A park scene collision probability weighted dynamic potential field gradient descent motion preview method

Abstract

The application discloses a kind of garden scene collision probability weighted dynamic potential field gradient descent motion preview method, and vehicle and environmental state information are collected;Determine the number of preview steps and generate reference point sequence;Risk value is calculated to preview step and adaptively normal sampling;According to the approach rate, determine the collision probability, keep the potential field action range unchanged, the obstacle repulsive potential field is weighted by collision probability, the optimal initial sample is screened in combination with gravitational potential field and feasibility penalty term;Gradient descent algorithm is used to iterate and refine, collision probability and obstacle repulsive potential field gradient are updated in real time during iteration process, lateral coordinate is updated and projected to feasible region;Determine the planning point sequence and planning speed, generate feasible trajectory based on quadratic programming method.The application couples collision probability and potential field depth under the premise of not changing potential field action range, and dynamically updates potential field gradient in gradient descent process, and gives consideration to obstacle avoidance safety, trajectory smoothness and vehicle real-time performance.

Description

Technical Field

[0001] This invention relates to a dynamic potential field gradient descent motion pre-aiming method with collision probability weighting for a park scene. Background Technology

[0002] In park scenarios, trajectory planning for autonomous vehicles must simultaneously meet the requirements of high real-time performance, obstacle avoidance safety, smooth driving, and close adherence to a reference path. Currently, commonly used trajectory planning technologies mainly have the following characteristics and shortcomings.

[0003] Traditional artificial potential field methods are simple in structure and require little computation, making them widely used for local obstacle avoidance. In this method, the repulsive force of obstacles is determined by the relative distance between the vehicle and the obstacle. It can only reflect the degree of danger in spatial location and cannot distinguish between scenarios where the vehicle is close but will not collide with the obstacle and scenarios where the vehicle is close but will inevitably collide with the obstacle. This can easily lead to problems such as over-avoidance and trajectory jitter.

[0004] Existing technologies often combine artificial potential fields with sampling points to achieve early obstacle avoidance. Most of these schemes use fixed potential field parameters, or only adjust the lateral and longitudinal range of the potential field according to distance and velocity. They do not introduce collision probabilities at the motion aiming step level to weight and couple the potential field, making it difficult to adaptively adjust the obstacle avoidance intensity according to future collision probabilities, resulting in a trade-off between safety and ride comfort.

[0005] Existing solutions generally use a fixed artificial potential field for iterative solutions, without updating the pre-collision probability and potential field gradient in real time based on the iteration position during gradient descent. Some adaptive potential field solutions, in order to enhance obstacle avoidance, dynamically modify the lateral and longitudinal ranges of the potential field according to the risk, which can easily cause abrupt changes in trajectory curvature and poor smoothness.

[0006] Furthermore, methods based on preview theory for planning functions focus more on the positional state at the preview moment and pay insufficient attention to the process within the preview time. Such methods are inefficient at high sampling densities, and easily get trapped in local optima when the sampling density is reduced.

[0007] In summary, existing trajectory planning techniques suffer from the following core problems: the artificial potential field and the pre-aimed collision probability are poorly coupled, making it impossible to adaptively adjust the obstacle avoidance intensity according to the actual collision probability; during gradient descent iteration, the potential field cannot be updated in real time with the iteration position and collision probability, making it difficult to balance trajectory safety and smoothness; when using motion pre-aiming to achieve trajectory planning, it is difficult to adaptively sample according to the obstacle state, making it difficult to balance solution efficiency and optimality.

[0008] Therefore, there is an urgent need in this field for a planning method that couples the collision probability with the potential field depth without changing the range of the potential field, and dynamically updates the potential field during the gradient descent process, in order to solve the above-mentioned defects of the existing technology. Summary of the Invention

[0009] This invention provides a dynamic potential field gradient descent motion pre-aiming method for collision probability weighting in a park scene in order to solve the problems existing in the prior art.

[0010] The technical solutions adopted in this invention are as follows:

[0011] A dynamic potential field gradient descent motion prediction method for collision probability weighting in a park scene includes the following steps:

[0012] S1: Collect vehicle and environmental status information;

[0013] S2: Determine the number of aiming steps based on the speed of the main vehicle and the longitudinal distance between the main vehicle and the obstacle, determine the aiming range based on the speed of the main vehicle and the aiming time, use the center line of the lane as a reference, calculate the length of each step based on the aiming range and the number of aiming steps, generate the reference point for each aiming step, and form a reference trajectory from all the reference points.

