Rule-based degradation control method and system considering trajectory prediction uncertainty

Abstract

The application discloses a rule degradation control method considering trajectory prediction uncertainty, comprising the following steps: inputting a historical driving trajectory into a pre-trained neural network to output a Gaussian distribution probability of a predicted horizontal and longitudinal position; multiplying the Gaussian distribution probability with a preset lateral passing artificial potential field function to output a lateral passing risk field with time information; combining the lateral passing risk field with rules other than the collision avoidance rule to form a relaxed rule constraint set of a target vehicle at a future time; designing a target function, introducing the formed relaxed rule constraint set to form a quadratic programming solving problem; when the quadratic programming has no solution under the existing constraint set, relaxing the constraint according to the importance of the rule and re-solving in sequence, and finally obtaining a feasible control sequence satisfying the rule to control the driving of the target vehicle. The application considers the uncertainty of the lateral passing prediction trajectory, maximizes the satisfaction of other rules under the premise of guaranteeing the collision avoidance safety rule, and smoothly generates a safe and feasible rule control instruction.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to a rule-based degrading control method and system that takes into account the uncertainty of trajectory prediction. Background Technology

[0002] With the rapid development of technologies such as automation and artificial intelligence, safety requirements have gradually gained attention in various engineering fields, leading to an increasing number of regulatory requirements, particularly in the field of autonomous driving, where adherence to reasonable rules is essential. In addition to general rules for vehicle control, such as minimizing energy consumption, maximizing commuting efficiency, and acceleration and turning angle constraints, autonomous vehicles must also comply with complex, conflicting, and prioritized traffic rules, such as collision avoidance, lane speed limits, following distance, and lane priority. This necessitates predicting the future trajectories of dynamic traffic participants and fully considering the uncertainties of prediction to ensure the implementation of the highest-priority collision avoidance rules. Furthermore, methods that directly abandon all rules except collision avoidance to generate emergency safety controls when a feasible solution cannot be generated under too many constraints significantly impact the continuity of vehicle control, passenger comfort, and the normal driving of other traffic participants.

[0003] Therefore, it is still necessary to design a rule-based degradation control method that can address the uncertainty in predicting the trajectory of adjacent vehicles. Summary of the Invention

[0004] This invention addresses the problem that existing autonomous driving control methods struggle to generate collision-free control sequences while simultaneously satisfying other rule constraints under set rules. It provides a rule-downgraded control method and system that considers trajectory prediction uncertainty.

[0005] The technical solution adopted in this invention is:

[0006] A rule-based degradation control method considering trajectory prediction uncertainty is provided, comprising the following steps:

[0007] S1. Input the driving trajectories of the target vehicle and its adjacent vehicles within a specific time period in the past into a pre-trained neural network, and output the Gaussian distribution probability representing the predicted horizontal and vertical positions of the target vehicle and its adjacent vehicles.

[0008] S2. Multiply the Gaussian distribution probability of the vehicle to be parked by the preset artificial potential field function of the vehicle to output the vehicle risk field with time information.

[0009] S3. Combine the adjacent vehicle risk field with the rules other than the collision avoidance rules to form a relaxed rule constraint set for the target vehicle at future moments;

[0010] Among them, the rule corresponding to the side vehicle risk field is no collision (r1). The rules other than the collision avoidance rule include no collision (r1), lane keeping (r2), meeting the maximum speed limit (r3), meeting the minimum speed limit (r4), and smooth and comfortable driving (r5). Then the total rule set S = {r1,r2,r3,r4,r5}.

[0011] The total rule set S is divided into multiple equivalence class combinations according to their importance, including C1 = {r4}, C2 = {r5}, C3 = {r2, r3}, and C4 = {r1}. The constraint sets for relaxation rules are constructed according to the importance of the equivalence class combinations: R1 = S, R2 = S / {C1}, R3 = S / {C1, C2}, R4 = S / {C3}, R5 = S / {C1, C3}, and R6 = S / {C1, C2, C3}. When there is no feasible solution under the constraint set R1 of the relaxation rules, the constraint set is switched to R2 to continue solving, and this process is repeated recursively.

[0012] S4. Design the objective function, introduce the constructed relaxed rule constraint set for condition constraints, and form a quadratic programming problem.

