1. Introduction
Modern developments in the domain of autonomous vehicle (AV) motion planning are distinguished by a wide range of proposed techniques and methods having advantages and drawbacks depending on the forecasting goals under current conditions. The main planning challenge concerns the need for simultaneously satisfying multiple requirements that guarantee both the motion feasibility and its safety, including the controllability and stability within the subsequent tracking process.
The basic planning task consists of generating a forecast of trajectory and speed mode. On the one hand, it must conform to the principles of simplicity, smoothness, and unambiguity. For these purposes, analytical functions of polynomial and trigonometric types are often used [
1], which change shape depending on parameters. The advantage of this approach implies that, with known parameters, any trajectory’s point may be directly computed. However, there are also problems associated with the fact that trajectories represented by polynomials of small degrees do not provide high differentiability. For large degrees, the polynomials may be unstable, and the sensitivity of DOFs corresponding to higher derivatives increases. This complicates the optimization procedure since the determined parameters, in this case, may differ by an order of magnitude.
On the other hand, the trajectory curve must meet the requirements of continuity and differentiability. Since the forecast optimality criteria themselves depend on the trajectory curvature’s derivatives, it is also necessary to ensure the equality of these derivatives at the point of changing planning cycles. In this case, most often it is impossible to ensure the continuity of higher speed derivatives, as well as kinematic parameters based on them such as yaw rate, angular acceleration, and jerks. In most of the existing techniques, these aspects are neglected, which affects the irregular and intermittent nature [
2] of the listed parameters at transient points. However, these parameters are associated with vehicle control systems and actuators. Therefore, it is desirable to ensure the continuity of their functions and derivatives for transient processes.
To guide the AV on a roadway, two approaches may be used. The first one involves the formation of roadway boundaries and motion zones in the Cartesian coordinate system [
3]. In this case, a space available for predicting the AV movement may be obtained by limiting motion zones for other traffic participants in such a way that a minimum safe distance is guaranteed. However, most often in this case, there is a need for separating the planning stage into sequential phases of determining the trajectory and distributing the speed mode. Another approach allows for the consideration of the formation of AV motion in Frenet coordinates [
4,
5]. In this case, usually, the midline of a road lane is in some way expressed as a reference path function. Thus, instead of Cartesian coordinates, two Frenet coordinates arise, such as the lane reference curve’s length and the lateral displacement perpendicular to the reference curve’s tangent. In this case, it is much easier to work with spatial constraints by imposing them on the lateral displacement within the boundaries of a road lane. However, we still need to make the transition to the AV’s local Cartesian system to decompose the components of kinematic parameters into longitudinal and transversal ones. Due to curvature, the full representation of the motion model in curvilinear coordinates becomes more complicated. Therefore, in this work, we will use a mixed approach in which there is a reference path curve in the Frenet coordinates, and the trajectory is predicted in the Cartesian coordinates.
Forecasting the speed mode [
6] involves obtaining several kinematic parameters (linear and angular speeds, linear and angular accelerations, linear jerks) characterizing the AV dynamics and providing direct ties with vehicle sensors’ data (accelerometers, gyroscopes, radars), which allows for the monitoring and correcting of the AV control at the tracking stage. These same parameters may be used as criteria for optimizing the AV motion. The main task of speed planning aims at compromising the vehicle’s maximum performance, safety, and motion smoothness.
To implement the above, two different approaches may be used. The first one is represented by discrete modeling [
7], where a system of equations based on vehicle dynamics (or kinematics) allows simultaneous optimization [
8] of the trajectory and speed related differentially to each other. However, to ensure good convergence of the numerical solution, it is necessary to reduce the integration step, which increases iteration time. Another approach proceeds from representing the trajectory and speed by piecewise analytical functions [
9], forms of which are known in advance and only the parameters that determine the nature of these functions are optimized. In this case, there are no problems with step control, but the motion can only be represented by a vehicle kinematic model.