[0014] S3: Determine the aiming step time based on the vehicle speed and the distance from the reference point to the vehicle, extract the motion state of the obstacle at the aiming step time, calculate the aiming step risk value based on the longitudinal distance and relative longitudinal speed between the vehicle and the obstacle at the aiming step time, determine the adaptive normal sampling step size based on the aiming step risk value, and generate aiming step sampling points along the normal of the reference trajectory.

[0015] S4: Estimate the motion time based on the speed of the main vehicle and the distance from the pre-aiming step sampling point to the main vehicle, extract the motion state of the obstacle during the motion time, calculate the left approach rate and right approach rate based on the pose and speed of the main vehicle at the pre-aiming step sampling point and the motion state of the obstacle during the motion time, determine the integration interval based on the left approach rate, right approach rate and preset approach rate threshold and calculate the collision probability.

[0016] Keeping the range of the potential field unchanged, the obstacle repulsive potential field is weighted and adjusted according to the collision probability. The total potential energy of the pre-aiming step sampling point is determined according to the gravitational potential field, the weighted obstacle repulsive potential field, and the feasibility penalty term. The optimal initial sampling point is selected according to the principle of minimizing the total potential energy.

[0017] S5: The gradient descent algorithm is used to iteratively refine the optimal initial sample points of each aiming step. During the iteration process, the collision probability is updated in real time according to the current iteration coordinates, and the gradient of the obstacle repulsive potential field is updated synchronously. The lateral coordinates of the iteration point are iteratively updated according to the updated total potential field gradient, and the updated lateral coordinates are projected to the feasible region until the change in lateral coordinates satisfies the convergence condition.

[0018] S6: Determine the planning point sequence and planning speed based on the iterative refinement results, and generate a smooth feasible trajectory based on the planning point sequence and planning speed.

[0019] Furthermore, in S1, the vehicle and environmental status information includes the current position and speed of the main vehicle, the position and speed of the obstacle during the pre-aiming time, the lane lines, the initial reference speed, and the vehicle's maneuvering parameters.

[0020] Furthermore, in S2, the number of aiming steps is determined based on the vehicle speed and the longitudinal distance between the vehicle and the obstacle, including:

[0021] When the main vehicle speed is less than 8m / s, the basic pre-aiming steps are 5~10;

[0022] When the main vehicle speed is greater than or equal to 8m / s, the basic pre-aiming steps are 10~15;

[0023] When the longitudinal distance between the main vehicle and the obstacle is greater than or equal to 8m, adjust the pre-aiming steps by subtracting 2 to 3 from the base pre-aiming steps.

[0024] When the longitudinal distance between the main vehicle and the obstacle is less than 8m, the number of aiming steps is adjusted to equal the basic aiming steps; the aiming time is 1s~2s.

[0025] Furthermore, in S3, the pre-aiming step risk value is calculated based on the longitudinal distance and relative longitudinal velocity between the main vehicle and the obstacle at the moment of the pre-aiming step, including using the following formula:

[0026] ,

[0027] In the formula, The relative velocity influence coefficient. This is the distance attenuation coefficient. The relative longitudinal velocity, This refers to the relative distance.

[0028] Further, in S3, determining the adaptive normal sampling step size based on the pre-aiming step risk value includes:

[0029] When the risk value of the aiming step is less than 0.3, the sampling step size is 0.4m; when the risk value of the aiming step is greater than or equal to 0.3 and less than or equal to 0.6, the sampling step size is 0.3m; when the risk value of the aiming step is greater than 0.6, the sampling step size is 0.2m.

[0030] Furthermore, in S4, before calculating the left and right approximation rates based on the pose and speed of the main vehicle at the pre-aiming step sampling point and the motion state of the obstacle during the motion time, the following steps are also included:

[0031] Based on the obstacle's center coordinates, dimensions, and heading angle, a rotation matrix is ​​used to calculate the four corner points of the obstacle's rectangular outline. Two external tangent lines are drawn from the main vehicle's position as the external point to the rectangular outline; the intersections of these external tangent lines and the rectangular outline are the left and right tangent points. The approximation rate threshold is calculated based on the limit approximation rate, which is calculated using the following formula:

[0032] ,

[0033] In the formula, Main vehicle speed, D represents the width of the car. lim This is the maximum collision avoidance distance.