[0013] S5. When there is no solution to the quadratic programming under the existing constraint set, the constraints are relaxed in turn according to the importance of the rules and the solution is solved again. Finally, a set of feasible control sequences that satisfy the rules is obtained to control the driving of the target vehicle.

[0014] Following the above technical solution, the neural network is an LSTM neural network.

[0015] Following the above technical solution, the neural network has n attention heads. The output of each attention head will decode a set of predicted trajectories of the target vehicle in the future time T, with a probability close to the true value. Finally, the set of trajectories with the closest probability is selected.

[0016] Following the above technical solution, the artificial potential field function U of the side vehicle o as follows:

[0017]

[0018] In the formula, a i and b i Let X and Y represent the potential field strength parameters and shape parameters of the i-th vehicle; (X, Y) represent the coordinates of any position in the artificial potential field of the vehicle. and Let x and y be the horizontal and vertical coordinates of the i-th adjacent vehicle, respectively. and These represent the longitudinal and lateral safe distances from the i-th adjacent vehicle, respectively.

[0019] Following the above technical solution, and Defined as:

[0020]

[0021] In the formula, X0 is the minimum longitudinal distance, Y0 is the minimum lateral distance, T0 represents the safety time interval, and u e θ represents the relative longitudinal velocity between the vehicle and the adjacent vehicle. e The relative heading angle between the vehicle and the adjacent vehicle.

[0022] Following the above technical solution, step S2 specifically involves: sampling the distribution probability of the predicted trajectory of the adjacent vehicle at each prediction time, where the probability is P. μ The probability of the mean position point and its four symmetrical points (front, back, left, and right) is 0.7P. μ The location points are determined, and the probability values ​​of the five selected location points are respectively compared with the vehicle potential field function U. o Multiplying these values ​​yields five new potential fields. Finally, the maximum value at the point where the five new potential fields overlap is taken as the potential field value of the adjacent vehicle generated by the vehicle at that moment. The potential field values ​​of all adjacent vehicles together constitute the adjacent vehicle risk field for collision avoidance of the vehicle, which is used to form the position constraints of the collision-free rule.

[0023] Following the above technical solution, the objective function designed in step S4 is:

[0024]

[0025] In the formula, N t To control the time domain, u k Let y be the control vector at time k, which includes the longitudinal control input acceleration and the lateral control input front wheel steering angle acceleration. k Let be the state error vector at time k, which includes lateral tracking error and heading angle tracking error, and its reference trajectory is the centerline of the target lane.

[0026] Following the above technical solution, when predicting the trajectory of a vehicle, the vehicle being predicted is taken as the target vehicle, and a limited number of vehicles around the target vehicle are considered to form a prediction network.

[0027] This invention also provides a rule-based degradation control system that considers trajectory prediction uncertainty, comprising:

[0028] The module for predicting the future occupancy area of ​​adjacent vehicles is used to input the driving trajectories of the target vehicle and its adjacent vehicles within a specific time period in the past into a pre-trained LSTM neural network, and output a Gaussian distribution probability representing the predicted horizontal and vertical positions of the target vehicle and its adjacent vehicles.

[0029] The vehicle risk field construction module is used to multiply the Gaussian distribution probability of the vehicle by the preset artificial potential field function of the vehicle, and output the vehicle risk field with time information.

[0030] The relaxed rule constraint module combines the adjacent vehicle risk field with rules other than collision avoidance rules to form a relaxed rule constraint set for the target vehicle at future moments;

[0031] Among them, the rule corresponding to the adjacent vehicle risk field is no collision (r1), that is, no collision with adjacent vehicles is allowed during driving; the rules other than the collision avoidance rule include lane keeping (r2), that is, not leaving the current lane and trying to stay in the center of the lane; meeting the maximum speed limit (r3); meeting the minimum speed limit (r4); smooth and comfortable driving (r5), that is, not generating large lateral or longitudinal acceleration. Then the total rule set S = {r1,r2,r3,r4,r5}.

[0032] The total rule set S is divided into multiple equivalence class combinations according to their importance, including C1 = {r4}, C2 = {r5}, C3 = {r2, r3}, and C4 = {r1}. The constraint sets for relaxation rules are constructed according to the importance of the equivalence class combinations: R1 = S, R2 = S / {C1}, R3 = S / {C1, C2}, R4 = S / {C3}, R5 = S / {C1, C3}, and R6 = S / {C1, C2, C3}. When there is no feasible solution under the constraint set R1 of the relaxation rules, the constraint set is switched to R2 to continue solving, and this process is repeated recursively.

