A risk lane-changing scene generalization generation method fusing vehicle dynamics constraints

Abstract

The application relates to a risk lane-changing scene generalization generation method fusing vehicle dynamics constraints, test scene data is generated by using a TimeGAN model fusing dynamics constraints based on a vehicle lane-changing trajectory sequence, and the training method of the model is as follows: trajectory data of a lane-changing scene is subjected to clustering analysis of lane-changing styles to construct a training data set; a vehicle control dynamics model, vehicle lane-changing stability constraints and vehicle actuator capability constraints are constructed; the TimeGAN is trained by using the training data set, and a comprehensive loss function is introduced in the training process to update model parameters; the comprehensive loss function is a dynamics constraint loss function, and the dynamics model is constructed based on the vehicle control dynamics model, the vehicle lane-changing stability constraints and the vehicle actuator capability constraints. Compared with the prior art, the application can generate vehicle lane-changing test scenes which consider the authenticity of lane-changing styles and the controllable confrontation intensity close to the critical point.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving car testing, in particular to a risk lane-changing scene generalization generation method fusing vehicle dynamics constraints. BACKGROUND

[0002] With the rapid development of automatic driving technology, testing plays an increasingly important role in verifying the safety of automatic driving systems. As a typical representative of high-risk interaction situations, the lane-changing scene often involves complex decision-making games and trajectory intersections between the host vehicle and the target vehicle, and becomes a key to the safety evaluation of automatic driving. Existing risk lane-changing scene generation methods mainly include polynomial / spline curve optimization-based methods and data-driven learning methods based on real lane-changing trajectories. For example, Chinese patent application CN120278193A provides an automatic driving risk lane-changing test scene generation method, which generates lane-changing background vehicle (BV) trajectories with human driving characteristics by improving TimeGAN, and then calculates the initial state of the measured automatic driving vehicle (AV) through a critical safety distance model, finally generates a high-risk critical lane-changing test scene. Although this method solves the problem of generating multi-directional lane-changing angles in the prior art and meets the demand for automatic driving high-risk testing, it still has the following problems: 1) only aiming to fit the distribution of real trajectory data, and improving the similarity between generated trajectories and real trajectories by optimizing the network structure and loss function, which may cause the generated BV lane-changing trajectories to exceed the dynamics capability of real vehicles, resulting in that the test results of the automatic driving system in this scene cannot reflect its actual road performance, and the generated trajectories cannot be reproduced by real vehicles; 2) the critical safety distance model only focuses on the geometric constraints and speed constraints of AV and BV to avoid collision, but does not consider the stability boundary of the vehicle when changing lanes, which may lead to test results that cannot verify the performance of the vehicle under real physical constraints.

[0003] Therefore, how to provide a test scene generation method that can be accurately reproduced and improve the credibility of test results is a technical problem to be solved. SUMMARY

[0004] The purpose of the present application is to overcome the defects of the prior art and provide a risk lane-changing scene generalization generation method fusing vehicle dynamics constraints. On the basis of maintaining data-driven flexibility and trajectory diversity, the model is trained by introducing a dynamics prior model and a stability constraint loss function, so that the test scene generated by the model can be reproduced and the feasibility of the test is guaranteed.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] The application provides a risk lane-changing scene generalization generation method fusing vehicle dynamics constraints.

[0007] Obtain lane-changing scene trajectory data, perform lane-changing style cluster analysis on the lane-changing scene trajectory data, and construct a training data set based on the cluster analysis result.

[0008] A vehicle control dynamics model is constructed in a Frenet coordinate system, and vehicle lane-changing stability constraints and vehicle actuator capability constraints are constructed; the vehicle lane-changing stability constraints include a mass center side slip angle constraint and a yaw angular velocity constraint.

[0009] The training data set is used to train TimeGAN, and model parameter updating is performed based on a comprehensive loss function during the training process to obtain a TimeGAN model fusing dynamics constraints; the comprehensive loss function includes a reconstruction loss function, an adversarial loss function, a supervised loss function and a dynamics constraint loss function; a tracking loss term of the dynamics constraint loss function is constructed based on the vehicle control dynamics model and the reconstructed vehicle lane-changing trajectory, a lower limit constraint set and an upper limit constraint set are constructed based on the vehicle lane-changing stability constraints and the vehicle actuator capability constraints, and a lower limit constraint loss term and an upper limit constraint loss term of the dynamics constraint loss function are constructed based on the lower limit constraint set and the upper limit constraint set.