[0034] Further, in S4, the integration interval is determined based on the left approximation rate, the right approximation rate, and a preset approximation rate threshold, and the collision probability is calculated. This includes determining the collision probability by integrating a probability density function of a triangular distribution within the integration interval. The probability density function of the triangular distribution is:

[0035] ,

[0036] In the formula, Let the integral variable be the approximation rate. and These are the left-limit approximation rate and the right-limit approximation rate, respectively.

[0037] Furthermore, in S4, the feasibility penalty term is:

[0038] ,

[0039] In the formula, For local curvature, To limit the curvature;

[0040] The local curvature was calculated using the three-point geometric method.

[0041] Furthermore, in S6, the planned vehicle speed is calculated based on the collision probability at each planning point location, including using the following formula:

[0042] ,

[0043] In the formula, The initial reference speed, This represents the collision probability.

[0044] Further, in S6, generating a smooth feasible trajectory based on the planned point sequence and the planned vehicle speed includes:

[0045] A smooth trajectory curve is obtained by applying cubic spline processing to the planned point sequence and planned vehicle speed. This smooth trajectory curve is used as the initial solution. A quadratic programming method is then used to fine-tune the lateral coordinates of each planned point in the initial solution, while keeping the longitudinal mileage unchanged, to obtain the final feasible trajectory. The objective function of the quadratic programming method is:

[0046] ,

[0047] In the formula, , , For each weight, The horizontal coordinate to be optimized. The initial position is unconstrained. For the curvature of the position to be optimized, To maximize the curvature, The lateral acceleration to be optimized Lateral acceleration is limited.

[0048] The present invention has the following beneficial effects:

[0049] 1) This invention deeply weights and couples the collision probability within the target range with the artificial potential field. Unlike potential field methods that rely solely on distance and velocity, this invention can adapt to complex scenarios such as no collision risk at close range and collision risk at distant range, thus mitigating the inherent defects of traditional potential fields that rely solely on distance from a mechanistic perspective.

[0050] 2) This invention adopts a mechanism of fixing the range of the potential field and only dynamically adjusting the strength of the potential field, which avoids the trajectory abrupt change and curvature jump problem caused by the existing technology of expanding or shrinking the lateral and longitudinal range of the potential field to achieve obstacle avoidance adjustment, and helps to improve trajectory smoothness.

[0051] 3) In the gradient descent iteration process, the collision probability and potential field gradient are updated synchronously with the iteration coordinates, enabling the trajectory optimization process to have real-time risk perception capability; combined with the adaptive pre-aiming sampling strategy, the computational burden is reduced while ensuring the solution accuracy, which helps to balance the solution efficiency and optimality, and adapts to the real-time requirements of the vehicle embedded platform.

[0052] 4) This invention forms a complete closed loop of pre-aiming sampling, collision probability calculation, vehicle speed adjustment, initial sample point evaluation and screening, and optimization solution. It is applicable to park scenarios, helps to achieve smooth trajectory and reliable obstacle avoidance, and has engineering feasibility. Attached Figure Description

[0053] Figure 1 This is a functional architecture diagram of the present invention.

[0054] Figure 2 This is a flowchart of the present invention.

[0055] Figure 3 A schematic diagram illustrating the principle of calculating the pre-aiming approximation rate and collision probability.

[0056] Figure 4 This is a schematic diagram of the probability density function and integration interval of a triangular distribution; Figure 5 This is a schematic diagram of risk normal sampling based on the pre-aiming step. Detailed Implementation

[0057] The invention will now be further described with reference to the accompanying drawings.

[0058] This embodiment applies to trajectory planning for low-speed autonomous vehicles in a closed park environment. It employs a collision probability-weighted dynamic potential field gradient descent motion prediction method to achieve safe, smooth, and real-time autonomous driving. The functional architecture of this method is as follows: Figure 1 As shown, the overall process is as follows: Figure 2 As shown, the specific implementation process is as follows.