[0033] The rule-downgraded control module is used to design the objective function, introduce a set of relaxed rule constraints to impose conditional constraints, and form a quadratic programming problem. When there is no solution to the quadratic programming under the existing constraint set, the constraints are relaxed in sequence according to the importance of the rules and the problem is solved again. Finally, a set of feasible control sequences that satisfy the rules are obtained to control the driving of the target vehicle.

[0034] The present invention also provides a computer storage medium storing a computer program executable by a processor, the computer program executing the rule-based degradation control method for considering trajectory prediction uncertainty as described in the above technical solution.

[0035] The beneficial effects of this invention are as follows: This invention mainly addresses the various rule constraints that autonomous vehicles should meet. It utilizes a neural network to generate a bivariate Gaussian probability distribution function for the future trajectory of a neighboring vehicle, and combines this with a potential field function to generate the driving space that the neighboring vehicle will occupy. This serves as the highest collision avoidance safety constraint, combined with other safety and comfort-related driving rule information as supplementary constraints, and combined with the objective function, to generate a smooth control sequence that the autonomous vehicle can execute in the future. Within a specified solution time Δt, if there is no feasible solution under the current constraints, the rule constraints are relaxed and the solution is recalculated. This process is iterated until a feasible solution is found, thereby ensuring the smooth generation of the control sequence under multiple constraints. This avoids the impact on the continuity of autonomous vehicle control, passenger comfort, and the normal driving of other traffic participants caused by methods that directly abandon all rules except collision avoidance to generate emergency safety control. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart of a rule-based degradation control method for trajectory prediction uncertainty in an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the target vehicle trajectory prediction network model according to an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram of the artificial potential field function of the vehicle in an embodiment of the present invention;

[0040] Figure 4 This is a block diagram of the rule degradation logic in an embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0042] Example 1

[0043] like Figure 1 As shown, the data-driven feedforward plus feedback compensation rotation angle optimization method of this invention mainly includes the following steps:

[0044] S1. Input the driving trajectories of the target vehicle and its adjacent vehicles within a specific time period in the past into a pre-trained neural network, and output the Gaussian distribution probability representing the predicted horizontal and vertical positions of the target vehicle and its adjacent vehicles.

[0045] S2. Multiply the Gaussian distribution probability of the vehicle to be parked by the preset artificial potential field function of the vehicle to output the vehicle risk field with time information.

[0046] S3. Combine the adjacent vehicle risk field with the rules other than the collision avoidance rules to form a relaxed rule constraint set for the target vehicle at future moments;

[0047] Among them, the rule corresponding to the side vehicle risk field is no collision (r1). The rules other than the collision avoidance rule include no collision (r1), lane keeping (r2), meeting the maximum speed limit (r3), meeting the minimum speed limit (r4), and smooth and comfortable driving (r5). Then the total rule set S = {r1,r2,r3,r4,r5}.

[0048] The total rule set S is divided into multiple equivalence class combinations according to their importance, including C1 = {r4}, C2 = {r5}, C3 = {r2, r3}, and C4 = {r1}. The constraint sets for relaxation rules are constructed according to the importance of these equivalence class combinations: R1 = S, R2 = S / {C1}, R3 = S / {C1, C2}, R4 = S / {C3}, R5 = S / {C1, C3}, and R6 = S / {C1, C2, C3}. When there is no feasible solution under the constraint set R1 of the relaxation rules, the constraint set is switched to R2 to continue solving. If there is still no feasible solution, the constraint set is switched to R3 to continue solving. If there is still no feasible solution, the constraint set is switched to R4 to continue solving. If there is still no feasible solution, the constraint set is switched to R5 to continue solving. If there is still no feasible solution, the constraint set is switched to R6 to continue solving, until the corresponding solution is finally obtained.

[0049] S4. Design the objective function, introduce the constructed relaxed rule constraint set for condition constraints, and form a quadratic programming problem.

[0050] S5. When there is no solution to the quadratic programming under the existing constraint set, the constraints are relaxed in turn according to the importance of the rules and the solution is solved again. Finally, a set of feasible control sequences that satisfy the rules is obtained to control the driving of the target vehicle.