Constraints are a separate issue. There are quite a number of restrictions stipulated by both the parameter nature and the possibility of its assessment. Thus, geometric restrictions [
10] imply a non-violation of the roadway boundaries (permissible zone) and keep safe positions relative to other traffic participants. Kinematic constraints concern parameters based on speed [
11] and its derivatives. Dynamic restrictions are determined by limits imposed on the vehicle’s powertrain system to realize its traction potential. The vehicle acceleration naturally decreases with speed, which must be reflected in the driving mode being planned. Physical limitations are based on the tire’s maximum adhesion property and the potential for motion stability. All these restrictions may be represented as hard and/or soft constraints. In the first case, satisfying limiting conditions is mandatory and formed by additional linear or nonlinear equalities and inequalities. In the second case, the restrictions themselves are included in the objective function [
12], which determines the influence of the criteria but allows for violations of the limits for exceptional reasons. For example, this approach is appropriate in cases when a safe distance between vehicles in the same lane is not ensured.
Obstacles: In real conditions, the AV must maneuver among moving and stationary obstacles [
13]. One of the main tasks on AV’s motion-planning strategy is to prevent collisions and maintain safe distances [
14]. The issue is complicated by the fact that it is difficult to accurately estimate the motion directions and speeds of other participants for a relatively long time. In this regard, it may be assumed that during AV’s maneuvering, other vehicles barely change the nature of their movement. Thus, the task is to maximize the distances between vehicles while maneuvering, considering vehicles’ safety contours.
This study continues a series of our research on developing algorithms for AV motion planning. Initially, in paper [
15], we outlined the main approaches to modeling trajectory and speed modes based on the vehicle kinematics model. At the same time, the sequential approach to optimizing the trajectory and speed was used. In the next research [
16], we stipulated an advantage of using integral forms of the equality constraints instead of point ones. In articles [
17], we tested various numerical integration schemes in optimization problems of motion planning.
Based on the above, we shall consider a motion-planning technique that involves the simultaneous prediction of trajectory and speed mode within a single finite element (FE) of the variable reference length. At the same time, depending on the number of lanes and presence of obstacles, trajectory variants are to be analyzed and estimated followed by possible selection of an optimal strategy in specific initial conditions. The general strategy of the technique is shown in
Figure 1.
2. Motion-Planning Concept
Let us assume that AV forms a motion trajectory relative to the midline of a lane on which it is currently located [
18,
19] (
Figure 2). Such a line may be obtained, for example, by the cubic spline interpolation providing sufficient smoothness and representing the curvature of a road section. If functional dependencies
x =
x(
sr),
y =
y(
sr) are built for a section, then, varying the parameter
sr relative to the initial path length
sr0, it is possible to determine the values
xsr,
ysr corresponding to the zero point of the lateral displacement
L along the road width. Thus, the coordinates
xf,
yf of the AV mass center’s final position can be calculated as
where
αf = the midline curve’s tangent slope at the point
xsr, and
Lf = AV mass center’s lateral displacement at the final point.
Since the transverse coordinate L is always perpendicular to the curve sr, the AV’s final angular position on any lane has ideally the same value αf. However, the angle α actual value may for various reasons deviate from the ideal value, which is associated with the desired angle αd. In the transverse direction, AV may be located in any lane by introducing such a parameter as the desired lateral displacement (position) Ld. Thus, the trajectory can be represented as a piecewise polynomial function with the basis D = xf − x0.
Trajectory. Let us apply the Hermitian interpretation of Lagrange polynomials [
20] using basis functions and nodal degrees of freedom (DOFs). To model the trajectory as a function
y(
x), the fifth-degree polynomial (
p = 5) may be considered, which requires three DOFs in a node. Then,
y(
x) may be expressed in the vector form
where
cj = polynomial coefficient,
j ∈ [0,
p].
If define
k = ((
p + 1)/2 − 1) first derivatives of the vector
X (where
p is odd) and substitute the
x-coordinates of the initial (0) and final (
D) nodes, the matrix
B is composed as
where
Qtr = set of trajectory nodal parameters.
As a result, expressions of Equations (2) and (3) are linked in the form
Considering one final element (FE) of the length
D, the function
y(
x) may be represented by the sets of shape functions and DOF values. Then
where
fj = shape function and
qj = weight coefficient or DOF,
j ∈ [1,
p+1].