[0059] First, vehicle and environmental status information is collected. Lane line information is obtained through the vehicle perception module, and the current pose and speed of the main vehicle are obtained through the positioning module. The pose and speed of the obstacle during the preview time are obtained through the obstacle tracking and prediction module. At the same time, the initial reference speed of the vehicle and vehicle maneuver parameters, including vehicle width, limit collision avoidance distance, maximum curvature limit, and maximum lateral acceleration limit, are read. The preview time is a fixed constant of 1-2 seconds, which is derived from the preview following theory, and the input information of the planning module is collected.

[0060] The number of aiming steps is determined based on the vehicle's speed and the longitudinal distance between the vehicle and the obstacle. Simultaneously, the aiming range is determined based on the vehicle's speed and aiming time. Using the lane centerline as a reference, the length of each step is calculated based on the aiming range and the number of aiming steps, generating reference points for each aiming step. All reference points then form a reference trajectory. The rule for determining the number of aiming steps is as follows:

[0061] When the main vehicle speed is less than 8m / s (v ego <8m / s), the basic pre-aiming steps are 5~10 (N_base=5~10);

[0062] When the main vehicle speed is greater than or equal to 8m / s, the basic pre-aiming step count is 10~15 (N_base=10~15).

[0063] When the longitudinal distance between the main vehicle and the obstacle is greater than or equal to 8m, adjust the aiming step number by subtracting 2 to 3 from the base aiming step number; when the longitudinal distance between the main vehicle and the obstacle is less than 8m, adjust the aiming step number to equal the base aiming step number N = N_base.

[0064] Each aiming step k corresponds to a reference point, with the lateral coordinate of the reference point being 0m. All reference points are evenly distributed along the lane centerline, forming the reference path for vehicle travel, such as... Figure 3 As shown.

[0065] The aiming step time is determined based on the vehicle speed and the distance from the reference point to the vehicle. The motion state of the obstacle at that aiming step time is extracted. The aiming step risk value is calculated based on the longitudinal distance and relative longitudinal velocity between the vehicle and the obstacle at that aiming step time. The aiming step risk value is calculated using the following formula:

[0066] ,

[0067] In the formula, The relative velocity influence coefficient. This is the distance attenuation coefficient. The relative longitudinal velocity, This refers to the relative distance.

[0068] The adaptive normal sampling step size is determined based on the pre-aiming step risk value, and the specific rules are as follows:

[0069] When the risk value of the aiming step is less than 0.3, the sampling step size is 0.4m; when the risk value of the aiming step is greater than or equal to 0.3 and less than or equal to 0.6, the sampling step size is 0.3m; when the risk value of the aiming step is greater than 0.6, the sampling step size is 0.2m.

[0070] Preview step sampling points are generated along the normal to the reference trajectory. The formula for calculating the sampling points is as follows:

[0071] ,

[0072] ,

[0073] In the formula, To achieve adaptive sampling step size, it is limited to the maximum executable lateral displacement range of the vehicle. For sampling sequence number, Let be the normal unit vector of the reference trajectory. The reference trajectory is the direction angle of this preview step, and the normal sampling method based on the preview step risk is as follows: Figure 5 As shown, the higher the risk, the denser the sampling, balancing solution efficiency and planning accuracy.

[0074] The movement time is determined based on the vehicle speed and the distance from the pre-aiming step sampling point to the vehicle. The movement state of the obstacle during this time is extracted. Before calculating the left and right approximation rates, the four corner points of the obstacle's rectangular outline are calculated using a rotation matrix based on the obstacle's center coordinates, size, and heading angle. Two external tangent lines are drawn from the vehicle's position as the external point to the rectangular outline. The intersection of these external tangent lines with the rectangular outline are the left and right tangent points. Figure 4 As shown.

[0075] Based on the pose and speed of the main vehicle at the pre-aiming step sampling point, and the motion state of the obstacle during the motion time, the left approximation rate and the right approximation rate are calculated, and the calculation formulas are as follows:

[0076] ,

[0077] ,

[0078] In the formula, and Let k be the approximation rate of the main vehicle to the left / right tangent point of the obstacle in the case of the j-th sampling point of the pre-aiming step. and The left and right tangential point position vectors of the virtual outline of the obstacle. and The velocity vector between the main vehicle and the obstacle. The position vector of the vehicle's center of mass.