[0051] In step S1, the driving trajectories (including position, heading angle, speed, acceleration, etc.) of the target vehicle and its adjacent vehicles within the past tp time period can be input into a neural network (such as LSTM). After being encoded by the neural network, the data is then input into a multi-head attention pool, and the position prediction probability distribution parameters of the target vehicle are decoded and output. These parameters are binary Gaussian distribution parameters, which represent the future horizontal and vertical position distribution probability of the target vehicle.

[0052] The reason this invention considers the uncertainty of trajectory prediction is that the trajectory prediction module receives environmental observation information from the perception module, such as the position, heading angle, speed, and acceleration of the adjacent vehicle, and maps the future trajectory distribution of the adjacent vehicle. This process is subject to interference from observation noise, resulting in some inaccuracy in the input of the trajectory prediction module. Furthermore, trajectory prediction cannot be perfect; over-reliance on the results of trajectory prediction can make it difficult for the lower-level control module to provide a set of rules that satisfy the highest priority of collision avoidance.

[0053] The reason why this invention adopts a rule-downgraded control method is that a control method that switches back and forth between a high-performance and high-safety controller based on a certain decision logic will abandon other rule requirements except for collision avoidance rules when it touches the defined boundary, generating an unsmooth and irregular control sequence, which has an adverse effect on the continuity of vehicle control, passenger comfort, and the normal driving of other traffic participants.

[0054] This invention employs a rule-downgrading control method, taking into account the uncertainty of trajectory prediction. This method fully utilizes the probability distribution information output by the trajectory prediction module to safely construct a side-vehicle risk field that meets the collision avoidance rule requirements, ensuring the highest priority collision avoidance rule is always used during the control process. Under this collision avoidance safety envelope, the rule-downgrading control method gradually releases lower-priority rule constraints when existing rule constraints are insufficient to generate feasible control sequences, smoothly solving for feasible control sequences while maximizing the satisfaction of rule constraints.

[0055] Example 2

[0056] The main steps of this embodiment are the same as those of Embodiment 1. The difference is that the trajectory prediction design considers uncertainty. In this embodiment, a trajectory prediction network model under multi-head attention is constructed based on the transformer model framework and combined with an LSTM encoding and decoding module. Figure 2 As shown.

[0057] Specifically, the historical trajectory information of the target vehicle and its adjacent vehicles, including the position, heading angle, speed, acceleration, and other trajectories of the target vehicle and its adjacent vehicles within the past time period (tp), is input into the LSTM encoder. The historical trajectory information of each vehicle is encoded into a feature vector, and the feature vectors encoded by all vehicles form a set of feature vectors. This set of feature vectors is then indiscriminately input into each attention head in the multi-head attention module. Each attention head encodes a set of vehicle interaction relationship vectors based on the importance of the adjacent vehicles' influence on the target vehicle's trajectory. This set of relationship vectors is then input into the LSTM decoder and the multilayer perceptron. The difference is that the LSTM decoder processes the relationship vectors given by each attention head separately and outputs a binary Gaussian distribution parameter, μ. t ,σ t They represent the mean vector and covariance matrix at time t, respectively:

[0058]

[0059] The above parameters can be used to represent the trajectory distribution probability of the target vehicle at time t:

[0060]

[0061] The multilayer perceptron, on the other hand, analyzes the relation vectors given by each attention head as a whole, and uses Softmax operations to obtain the probability of approximating the true value of the target vehicle trajectory distribution output by each head. Finally, it selects the predicted trajectory with the highest probability as the vehicle's future predicted trajectory in the time domain.

[0062] To fully utilize the trajectory prediction results and consider the uncertainty in prediction results caused by noise in the received data and network fitting errors during the prediction process, a risk field for oncoming vehicles at future time moments is generated to ensure the highest priority collision avoidance rule requirements. The sampling probability is P. μ The probability of the mean position point and its four symmetrical points (front, back, left, and right) is 0.7P. μ The location points, and the probability values ​​of the five selected location points are respectively compared with, as shown in the example. Figure 3 The artificial potential field function U shown is shown. o Multiplying these values ​​yields five new potential fields. The maximum value at the intersection of these five potential fields is then taken as the potential field value generated by the vehicle at that moment. The potential field values ​​of all adjacent vehicles together constitute the collision avoidance risk field of the vehicle, which is used to construct the position constraints without collision rules.