[0010] As a preferred technical solution, the cluster analysis method comprises:

[0011] For each lane-changing trajectory in the lane-changing scene trajectory data, key features are extracted, the key features including an average longitudinal velocity and a lane-changing duration; the average longitudinal velocity is calculated based on the longitudinal velocity of each sampling point between the starting point and the ending point of vehicle lane-changing, and the lane-changing duration is calculated based on the time when the vehicle reaches the starting point and the ending point of lane-changing;

[0012] The key features are converted into two-dimensional feature vectors, and the two-dimensional feature vectors are standardized to obtain standard feature vectors.

[0013] The lane-changing scene is divided into multiple regions, and for each region, DBSCAN is used to perform lane-changing style cluster analysis based on the standard feature vectors; the lane-changing style includes aggressive, normal and conservative.

[0014] As a preferred technical solution, the method for constructing the vehicle control dynamics model is:

[0015] The initial model under the Frenet coordinate system is constructed, and the initial model is:

[0016] ;

[0017] wherein, represents the speed of the lateral displacement error changing with time; represents the vehicle longitudinal speed; represents the yaw angle error of the current position of the vehicle and the nearest reference path point; represents the vehicle lateral speed; represents the speed of the heading error changing with time; represents the yaw rate; represents the curvature at the reference path point; represents the projection speed of the vehicle along the reference path direction; represents the lateral displacement error of the current position of the vehicle and the nearest reference path point; represents the speed of the vehicle longitudinal speed changing with time; represents the vehicle longitudinal acceleration; represents the speed of the vehicle lateral speed changing with time; represents the front wheel lateral force; represents the rear wheel lateral force; represents the total mass of the vehicle; represents the speed of the yaw rate changing with time; represents the longitudinal distance from the vehicle center of mass to the front axle; represents the longitudinal distance from the vehicle center of mass to the rear axle; represents the moment of inertia around the vertical direction;

[0018] Based on the initial model, a vehicle dynamic nonlinear function form is constructed, that is, a vehicle control dynamics model, and the expression is: , represents a state variable , and has ; represents a control input, and has , represents the front wheel steering angle rate.

[0019] As a preferred technical solution, the method for constraining the yaw rate is:

[0020] Under the small angle assumption, the vehicle center of mass side slip angle rate is calculated, and the expression is:

[0021] ;

[0022] wherein, denotes a vehicle mass center side slip angle rate of change; denotes a rear wheel lateral force; denotes a front wheel lateral force; denotes a vehicle total mass; denotes a vehicle longitudinal speed; denotes a yaw rate;

[0023] Let the vehicle mass center side slip angle rate of change be zero, the steady-state yaw rate of the vehicle driving process is calculated as:

[0024] ;

[0025] denotes a steady-state yaw rate;

[0026] When the vehicle has no longitudinal slip, the tire longitudinal force is zero, and the vehicle lateral force limit is , based on the vehicle lateral force limit and the steady-state yaw rate, the yaw rate limit boundary is calculated as:

[0027] ;

[0028] denotes a yaw rate limit boundary value; denotes a road adhesion coefficient; denotes a gravitational acceleration.

[0029] As a preferred technical solution, the method for constructing the mass center side slip angle constraint is:

[0030] The relationship function between the vehicle mass center side slip angle and the rear wheel side slip angle is constructed as:

[0031] ;

[0032] wherein, denotes a rear wheel side slip angle; denotes a vehicle lateral speed; denotes a rear axle track; denotes a yaw rate; denotes a vehicle longitudinal speed; denotes a vehicle mass center side slip angle;

[0033] Based on the Pacejka tire model, under the pure side slip condition, the tire maximum lateral force function is constructed as:

[0034] ;

[0035] denotes a tire maximum lateral force; denotes a road adhesion coefficient; represents the normal load of the tire;

[0036] linearizing the tire maximum lateral force function in a small slip angle region, we have:

[0037] ;

[0038] wherein, represents the lateral force of the tire; represents the equivalent cornering stiffness of the rear wheel; the small slip angle region refers to a region in which the vehicle mass center slip angle is less than a slip angle threshold value;

[0039] the approximate limit slip angle of the rear wheel when the vehicle generates the maximum lateral force is:

[0040] ;

[0041] wherein, represents the approximate limit slip angle of the rear wheel; represents the road adhesion coefficient; represents the gravity acceleration; represents the total mass of the vehicle; represents the front axle track; represents the rear axle track;

[0042] based on the approximate limit slip angle of the rear wheel, the mass center slip angle limit boundary is calculated as:

[0043] ;

[0044] wherein, represents the mass center slip angle limit boundary.