If we assume
x =
ξD, where parameter
ξ ∈ [0, 1], then the shape functions may be derived through basis functions for an element of unit length. Thus, the basis functions
Fξ of the argument
ξ ∈ [0, 1] are universal for the variable length
D of a roadway segment. Defining the vector
D and matrix
Dd
where
k = ((
p + 1)/2 – 1).
Thus, the functions
F of Equation (3) may be replaced by the basis functions of the parameter
ξThen, for the
k-th derivative,
Speed Model can be organized in the same way as the trajectory. We assume that the AV longitudinal speed
Vζ is distributed along the trajectory projection on the
x-coordinate. According to Equations (3)–(8), the longitudinal component and its
k-th derivative may be expressed as follows
where
Qsp = set of speed nodal parameters.
Full set of parameters. Since we consider one segment with initial (0) and final (f) points, each of them containing three DOFs, we need six variable parameters for each
Qtr and
Qsp to simultaneously search the trajectory
y(
x) and speed distribution
Vζ(
x). Then, introducing nodal vectors
The vectors Qtr and Qsp completely determine the configuration of the trajectory and speed distributions within the segment D. Nevertheless, not all of their members are variable parameters. Note that at the initial node the vectors qtr0, qsp0 are usually known from the previous solution. In turn, strict constraints may be imposed on the values of vectors qtr0, qsp0, which reduces the number of parameters to be optimized.
Note that the nodal parameter
yf is not independent in its pure form but is calculated from Equation (1), where the parameters
sr and
L are involved. Their variation determines a combination of
yf and
D. Thus, it is possible to form a set of variable parameters that best reflect both the trajectory shape and the speed distribution.
In the case of the strict finite transverse displacement
Lf and the corresponding angle
αf, these parameters depend only on the variable
sr, and the set of Equation (11) is shortened
Numerical Integration Technique. To simultaneously satisfy the rapidity and quality of calculations, we shall use the numerical integration based on the
N-point Gaussian quadrature scheme. Assume that some integrand
z(
x) is considered within a segment [
xi−1,
xi], then, expressing
x =
ξD
where
wk = integration weight in the
k-th point,
ϑk—
k-th point in the master–element coordinate system,
J = Jacobian,
k ∈ [1,
N], and
N = number of integration points.
The short expression for calculating the integral becomes
where
D = section length and
z = vector of integrands of 1
× N size.
In most cases, we will need not only final integral values but also intermediate ones corresponding to the points
xi (or
ξi). That is, the integration must also be conducted on the sections [
xi−1,
xi].
If we combine all increments Δ
ξi in a row vector
Δξ, then the sum of Equation (13) can be obtained as follows
Planning Strategy in the Presence of Obstacles. Let us consider the AV motion in the presence of other vehicles (
Figure 3). We assume, for instance, that AV is moving along the middle (2) lane and may perform three maneuver variants [
21]: lane changes to the first (1) and third (3) lanes and following the second lane. In terms of geometric sense, such a rearrangement may be considered the best, which ensures keeping the greatest distance from an obstacle. At the moment of starting the motion planning (1)–(4), AV fixes the midline points of the current lane in such a way to form smooth conjugations (splines) reflecting the road curvature. Then, considering the number of lanes, the longitudinal
sr and transverse
L coordinates, it is possible to build a curvilinear grid of the road segment space relative to which the motion is planned.
By using radars, AV can obtain distances to objects and angles of measuring beams. Then, the coordinates of vehicles’ initial positions may be calculated relative to the AV’s
xy coordinate system. To do this, we use the vehicle safety zones (contours) represented by conditional circles. Thus, for a passenger vehicle, three round areas of radius
r relative to the front, middle, and rear vehicle parts are enough to form the safety contour. The positions of the front and back parts’ centers are calculated as follows
where
xc,
yc = coordinates of the vehicle’s mass center,
hf,
hr = distances to the conditional centers of the vehicle’s front and rear parts, respectively,
ϕ = yaw angle.
Thus, considering the radar measurements, the positions of other vehicles’ centers can be estimated, and based on them and the predetermined road segment curvilinear grid, the lateral displacements Li,0 can be found. This helps to determine the corresponding traffic lane for each moving obstacle. If the lane’s midline as the vehicle’s current trajectory is considered, then, similarly to AV, we can compose interpolation dependencies xi = xi(si), yi = yi(si), i ∈ [1, Nv], where Nv is the number of fixed vehicles.