[0063] The artificial potential field function U of the side vehicle o for:

[0064]

[0065] In the formula, a i and b i Let X and Y represent the potential field strength parameters and shape parameters of the i-th vehicle; (X, Y) represent the coordinates of any position in the artificial potential field of the vehicle. and Let x and y be the horizontal and vertical coordinates of the i-th adjacent vehicle, respectively. and These represent the longitudinal and lateral safe distances from the i-th adjacent vehicle, respectively.

[0066] and Specifically defined as:

[0067]

[0068] In the formula, X0 is the minimum longitudinal distance, Y0 is the minimum lateral distance, T0 represents the safety time interval, and u e θ represents the relative longitudinal velocity between the vehicle and the adjacent vehicle. e The relative heading angle between the vehicle and the adjacent vehicle.

[0069] This constitutes the positional constraint that guarantees the highest priority no-collision (r1) rule requirement. The rules considered in this embodiment also include: lane keeping (r2), meeting the maximum speed limit (r3), meeting the minimum speed limit (r4), and smooth, comfortable driving (r5), collectively forming the constraint set: S = {r1, r2, r3, r4, r5}. The rules are divided into multiple equivalence classes according to their importance: C1 = {r4}, C2 = {r5}, C3 = {r2, r3}, C4 = {r1}. The relaxed rule constraint set R1 = S, R2 = S / {C1}, R3 = S / {C1, C2}, R4 = S / {C3}, R5 = S / {C1, C3}, R6 = / {C1, C2, C3}, as follows. Figure 4 As shown.

[0070] The objective function of the designed control method is:

[0071]

[0072] In the formula, N t To control the time domain, u k The control vector at time k includes the longitudinal control input jerk and the lateral control input front wheel steering angle acceleration, y k Let be the state error vector at time k, which includes lateral tracking error and heading angle tracking error, and its reference trajectory is the centerline of the target lane.

[0073] Imposing a set of constraints on the objective function constitutes a general quadratic programming problem. The difference is that, within a specified solution time Δt, if the current constraint set R... i If no feasible solution exists, then relax the constraints and use constraint set R. i+1 Solve the problem again, and repeat this process until a feasible solution is found.

[0074] Note that the initial constraint set is R1, i.e., the total constraint set. The control sequence obtained under this constraint is equivalent to the solution set obtained by the general regulatory method under the condition of satisfying the constraint. This shows that our rule-downgraded regulatory method has general applicability and can safely and smoothly give the control instruction set while satisfying more rule requirements.

[0075] This embodiment first conducted an experiment on a designed highway scenario based on the hardware-in-the-loop experimental platform in the Foshan Xianhu Park of the laboratory. By having human-driven vehicles actively create rule-violation scenarios, the feasibility and safety of the algorithm were verified. Secondly, the algorithm was deployed on the GAC Aion real vehicle platform in the laboratory. Within the park, the entry scenario was generated by remotely controlling the Ling Shi mini-car. To ensure safety, the speed of both vehicles was limited to within 5km / h.

[0076] Example 3:

[0077] The rule-based degradation control system for trajectory prediction uncertainty in this invention is mainly used to implement the above-mentioned method embodiments, and mainly includes:

[0078] The module for predicting the future occupancy area of ​​adjacent vehicles is used to input the driving trajectories of the target vehicle and its adjacent vehicles within a specific time period in the past into a pre-trained LSTM neural network, and output a Gaussian distribution probability representing the predicted horizontal and vertical positions of the target vehicle and its adjacent vehicles.

[0079] The vehicle risk field construction module is used to multiply the Gaussian distribution probability of the vehicle by the preset artificial potential field function of the vehicle, and output the vehicle risk field with time information.

[0080] The relaxed rule constraint module combines the adjacent vehicle risk field with rules other than collision avoidance rules to form a relaxed rule constraint set for the target vehicle at future moments;

[0081] Among them, the rules S other than collision avoidance rules include no collision (r1), lane keeping (r2), meeting the maximum speed limit (r3), meeting the minimum speed limit (r4), and smooth and comfortable driving (r5), S={r1,r2,r3,r4,r5};

[0082] Rule S is divided into multiple equivalence classes according to its importance, including C1 = {r4}, C2 = {r5}, C3 = {r2, r3}, and C4 = {r1}. The constraint sets for the relaxation rules are constructed according to the importance of the equivalence class combinations: R1 = S, R2 = S / {C1}, R3 = S / {C1, C2}, R4 = S / {C3}, R5 = S / {C1, C3}, and R6 = S / {C1, C2, C3}. When there is no feasible solution under the constraint set R1 of the relaxation rules, the constraint set is switched to R2 to continue solving, and this process is repeated recursively.