[0045] As a preferred technical solution, the vehicle actuator capability constraint comprises:

[0046] a front wheel maximum rotation angle constraint, i.e., the absolute value of the front wheel rotation angle is less than or equal to the maximum value of the front wheel rotation angle;

[0047] a longitudinal vehicle speed constraint, i.e., the longitudinal speed of the vehicle at any path point is a non-zero value less than or equal to the maximum longitudinal speed;

[0048] an acceleration constraint, i.e., the longitudinal acceleration value of the vehicle at any path point is between the maximum braking acceleration and the maximum driving acceleration; and the maximum braking acceleration is less than the maximum driving acceleration.

[0049] As a preferred technical solution, the reconstruction loss function is:

[0050] ;

[0051] wherein, an expected value of the vehicle lane-changing trajectory, a real data distribution of the vehicle lane-changing trajectory; a vehicle lane-changing trajectory at time t, a reconstructed vehicle lane-changing trajectory.

[0052] As a preferred technical solution, the adversarial loss function is:

[0053] ;

[0054] wherein, an adversarial loss function, an expected value of the vehicle lane-changing trajectory, a real data distribution of the vehicle lane-changing trajectory, a discriminative output probability of the TimeGAN for a real vehicle lane-changing trajectory, a mathematical expectation of the vehicle lane-changing trajectory under the probability distribution obtained by training, a discriminative output probability of the discriminative model for a real vehicle lane-changing trajectory, a probability distribution of the vehicle lane-changing trajectory obtained by training the TimeGAN.

[0055] As a preferred technical solution, the supervised loss function is:

[0056] ;

[0057] a supervised loss, an expected value of all static features and the vehicle lane-changing trajectory, a low-dimensional latent vector at time t obtained based on input data, an approximate expectation calculation function, denoted as , a random noise at time t, a random noise distribution, a probability distribution of the vehicle lane-changing trajectory obtained by training the TimeGAN, a latent variable generated by the TimeGAN at time t, a history latent variable sequence generated by the TimeGAN from 1 to , a low-dimensional latent state vector generated by the TimeGAN at time t-1.

[0058] As a preferred technical solution, the dynamics constraint loss function is:

[0059] ;

[0060] wherein, denotes a dynamics constraint loss function; denotes a tracking loss term; denotes a total time length; denotes a reconstructed vehicle lane-changing trajectory at time t+1; denotes a vehicle control dynamics model; denotes a reconstructed vehicle lane-changing trajectory at time t; denotes a control input of the reconstructed vehicle control dynamics model; denotes a lower bound constraint loss term; denotes an upper bound constraint loss term; denotes a constraint penalty coefficient; and each respectively denotes a lower bound constraint set and an upper bound constraint set, for there is:

[0061] ;

[0062] denotes a reconstructed front wheel steering angle at time t; denotes a front wheel steering angle maximum value; denotes a front wheel steering angle lower bound value; denotes a maximum braking acceleration; denotes a reconstructed longitudinal acceleration at time t; denotes a longitudinal acceleration lower bound value; denotes a reconstructed vehicle longitudinal speed at time t; denotes a vehicle longitudinal speed lower bound value; denotes a vehicle lateral speed minimum value; denotes a reconstructed vehicle lateral speed at time t; denotes a vehicle lateral speed lower bound value;

[0063] for there is:

[0064] ;

[0065] denotes a front wheel steering angle upper bound value; denotes a maximum driving acceleration; denotes a longitudinal acceleration upper bound value; denotes a vehicle longitudinal speed upper bound value; denotes a vehicle longitudinal speed maximum value; denotes a vehicle lateral speed maximum value; denotes a vehicle lateral speed upper bound value.