[0079] In this embodiment, the preset approximation rate threshold is the limit approximation rate. Since the left limit approximation rate and the right limit approximation rate have a symmetrical relationship of equal absolute values, the limit approximation rate is calculated using an approximate common formula in engineering implementation. There is no need to derive two independent sets of geometric analytical expressions for the left and right sides separately. The formula for calculating the limit approximation rate is:

[0080] ,

[0081] In the formula, Main vehicle speed, For the width of the car, The distance is the limit of collision avoidance; the left and right limit approximation rates are the positive and negative forms of the calculated result, respectively. It should be noted that this formula is an engineering approximation under the symmetry assumption that "the absolute values ​​of the left and right limit approximation rates are equal". In actual complex scenarios, there may be slight differences in the left and right geometric conditions, but in order to reduce the amount of calculation, this formula is used as a uniform approximation to meet the accuracy requirements and reduce the computational cost.

[0082] The integration interval is determined based on the left approximation rate, the right approximation rate, and a preset approximation rate threshold. The collision probability is determined by integrating the probability density function of the triangular distribution within the integration interval. The probability density function of the triangular distribution is as follows:

[0083] ,

[0084] The formula for calculating the collision probability is:

[0085] ,

[0086] In the formula, Let the integral variable be the approximation rate. and These are the left-limit approximation rate and the right-limit approximation rate, respectively, and their absolute values ​​are equal.

[0087] Keeping the effective range of the obstacle repulsive potential field unchanged, the obstacle repulsive potential field is weighted and adjusted according to the collision probability. The calculation methods for the gravitational potential field, the weighted obstacle repulsive potential field, the feasibility penalty term, and the total potential field energy are as follows:

[0088] ,

[0089] ,

[0090] ,

[0091] ,

[0092] In the formula, The horizontal coordinates of the pre-aimed sampling point, Let x be the lateral coordinate of the obstacle. The horizontal coordinate of the reference point For the fundamental coefficients of the repulsive potential field, The transverse Gaussian attenuation coefficient is... The gravitational potential coefficient is... For local curvature, the three-point geometric method is used for calculation. Due to curvature constraints, The penalty coefficient is used to select the optimal initial sample points from the pre-aiming step sampling points based on the principle of minimizing the total potential energy.

[0093] The gradient descent algorithm is used to iteratively refine the optimal initial sample points for each aiming step. During the iteration, the collision probability is updated in real time based on the current iteration coordinates, and the gradient of the obstacle repulsive potential field is updated synchronously. The total potential field gradient is the sum of the obstacle repulsive potential field gradient, the gravitational potential field gradient, and the penalty term gradient, and the calculation formula is as follows:

[0094] ,

[0095] The formula for iteratively updating the horizontal coordinate is:

[0096] ,

[0097] The updated horizontal coordinates are projected onto the feasible region in real time, with the following constraints:

[0098] ,

[0099] in and The maximum lateral displacement of the vehicle to the left and right; until the lateral coordinate change at each preview step iteration point satisfies the convergence condition, which is:

[0100] ,

[0101] In the formula,

[0102] This represents the total energy gradient after k iterations; , , Let represent the gradient of the repulsive potential field, the gradient of the gravitational potential field, and the gradient of the penalty term in the i-th iteration of the aiming step k. This represents the horizontal coordinate at the i-th iteration of the aiming step k; This represents the collision probability after i iterations when the aiming step k is initialized with the j-th sampling point; and Indicates the maximum lateral displacement to the left and right. To achieve the convergence threshold and refine the optimal initial sample points, the lateral coordinates of the iteration points gradually converge towards the low potential field region while remaining within the feasible region, without any sudden changes in trajectory curvature.