Therefore, we may estimate the vehicle motion parameters in discrete points over time intervals Δ
t and the total AV’s motion time
tf. If vehicles’ speeds
Vi are accepted to be constants, then, the paths for the intervals Δ
t are Δ
si =
ViΔt and for the total time
si =
Vitf. Knowing the paths, it is possible to determine coordinates
xi,
yi by the predefined interpolation. Then, considering the radius of safety zones
ri, the distance between the AV (0) and an
i-th obstacle vehicle can be evaluated as
Note that, strictly speaking, to determine the vehicle mass center’s position, it is necessary to identify a vehicle and estimate its overall dimensions and wheelbase, which somewhat complicates the calculations. In this regard, we assume that moving obstacles behind AV may be conditionally defined by circles centered at the vehicles’ fronts, and obstacles ahead of AV may be represented by circles centered at the vehicles’ rears. The corresponding centers can be easily found by estimating the overall width of a moving obstacle. Then, the parameters Li,0 can be replaced by Li,f and Li,r, respectively. Thus, the trajectories of moving obstacles can be considered as the trajectories of the safety contour circles’ centers.
Time. The time interval Δ
ti between AV’s positions
i − 1 and
i can be calculated by the integration, considering the trajectory
s and speed
V Each segment [
xi−1,
xi], in turn, may also be represented by a finite element corresponding to a range [
ξi−1,
ξi] and again considered in the segment
ϑ ∈ [0, 1]. Then
Passing to numerical integration according to Equation (17), we obtain
Thus, it is possible to form a vector of time increments
where
n = number of intervals.
The vector of time points can be obtained by the accumulative addition:
Provided that the segment
ξ ∈ [0, 1] is represented by the vector of increments
Δξ, the final time can also be obtained as follows
Spatial Constraints. The requirement of non-violating the roadway boundaries is the main limiting factor. It can be used as a hard or soft constraint. In the first case, it is assumed that there must be no safety contour exceeding the roadway boundaries. In the second case, it is possible to allow the crossing of the solid marking line subject to a critical approach to an obstacle. In this case, we shall use strict conditions. If the maximum lateral coordinate relative to an AV’s initial position as
LU (upper limit) is denoted for moving leftward and as
LL (lower limit) for the right bias, then the conditions for each point
p ∈ [0,
f,
r] of the central, front, and rear vehicle’s parts have the form
In the case of riding on the same lane, the safety conditions can be enhanced by the values LU and LL equal to half the lane width with the corresponding sign. As a result, by setting the upper (left) and lower (right) boundaries, it is possible to limit the lateral displacement for each trajectory variant. Note that the values of L0,p can be obtained directly by interpolating the known coordinates from Equation (13).
5. Simulation
We use the data of the Audi A4 3.2 FSI [
23] to represent the AV. All the calculations are accomplished by using MATLAB tools [
24]. The basic function to realize the optimization procedure is
fmincon. The five-point Gauss quadrature scheme is accepted for the numerical integration, and the relative path step is Δ
ξ = 0.1.
Let us define constraints and initial values. Speed
Acceleration
where
aζlim = function of the vehicle throttle response characteristic.
First, let us set the values of the weight coefficients
W in Equation (72).
Considering the order of variable parameters, the coefficient values were selected based on the condition of the influence proportionality but with an emphasis on safety and maneuver precision.
Let us define constraints and initial values. Speed
Acceleration
where
aζlim = function of the vehicle throttle response characteristic.
Experiment 1. Let us consider variants for the AV motion planning along a curved road section with three lanes (
Figure 4) under the condition of perfect adhesion with
φmax = 0.85. Note that many variants are possible regarding the lateral vehicle displacement
L. However, for the sake of compactness, we built three prediction variants related to the lane change along the lanes’ midlines. The AV is surrounded by four moving obstacles denoted as
vi,
i ∈ [1, 2, …, 4]. Further, Index 0 denotes the initial position, and
f—is the final position. The initial vehicles’ speeds were, respectively,
Vi0 = (45, 50, 50, 60) km/h.