[0066] Compared with the prior art, the present application has the following beneficial effects:

[0067] 1) The present application constructs a fusion kinetics constraint TimeGAN model for automatic driving risk lane changing scene generation. Specifically, the fusion kinetics constraint TimeGAN model is trained using lane changing trajectory data set with lane changing style label, and a kinetics constraint loss function is introduced to guide the model parameter update. The loss function is constructed based on the vehicle control kinetics model, vehicle lane changing stability constraint and vehicle actuator capability constraint, so that the TimeGAN model has the ability to generate trajectories that consider lane changing style authenticity and close-to-critical but controllable confrontation strength, and meet the vehicle lane changing stability realization boundary in the dimensions of vehicle lane changing stability and vehicle actuator capability, ensuring physical executability and practicality.

[0068] 2) The present application respectively constrains the vehicle lane changing stability and the vehicle actuator capability, generates trajectories that meet the vehicle realization boundary in the dimensions of speed, acceleration, yaw and lateral velocity by means of the joint constraint of vehicle dynamics and stability envelope, and improves the test executability. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1 The flowchart of the method provided by the present application is shown in the figure;

[0070] Figure 2 The stability phase plane envelope diagram of the present application is shown in the figure;

[0071] Figure 3 The TimeGAN training flowchart of the present application is shown in the figure. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the present application.

[0073] The present application provides a risk lane changing scene generalization generation method fusing vehicle dynamics constraint. The method generates test scene data based on vehicle lane changing trajectory sequence using a fusion kinetics constraint TimeGAN model, and after obtaining the scene data, uses TTC (Time to Collision) index as risk criterion, and converts the scene parameters, initial position and speed conditions, risk level and dynamics constraint in the above scene data into test case description file at the same time. The file can be directly imported into simulation or real vehicle test platform to realize batch risk evaluation of automatic driving system.

[0074] To avoid the test scene generated by the fusion dynamics constraint TimeGAN model finally having low landing execution and the generated lane change trajectory not meeting the stability constraint of vehicle driving, a loss function based on a physical dynamics prior control model and stability constraint is introduced in the present application to train the TimeGAN. The training method flow is as shown in Figure 1 The training method flow is as shown in

[0075] S1, lane change scene trajectory data is obtained, lane change style clustering analysis is performed on the lane change scene trajectory data, and a training data set is constructed based on the clustering analysis result.

[0076] S11, for each lane change trajectory in the lane change scene trajectory data, key features are extracted.

[0077] The key features include average longitudinal velocity and lane change duration, wherein the average longitudinal velocity is calculated based on the longitudinal velocity of each sampling point between the lane change start point and the end point of the vehicle, and the expression is:

[0078] ;

[0079] Vk represents the longitudinal velocity of the sampling point k, N represents the total number of sampling points, V represents the average longitudinal velocity.

[0080] The lane change duration is calculated based on the time when the vehicle reaches the lane change start point and the time when the vehicle reaches the lane change end point, and the expression is: , T represents the time when the vehicle reaches the lane change end point, T0 represents the time when the vehicle reaches the lane change start point.

[0081] S12, the key features of each lane change trajectory are converted into a two-dimensional feature vector Since the average longitudinal velocity and the lane change duration have different dimensions, direct calculation of the distance will cause deviation, so the two-dimensional feature vector is first standardized by Z-score, so that the two types of features are in the same scale, and a standard feature vector is obtained.

[0082] S13, the lane change scene is divided into multiple regions, and for each region, DBSCAN is used to perform lane change style clustering analysis based on the standard feature vector, and the lane change style includes aggressive, normal and conservative.

[0083] The neighborhood radius is set as ε With two core parameters of minimum neighborhood sample number Min_samples, DBSCAN is able to identify the areas with higher density in the normalized feature space and automatically form different lane-changing style clusters. When a large number of trajectories form a sufficiently dense distribution in a certain area, the area is identified as a natural form of lane-changing style.

[0084] In the embodiment, the lane-changing speed is fast and the time is short, which corresponds to the aggressive style; the speed is moderate and the time is moderate, which corresponds to the ordinary style; the speed is slow and the time is long, which represents the conservative style; and the abnormal trajectory with sparse distribution is marked as a noise point. Through this density structure-based clustering method, the number of style categories does not need to be set in advance, and the identification and labeling of lane-changing styles can be automatically completed according to the natural distribution of real data.

[0085] S2, a vehicle control dynamics model is constructed under the Frenet coordinate system, and a vehicle lane-changing stability constraint and a vehicle actuator capability constraint are constructed.

[0086] S21, a vehicle control dynamics model is constructed.