[0083] The rule-downgraded control module is used to design the objective function, introduce a set of relaxed rule constraints to impose conditional constraints, and form a quadratic programming problem. When there is no solution to the quadratic programming under the existing constraint set, the constraints are relaxed in sequence according to the importance of the rules and the problem is solved again. Finally, a set of feasible control sequences that satisfy the rules are obtained to control the driving of the target vehicle.

[0084] The functions of each module mainly correspond to the method implementation examples, and will not be elaborated here.

[0085] Example 4

[0086] This application also provides a computer-readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, server, app store, etc., which stores a computer program. When the program is executed by a processor, it performs corresponding functions. The computer-readable storage medium of this embodiment is used to implement the rule-based degradation control method considering trajectory prediction uncertainty in the method embodiment when executed by a processor.

[0087] In summary, this invention addresses the various rule constraints that autonomous vehicles should meet. It utilizes a pre-trained neural network to generate a bivariate Gaussian probability distribution function for the future trajectory of a neighboring vehicle, and combines this with a potential field function to generate the driving space that the neighboring vehicle will occupy. This serves as the highest collision avoidance safety constraint, supplemented by other safety and comfort-related driving rule information as additional constraints. Combined with the objective function, a smooth control sequence that the autonomous vehicle can execute in the future is generated. Within a specified solution time Δt, if no feasible solution exists under the current constraints, the rule constraints are relaxed and the solution is recalculated. This process iterates until a feasible solution is found, thus ensuring the smooth generation of the control sequence under multiple constraints. This avoids the impact on the continuity of autonomous vehicle control, passenger comfort, and the normal driving of other traffic participants caused by methods that directly abandon all rules except collision avoidance to generate emergency safety control.

[0088] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0089] The order of the steps in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0090] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A rule-based degradation control method considering trajectory prediction uncertainty, characterized in that, Includes the following steps: S1. Input the driving trajectories of the target vehicle and its adjacent vehicles within a specific time period in the past into a pre-trained neural network, and output the Gaussian distribution probability representing the predicted horizontal and vertical positions of the target vehicle and its adjacent vehicles. S2. Multiply the Gaussian distribution probability of the adjacent vehicle by the preset artificial potential field function of the adjacent vehicle to output the adjacent vehicle risk field with time information; specifically, at each prediction time, sample the distribution probability of the predicted trajectory of the adjacent vehicle. The probability of the mean position point and its four symmetrical points (front, back, left, and right) is 0.

7. The location points are determined, and the probability values ​​of the five selected location points are respectively compared with the vehicle potential field function. Multiplying them together, we get five new potential fields. Finally, we take the maximum value at the point where the five new potential fields overlap as the potential field value of the vehicle at that moment. The potential field values ​​of all the vehicles together constitute the risk field of the vehicles to avoid collisions, which is used to form the position constraints of the collision-free rules. S3. Combine the adjacent vehicle risk field with rules other than collision avoidance rules to form a relaxed rule constraint set for the target vehicle at future moments; Among them, the rule corresponding to the side-by-side risk field is no collision. Rules other than collision avoidance rules include no-collision rules. Lane keeping Meets maximum speed limits Meet the minimum speed limit Smooth and comfortable driving Then the total rule set ; The general rule set Based on importance, they are divided into multiple equivalence class combinations, including , , , Construct the constraint set for the relaxation rules according to the importance of equivalence class combinations: , / , / , / , / , / ; When in the set of constraints of relaxed rules When no feasible solution is found, switch the constraint set to Continue solving, and so on recursively; S4. Design the objective function, introduce the constructed relaxed rule constraint set for condition constraints, and form a quadratic programming problem. S5. When there is no solution to the quadratic programming under the existing constraint set, the constraints are relaxed in turn according to the importance of the rules and the solution is solved again. Finally, a set of feasible control sequences that satisfy the rules is obtained to control the driving of the target vehicle.