[0103] Based on the iterative refinement results, the sequence of planning points and the planned vehicle speed are determined. The sequence of planning points is represented as follows:

[0104] ,

[0105] The planned vehicle speed is calculated using the following formula:

[0106] ,

[0107] In the formula, The initial reference speed, This represents the collision probability at the corresponding iteration position; This represents the vehicle speed at position i at the aiming step k. Indicates the curvature of the road;

[0108] The final vehicle speed also needs to be constrained by the vehicle's lateral acceleration limit, and the calculation formula is as follows:

[0109] ,

[0110] In the formula, For the curvature of the position to be optimized, Lateral acceleration is a constraint. A smooth trajectory curve is obtained by cubic spline processing of the planned point sequence and the planned vehicle speed. This smooth trajectory curve is used as the initial solution. A quadratic programming method is then used to fine-tune the lateral coordinates of each planned point in the initial solution, while keeping the longitudinal mileage unchanged, to obtain the final feasible trajectory. The objective function of the quadratic programming method is:

[0111] ,

[0112] In the formula, , , For each weight, The horizontal coordinate to be optimized. The initial position is unconstrained. For the curvature of the position to be optimized, The lateral acceleration to be optimized To maximize the curvature, Lateral acceleration is limited.

[0113] After secondary planning optimization, the smoothness of the trajectory is further improved, and the final feasible trajectory not only meets the obstacle avoidance safety requirements, but also conforms to the vehicle's maneuverability constraints. It is suitable for the driving needs of low-speed autonomous driving scenarios in the park and can effectively solve the problems of excessive obstacle avoidance and trajectory jitter in traditional artificial potential field methods, as well as the inability of existing preview planning methods to take into account both safety and smoothness.

[0114] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A dynamic potential field gradient descent motion prediction method for collision probability weighting in a park scene, characterized in that: Includes the following steps: S1: Collect vehicle and environmental status information; S2: Determine the number of aiming steps based on the speed of the main vehicle and the longitudinal distance between the main vehicle and the obstacle, determine the aiming range based on the speed of the main vehicle and the aiming time, use the center line of the lane as a reference, calculate the length of each step based on the aiming range and the number of aiming steps, generate the reference point for each aiming step, and form a reference trajectory from all the reference points. S3: Determine the aiming step time based on the vehicle speed and the distance from the reference point to the vehicle, extract the motion state of the obstacle at the aiming step time, calculate the aiming step risk value based on the longitudinal distance and relative longitudinal speed between the vehicle and the obstacle at the aiming step time, determine the adaptive normal sampling step size based on the aiming step risk value, and generate aiming step sampling points along the normal of the reference trajectory. S4: Estimate the motion time based on the speed of the main vehicle and the distance from the pre-aiming step sampling point to the main vehicle, extract the motion state of the obstacle during the motion time, calculate the left approach rate and right approach rate based on the pose and speed of the main vehicle at the pre-aiming step sampling point and the motion state of the obstacle during the motion time, determine the integration interval based on the left approach rate, right approach rate and preset approach rate threshold and calculate the collision probability. Keeping the range of the potential field unchanged, the obstacle repulsive potential field is weighted and adjusted according to the collision probability. The total potential energy of the pre-aiming step sampling point is determined according to the gravitational potential field, the weighted obstacle repulsive potential field, and the feasibility penalty term. The optimal initial sampling point is selected according to the principle of minimizing the total potential energy. S5: The gradient descent algorithm is used to iteratively refine the optimal initial sample points of each aiming step. During the iteration process, the collision probability is updated in real time according to the current iteration coordinates, and the gradient of the obstacle repulsive potential field is updated synchronously. The lateral coordinates of the current iteration point are iteratively updated according to the updated total potential field gradient. The updated lateral coordinates are projected to the feasible region until the change in lateral coordinates satisfies the convergence condition. S6: Determine the planning point sequence and planning speed based on the iterative refinement results, and generate a smooth feasible trajectory based on the planning point sequence and planning speed.

2. The dynamic potential field gradient descent motion pre-aiming method for collision probability weighting in a park scene as described in claim 1, characterized in that: In S1, the vehicle and environmental status information includes the current position and speed of the main vehicle, the position and speed of the obstacle during the pre-aiming time, the lane lines, the initial reference speed, and the vehicle's maneuvering parameters.