[0087] For the geometric relationship between the vehicle motion process and the reference line, a vehicle control dynamics model under the Frenet coordinate system is established in combination with a single-track handling dynamics model of the vehicle, including:

[0088] S211, an initial model under the Frenet coordinate system is constructed, which is:

[0089] ;

[0090] wherein, represents the speed of the lateral displacement error changing with time; represents the vehicle longitudinal speed; represents the yaw angle error of the current position of the vehicle and the nearest reference path point; represents the vehicle lateral speed; represents the speed of the heading error changing with time; represents the yaw rate; represents the curvature at the reference path point; represents the projection speed of the vehicle along the reference path direction; represents the lateral displacement error of the current position of the vehicle and the nearest reference path point; represents the speed of the vehicle longitudinal speed changing with time; represents the vehicle longitudinal acceleration; represents the speed of the vehicle lateral speed changing with time; represents the front wheel lateral force; represents the rear wheel lateral force; represents the total mass of the vehicle; denotes the yaw rate; denotes the longitudinal distance from the vehicle center of mass to the front axle; denotes the longitudinal distance from the vehicle center of mass to the rear axle; denotes the moment of inertia around the vertical direction.

[0091] S212, based on the initial model, a vehicle dynamic nonlinear function form is constructed, that is, a vehicle control dynamics model, and its expression is: , denotes the state variable the first order derivative of time, and has ; denotes the control input, and has , denotes the front wheel steering angle rate.

[0092] S22, a vehicle lane changing stability constraint is constructed.

[0093] In order to ensure the stability of the vehicle during the lane changing execution process, it is necessary to analyze the nonlinear dynamics characteristics of steering instability and determine the stable region of lateral motion state convergence. In order to characterize the vehicle stable region, the phase plane method is used to determine the vehicle stability envelope, which is determined by the vehicle yaw rate r and the center of mass side slip angle . As long as the vehicle state is always within the envelope range as shown in Figure 2 , the stable driving state can be maintained, therefore, the vehicle lane changing stability constraint including the center of mass side slip angle constraint and the yaw rate constraint is constructed.

[0094] S221, a yaw rate constraint is constructed.

[0095] Under the small angle assumption, the vehicle center of mass side slip angle rate is calculated, and its expression is:

[0096] ;

[0097] wherein, denotes the vehicle center of mass side slip angle rate; denotes the rear wheel lateral force; denotes the front wheel lateral force; denotes the total mass of the vehicle; denotes the vehicle longitudinal speed; denotes the yaw rate.

[0098] When the vehicle center of mass side slip angle rate is zero, that is, , the steady-state yaw rate of the vehicle driving process is calculated, and is:

[0099] ;

[0100] denotes the steady-state yaw rate.

[0101] When the vehicle has no longitudinal slip, the tire longitudinal force is zero, then the vehicle lateral force limit is The yaw rate limit boundary is calculated based on the vehicle lateral force limit and the steady-state yaw rate, which is

[0102] ;

[0103] denotes the yaw rate limit boundary value; denotes the road adhesion coefficient; denotes the gravity acceleration.

[0104] S222, constructing the center of mass side slip angle constraint.

[0105] The relationship function between the vehicle center of mass side slip angle and the rear wheel side slip angle is constructed, which is

[0106] ;

[0107] wherein, denotes the rear wheel side slip angle; denotes the vehicle lateral velocity; denotes the rear axle track; denotes the yaw rate; denotes the vehicle longitudinal velocity; denotes the vehicle center of mass side slip angle.

[0108] Based on the Pacejka tire model, the tire maximum lateral force function is constructed under the pure side slip condition, which is

[0109] ;

[0110] denotes the tire maximum lateral force; denotes the road adhesion coefficient; denotes the tire normal load.

[0111] Linearizing the tire maximum lateral force function in the small side slip angle region, then there is

[0112] ;

[0113] wherein, denotes the tire lateral force; denotes the rear wheel equivalent side slip stiffness; the small side slip angle region refers to the region where the vehicle center of mass side slip angle is less than the side slip angle threshold.

[0114] The rear wheel approximate limit side slip angle when the vehicle generates the maximum lateral force is:

[0115]

[0116] wherein, represents the rear wheel approximate limit side slip angle; represents the road surface adhesion coefficient; represents the gravity acceleration; represents the total mass of the vehicle; represents the front axle track; represents the rear axle track.