As can be seen, moving to Lane 1 requires the greatest change in curvature and, accordingly, a larger basis than for passing to Lane 3. At the same time, the smallest change in curvature on Lane 2 and the distant position of the impeding vehicle allows for the maximum acceleration and length of the prediction basis. Note that while driving along Lane 2, the trajectory forecast does not tend to coincide with the centerline but is formed in such a way as to ensure the least curvature.
Figure 5 depicts the results of forecasts for speeds and accelerations along with restrictions. The speed limits under the sideslip condition are shifted to the maneuver’s beginning movement to Lane 1 and are shifted to the final phase for passing to Lane 3. The restrictions for following Lane 2 are significantly higher than the speed limit
Vζmax and are not reflected in the figure’s boundaries. However, the speed in Lane 2 does not increase as the acceleration reaches the limit of the powertrain’s potential, and the final state of AV requires zero longitudinal acceleration. Note that the curvatures in the maneuvers’ final phases are approximately the same, which is reflected in the final values of lateral accelerations that are close in modules (
Figure 5c). As seen in
Figure 5d, not a single total acceleration curve exceeds the limit
amax by the adhesion properties.
Figure 6 presents the kinematic and geometric parameters used to impose restrictions. As seen in
Figure 6a, all the longitudinal jerk curves are well below the upper limit and do not exceed the lower limit. An important indicator is the smoothness and unambiguity of the yaw rate and acceleration curves (
Figure 6b,c). This characterizes the strict stability of the yaw angle and the absence of uncontrolled dynamic phenomena.
Figure 6d shows that AV increases the distance from vehicles located behind and reduces gaps with the vehicles ahead while keeping safe spaces.
Figure 7 reflects the geometric parameters characterizing the maneuvers’ trajectories. As seen in
Figure 7a, all curvature curves are extremely smooth and concordant with the polynomial used for the trajectory model. The steering angle (
Figure 7d) scales the curvature due to the ideal kinematic model. The curvature derivatives (
Figure 7b,c) are also smooth without intense oscillations and sharp spikes, characterizing a consistent speed distribution along the trajectory’s curvature.
Experiment 2. Now, we consider the motion-planning variants (
Figure 8) under the same conditions as for Experiment 1 but with worse adhesion
φmax = 0.5. In this case, more time and distance are needed to complete the maneuver since the traction potential is limited by an external factor. Therefore, the length of the required path
s is closer to the upper limit of the reference
sr length.
As seen, the nature of the maneuvers differs in phase. Thus, on Lane 2, there is no significant change in curvature, and AV may be quickly accelerated. For the maneuver to Lane 1, the intensity is shifted to the initial phase, and for the maneuver to Lane 2, it is shifted to the final phase.
Figure 9 depicts the results of forecasting speeds and accelerations under the condition of reduced tire-road adhesion. Accordingly, the critical speed values, caused by the locally increased curvature, also decrease and limit the increase in AV speed for maneuvers to Lanes 1 and 3 (
Figure 9a). At the same time, the curvature of the trajectory for driving on Lane 2 changes little, which allows the speed to rise since the critical one, in this case, is greater than the preset maximum. The same is noted for longitudinal accelerations with a difference that the vehicle, naturally, cannot realize its full traction potential. As seen in
Figure 9d, the maximum acceleration of the maneuver to Lane 1 is formed at the beginning of the lane change, and for Maneuver 3 at the end. With that, accelerations do not exceed the limit achievable under the conditions of full adhesion.
Figure 10 presents the kinematic parameters used as constraint criteria. As seen, all curves meet the requirements of smoothness and uniqueness, and their extremum values are significantly less than the preset limit ones. The yaw rate and angular acceleration for maneuvering along Lane 2 are noticeably lower than others due to the stability of the trajectory curvature. This also allows using a wider range of the jerk than for passing to Lanes 1 and 3.
Figure 11 shows the trajectory geometric parameters relative to the time. The greatest work on vehicle control occurs during the maneuver to Lane 1. However, the adhesion is reduced, the curves are stable, situated within the established limits, and the vehicle control is completely consistent with the kinematic characteristics in
Figure 10.