ai-guo, et al. A simulation on path
RRT-Connect 36.4500 129.9120 0.0040 planning of intelligent vehicles based on artificial potential field
RRT∗ 450 101.8644 0.7169 algorithm [J]. Automotive Engineering 2017;39:1451–6.
Improved RRT-Connect 36.1500 100.6166 0.0340
4. Yang Xue-yan, Qian Yu-bin. Research on obstacle avoidance
path planning of vehicle based on improved artificial potential
designed. Secondly, the heuristic search idea is added on the basis field method [J/OL]. Agricultural Equipment & Vehicle Engineering
of the RRT-Connect algorithm to reduce the randomness of the 2020;1–7.
path search. Then, the post-processing method of the path is put 5. BoroujenI Z, Goehring D, Ulbrich F, et al. Flexible unit A-star tra-
forward, and the path point extraction method is proposed con- jectory planning for autonomous vehicles on structured road
sidering the requirements of path smoothness. Finally, combined maps[C]. IEEE International Conference on Vehicular Electronics &
with the B-spline curve fitting method and curve control point in- Safety, IEEE, 2017.
terpolation method, the executable path is obtained. Simulation 6. Majumder S, Prasad M S. Three dimensional D ∗ algorithm for in-
tests in overtaking and parking scenarios show that path length cremental path planning in uncooperative environment[C]. 2016
can be reduced by 19.7%, 29.3% and 1% respectively, compared 3rd International Conference on Signal Processing & Integrated Net-
to algorithms based on RRT, RRT-Connect and RRT∗. The max- works, IEEE, 2016.
imum path curvature planned by algorithm here is 0.0796 m−1 7. QI Yao, Xu You-chun, LI Hua, et al. Hybrid A∗algorithm-based
and 0.1512 m−1 respectively, both meet the vehicle dynamics con- path planning for intelligent vehicle [J]. Journal of Academy of Mil-
straints. In the complex parking environment, the average plan- itary Transportation 2018;20:85–90.
ning time of the improved RRT-Connect algorithm is only 0.03 s 8. Kavraki L, KolountzakiS M N, Latombe J. Analysis of probabilistic
longer than that of RRT-Connect and 0.16 s longer than that of roadmaps for path planning[C] //Robotics and Automation, IEEE
the basic RRT. Although the improved RT-Connect algorithm in- International Conference on. Minneapolis: IEEE, 1996;4:3020–5.
creases the planning time a little, it greatly improves the quality 9. Li X, Sun Z, Cao D, et al. Real-time trajectory planning for au-
of the planned path. Simulation results show that the improved tonomous urban driving: framework, algorithms, and verifica-
RRT-Connect algorithm has better overall performance than the tions[J]. IEEE/ASME Trans Mechatron 21:740–53.
RRT, RRT-Connect and RRT∗ algorithms in complex environments. 10. Zhu Ying-jie, Li Jian-shi, Feng Ming-yue, et al. Autonomous park-
It can be used to plan more suitable driving paths for autonomous ing and path planning algorithm for intelligent vehicles based
vehicles. on numerical optimization [J]. Journal of Military Transportation In-
stitute 2019;21: 84–89.
5.2 Further works 11. LaValle SM. Rapidly-exploring Random Trees: a new tool for path
planning[J]. Algorithmic & Computational Robotics New Directions,
Without considering the influence of the speed of vehicle dynam-
1998;293–308.
ics, path planning algorithm proposed in this paper is difficult to
12. Du Ming-bo, Mei Tao, Chen Jia-jia, et al. RRT-based motion plan-
apply on high speed car. In future work, we will further consider
ning algorithm for intelligent vehicle in complex environments
the vehicle motion state to design the cost function and improve
[J]. Robot 2015;37:443–50.
the applicability of the algorithm for high-speed moving vehi-
13. Liu En-hai, GaO Wen-bin, Kong Rui-ping, et al. Improved RRT
cles. In addition, this paper assumes that the obstacles are static,
path planning algorithm [J]. Computer Engineering and Design
which lacks certain adaptability to complex dynamic scenes. Fu-
2019;40:2253–8.
ture work will increase the optimization of dynamic scenes by al-
14. Zhu Jin-hui, Liang Ming-jie, Liang Ying-ju, et al. Adaptive
gorithm.
weighted Rapidly-exploring Random Tree algorithm [J]. Computer
Engineering 2010;36:16–18.
Funding 15. Sun Qin-peng, Li Meng, Wang Zhong-hua. Mobile mobot path
planning based on improved Rapidly-exploring Random Tree al-
Supported by Guangdong Natural Science Foundation (Grant No.
gorithm [J]. Journal of University of Jinan (Science and Technology)
2015A030310411), Guangdong University Characteristic Innova-
2019;33:431–8.
10 | Transportation Safety and Environment, 2023, Vol. 5: No. 3

16. Wang Dao-wei, Zhu Ming-fu, Liu Hui. Rapidly-exploring Random 18. Elbanhawi M, Simic M, Jazar R. Randomized bidirectional b-
Tree algorithm based on dynamic step [J]. Computer Technology spline parameterization motion planning[J]. IEEE Trans Intell
and Development 2016;26:105–7 112. Transp Syst 2016;17:406–19.
17. Jiang Yuan-yuan, Tao De-jun, Shi Mei-le, et al. Path smoothing 19. Gómez-Bravo F, Cuesta F, Ollero A, et al. Continuous curva-
algorithm of mobile robot based on DP-B spline [J/OL]. Mechanical ture path generation based on β -spline curves for parking ma-
Science and Technology for Aerospace Engineering 1-7[2019-10-09]. noeuvres[J]. Rob Autom Syst 2008;56:360–72.

Downloaded from https://academic.oup.com/tse/article/5/3/tdac061/6955827 by guest on 08 January 2025

Received: June 7, 2022. Revised: September 20, 2022. Accepted: October 5, 2022
© The Author(s) 2022. Published by Oxford University Press on behalf of Central South University Press. This is an Open Access article distributed under the terms of the Creative Commons Attribution
License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.