[0117] Based on the rear wheel approximate limit side slip angle, the center of mass side slip angle limit boundary is calculated as:

[0118]

[0119] wherein, represents the center of mass side slip angle limit boundary.

[0120] S23, vehicle actuator capability constraints.

[0121] The scenario execution is subject to the vehicle bottom layer actuator completion, so the vehicle actuator capability constraints also need to be considered, and the vehicle actuator capability constraints include:

[0122] The front wheel maximum turning angle constraint, i.e., the absolute value of the front wheel turning angle is less than or equal to the maximum front wheel turning angle, can be represented as: , represents the front wheel turning angle; represents the maximum front wheel turning angle.

[0123] The longitudinal vehicle speed constraint, i.e., the longitudinal speed of the vehicle at any path point is a non-zero value less than or equal to the maximum longitudinal speed , i.e.: .

[0124] The acceleration constraint, i.e., the longitudinal acceleration value of the vehicle at any path point is between the maximum braking acceleration and the maximum driving acceleration, and the maximum braking acceleration is less than the maximum driving acceleration , can be represented as: .

[0125] S3, training TimeGAN using the training data set, the process is as shown in Figure 3 , and the model parameter update is carried out based on the comprehensive loss function during the training process to obtain the fusion dynamics constraint TimeGAN model.

[0126] ​​A set of training data is randomly sampled from the training data set for TimeGAN training, specifically, TimeGAN includes an encoder, a generator, a decoder, a supervisor and a discriminator, and the following steps are performed in TimeGAN:

[0127] i) The lane-changing trajectory sequence of the vehicle in the training data and the lane-changing style label corresponding to the trajectory are spliced to obtain input data, which is input into the encoder, and the high-dimensional trajectory state is encoded into a low-dimensional latent vector sequence through a multi-layer recurrent neural network structure , represents a low-dimensional forward vector sequence.

[0128] ii) A set of random noise sequences is randomly generated, including random noise and corresponding lane-changing style labels, and the generator generates a latent trajectory sequence based on the random noise sequence , represents a latent trajectory.

[0129] iii) The supervisor receives the latent trajectory sequence and outputs the time-dependent latent representation of the next time point of each latent trajectory through a one-step time prediction structure , represents the processing process of the supervisor.

[0130] iiii) The decoder receives the time prediction latent representation from the supervisor , and decodes to generate a complete trajectory sequence .

[0131] iiiii) The discriminator receives , and and classifies them to push the generator and the supervisor to constantly approach the real trajectory distribution.

[0132] After the above process is performed, a comprehensive loss function is calculated, which includes:

[0133] The reconstruction loss function is used to measure the difference between the reconstructed data and the original data, and its expression is:

[0134] ;

[0135] wherein, represents the expected value of the vehicle lane-changing trajectory, represents the real data distribution of the vehicle lane-changing trajectory; represents the vehicle lane-changing trajectory at time t; represents the reconstructed vehicle lane-changing trajectory; , represents the Euclidean distance between the real vehicle lane-changing trajectory and the reconstructed vehicle lane-changing trajectory.

[0136] The adversarial loss function is expressed as follows:

[0137] ;

[0138] in, Represents the adversarial loss function; This represents the expected value of the vehicle's lane-changing trajectory; The actual data distribution representing the vehicle lane-changing trajectory; This represents the probability output by TimeGAN that the lane-changing trajectory of a real vehicle is a real sample. This represents the mathematical expectation of the vehicle lane-changing trajectory under the probability distribution obtained from training. This represents the probability output by the discriminant model that the reconstructed vehicle lane-changing trajectory is a real sample; This represents the probability distribution of vehicle lane-changing trajectories obtained from TimeGAN training.

[0139] The supervised loss function is expressed as follows:

[0140] ;

[0141] Indicates monitoring losses; This represents the expected value of all static features and vehicle lane-changing trajectories; This represents the low-dimensional latent vector obtained at time t based on the input data; The approximate expectation calculation function is expressed as: , This represents random noise at time t. Represents a random noise distribution. This represents the probability distribution of vehicle lane-changing trajectories obtained from TimeGAN training. This represents the latent variable generated by TimeGAN at time t. This indicates that the values ​​generated by TimeGAN range from 1 to... Historical latent variable sequence, This represents the low-dimensional latent state vector generated by TimeGAN at time t-1.