3. The dynamic potential field gradient descent motion prediction method for collision probability weighting in a park scene as described in claim 1, characterized in that: In S2, the number of aiming steps is determined based on the vehicle's speed and the longitudinal distance between the vehicle and the obstacle, including: When the main vehicle speed is less than 8m / s, the basic pre-aiming steps are 5~10; When the main vehicle speed is greater than or equal to 8m / s, the basic pre-aiming steps are 10~15; When the longitudinal distance between the main vehicle and the obstacle is greater than or equal to 8m, adjust the pre-aiming steps by subtracting 2 to 3 from the base pre-aiming steps. When the longitudinal distance between the main vehicle and the obstacle is less than 8m, the number of aiming steps is adjusted to equal the basic aiming steps; the aiming time is 1s~2s.

4. The dynamic potential field gradient descent motion prediction method for collision probability weighting in a park scene as described in claim 1, characterized in that: In S3, the risk value of the pre-aiming step is calculated based on the longitudinal distance and relative longitudinal velocity between the main vehicle and the obstacle at the moment of the pre-aiming step, including the following formula: ， In the formula, The relative velocity influence coefficient. This is the distance attenuation coefficient. The relative longitudinal velocity, This refers to the relative distance.

5. The dynamic potential field gradient descent motion prediction method for collision probability weighting in a park scene as described in claim 4, characterized in that: In S3, the adaptive normal sampling step size is determined based on the pre-aiming step risk value, including: When the risk value of the aiming step is less than 0.3, the sampling step size is 0.4m; when the risk value of the aiming step is greater than or equal to 0.3 and less than or equal to 0.6, the sampling step size is 0.3m; when the risk value of the aiming step is greater than 0.6, the sampling step size is 0.2m.

6. The dynamic potential field gradient descent motion pre-aiming method for collision probability weighting in a park scene as described in claim 1, characterized in that: In S4, before calculating the left and right approximation rates based on the pose and speed of the main vehicle at the pre-aiming step sampling point and the motion state of the obstacle during the motion time, the following steps are also included: Based on the obstacle's center coordinates, dimensions, and heading angle, a rotation matrix is ​​used to calculate the four corner points of the obstacle's rectangular outline. Two external tangent lines are drawn from the main vehicle's position as the external point to the rectangular outline; the intersections of these external tangent lines and the rectangular outline are the left and right tangent points. The approximation rate threshold is calculated based on the limit approximation rate, which is calculated using the following formula: ， In the formula, Main vehicle speed, D represents the width of the car. lim This is the maximum collision avoidance distance.

7. The dynamic potential field gradient descent motion pre-aiming method for collision probability weighting in a park scene as described in claim 1, characterized in that: In S4, the integration interval is determined based on the left approximation rate, the right approximation rate, and a preset approximation rate threshold, and the collision probability is calculated. This includes determining the collision probability by integrating a probability density function of a triangular distribution within the integration interval. The probability density function of the triangular distribution is: ， In the formula, Let the integral variable be the approximation rate. and These are the left-limit approximation rate and the right-limit approximation rate, respectively.

8. The dynamic potential field gradient descent motion pre-aiming method for collision probability weighting in a park scene as described in claim 1, characterized in that: In S4, the feasibility penalty item is: ， In the formula, For local curvature, To limit the curvature; The local curvature was calculated using the three-point geometric method.

9. The dynamic potential field gradient descent motion pre-aiming method for collision probability weighting in a park scene as described in claim 1, characterized in that: In S6, the planned vehicle speed is calculated based on the collision probability at each planning point location, including using the following formula: ， In the formula, The initial reference speed, This represents the collision probability.

10. The dynamic potential field gradient descent motion pre-aiming method for collision probability weighting in a park scene as described in claim 1, characterized in that: In S6, a smooth feasible trajectory is generated based on the planned point sequence and the planned vehicle speed, including: A smooth trajectory curve is obtained by applying cubic spline processing to the planned point sequence and planned vehicle speed. This smooth trajectory curve is used as the initial solution. A quadratic programming method is then used to fine-tune the lateral coordinates of each planned point in the initial solution, while keeping the longitudinal mileage unchanged, to obtain the final feasible trajectory. The objective function of the quadratic programming method is: ， In the formula, , , For each weight, The horizontal coordinate to be optimized. The initial position is unconstrained. For the curvature of the position to be optimized, To maximize the curvature, The lateral acceleration to be optimized Lateral acceleration is limited.

Images