2. The rule-based degradation control method considering trajectory prediction uncertainty according to claim 1, characterized in that, The neural network is an LSTM neural network.

3. The rule-based degradation control method considering trajectory prediction uncertainty according to claim 2, characterized in that, The neural network has n attention heads. The output of each attention head decodes a set of predicted trajectories of the target vehicle in the future time T, with a probability that approximates the true value. Finally, the set of trajectories with the closest probability is selected.

4. The rule-based degradation control method considering trajectory prediction uncertainty according to claim 1, characterized in that, Artificial potential field function of the vehicle as follows: In the formula, The first Potential field strength parameters and shape parameters of the vehicle next to it; () represents the coordinates of any position in the artificial potential field of the vehicle. and The first The horizontal and vertical coordinates of the vehicle next to it; and They are respectively the distances from the first The longitudinal and lateral safety distances of vehicles alongside each other.

5. The rule-based degradation control method considering trajectory prediction uncertainty according to claim 4, characterized in that, Will and Defined as: In the formula, It is the minimum vertical distance. It is the minimum lateral distance. Indicated as a safety time interval, Represents the relative longitudinal velocity between the vehicle and the adjacent vehicle. The relative heading angle between the vehicle and the adjacent vehicle.

6. The rule-based degradation control method considering trajectory prediction uncertainty according to claim 1, characterized in that, The objective function designed in step S4 is: In the formula, To control the time domain, The control vector at time k includes the longitudinal control input acceleration and the lateral control input front wheel steering angle acceleration. Let be the state error vector at time k, which includes lateral tracking error and heading angle tracking error, and its reference trajectory is the centerline of the target lane.

7. The rule-based degradation control method considering trajectory prediction uncertainty according to claim 1, characterized in that, When predicting the trajectory of a vehicle, the vehicle being predicted is taken as the target vehicle, and a limited number of vehicles around the target vehicle are considered to form a prediction network.

8. A rule-based degradation control system considering trajectory prediction uncertainty, characterized in that, include: The module for predicting the future occupancy area of ​​adjacent vehicles is used to input the driving trajectories of the target vehicle and its adjacent vehicles within a specific time period in the past into a pre-trained LSTM neural network, and output a Gaussian distribution probability representing the predicted horizontal and vertical positions of the target vehicle and its adjacent vehicles. The vehicle-side risk field construction module is used to multiply the Gaussian distribution probability of the vehicle-side vehicle with the preset artificial potential field function of the vehicle-side vehicle, and output the vehicle-side risk field with time information; specifically, it samples the distribution probability of the predicted trajectory of the vehicle-side vehicle at each prediction time. The probability of the mean position point and its four symmetrical points (front, back, left, and right) is 0.

7. The location points are determined, and the probability values ​​of the five selected location points are respectively compared with the vehicle potential field function. Multiplying them together, we get five new potential fields. Finally, we take the maximum value at the point where the five new potential fields overlap as the potential field value of the vehicle at that moment. The potential field values ​​of all the vehicles together constitute the risk field of the vehicles to avoid collisions, which is used to form the position constraints of the collision-free rules. The relaxed rule constraint module combines the adjacent vehicle risk field with rules other than collision avoidance rules to form a relaxed rule constraint set for the target vehicle at future moments; Among them, the rule corresponding to the side-by-side risk field is no collision. Rules other than collision avoidance rules include no-collision rules. Lane keeping Meets maximum speed limits Meet the minimum speed limit Smooth and comfortable driving Then the total rule set ; The general rule set Based on importance, they are divided into multiple equivalence class combinations, including , , , Construct the constraint set for the relaxation rules according to the importance of equivalence class combinations: , / , / , / , / , / ; When in the set of constraints of relaxed rules When no feasible solution is found, switch the constraint set to Continue solving, and so on recursively; The rule-downgraded control module is used to design the objective function, introduce a set of relaxed rule constraints to impose conditional constraints, and form a quadratic programming problem. When there is no solution to the quadratic programming under the existing constraint set, the constraints are relaxed in sequence according to the importance of the rules and the problem is solved again. Finally, a set of feasible control sequences that satisfy the rules are obtained to control the driving of the target vehicle.

9. A computer storage medium, characterized in that, It contains a computer program that can be executed by a processor, which performs the rule-based degradation control method considering trajectory prediction uncertainty as described in any one of claims 1-7.

Images