[0142] The dynamic constraint loss function includes a tracking loss term, a lower bound constraint loss term, and an upper bound constraint loss term. Specifically, the tracking loss term of the dynamic constraint loss function is constructed based on the vehicle control dynamics model and the reconstructed vehicle lane-changing trajectory; the lower bound constraint set and the upper bound constraint set are constructed based on the vehicle lane-changing stability constraint and the vehicle actuator capability constraint; and the lower bound constraint loss term and the upper bound constraint loss term of the dynamic constraint loss function are constructed based on the lower bound constraint set and the upper bound constraint set.

[0143] Its expression is:

[0144] ;

[0145] wherein, denotes a kinematic constraint loss function; denotes a tracking loss term; denotes a total time length; denotes a reconstructed vehicle lane-changing trajectory at time t+1; denotes a vehicle control dynamics model; denotes a reconstructed vehicle lane-changing trajectory at time t; denotes a control input of the reconstructed vehicle control dynamics model; denotes a lower bound constraint loss term; denotes an upper bound constraint loss term; denotes a constraint penalty coefficient; and each respectively denotes a lower bound constraint set and an upper bound constraint set, for there is:

[0146] ;

[0147] denotes a reconstructed front wheel steering angle at time t; denotes a front wheel steering angle maximum value; denotes a front wheel steering angle lower bound value; denotes a maximum braking acceleration; denotes a reconstructed longitudinal acceleration at time t; denotes a longitudinal acceleration lower bound value; denotes a reconstructed vehicle longitudinal speed at time t; denotes a vehicle longitudinal speed lower bound value; denotes a vehicle lateral speed minimum value; denotes a reconstructed vehicle lateral speed at time t; denotes a vehicle lateral speed lower bound value;

[0148] for there is:

[0149] ;

[0150] denotes a front wheel steering angle upper bound value; denotes a maximum driving acceleration; denotes a longitudinal acceleration upper bound value; denotes a vehicle longitudinal speed upper bound value; denotes a vehicle longitudinal speed maximum value; denotes a vehicle lateral speed maximum value; denotes a vehicle lateral speed upper bound value.

[0151] The present application also provides an electronic device including a central processing unit (CPU) that can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). Various programs and data required for device operation can also be stored in the RAM. The CPU, the ROM, and the RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0152] A plurality of components in the device are connected to the I / O interface, including an input unit such as a keyboard, a mouse, and the like; an output unit such as various types of displays, a speaker, and the like; a storage unit such as a magnetic disk, an optical disk, and the like; and a communication unit such as a network card, a modem, a wireless communication transceiver, and the like. The communication unit allows the device to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0153] The processing unit performs the various methods and processes described above, such as the methods S1-S3. For example, in some embodiments, the methods S1-S3 can be implemented as a computer software program that is tangibly embodied in a machine-readable medium, such as the storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed on the device via the ROM and / or the communication unit. When the computer program is loaded into the RAM and executed by the CPU, one or more steps of the methods S1-S3 described above can be performed. Alternatively, in other embodiments, the CPU can be configured to perform the methods S1-S3 by any other appropriate means, such as by means of firmware.

[0154] The functionality described above in this document can be performed, at least in part, by one or more hardware logic components. For example, and without limitation, example types of hardware logic components that can be used include field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip systems (SOCs), complex programmable logic devices (CPLDs), etc.

[0155] Program code for carrying out the methods of the present application can be written in any combination of one or more programming languages. The program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, causes the machine to perform the functions / acts specified in the flowcharts and / or block diagrams. The program code can execute entirely on a machine, partly on a machine, as a stand-alone software package, partly on a machine and partly on a remote machine or entirely on a remote machine or server.

[0156] In the context of the present application, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable storage medium can include, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of machine-readable storage medium would include one or more lines of electrical wire, portable computer diskette, hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the foregoing.

[0157] The above description is only specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A generalized generation method for risky lane-changing scenarios that integrates vehicle dynamics constraints, characterized in that, The method described above generates test scenario data based on vehicle lane-changing trajectory sequences using a fusion dynamics-constrained TimeGAN model. The training method for the fusion dynamics-constrained TimeGAN model includes: Acquire lane-change scene trajectory data, perform lane-change style cluster analysis on the lane-change scene trajectory data, and construct a training dataset based on the cluster analysis results; A vehicle control dynamics model is constructed in the Frenet coordinate system, and vehicle lane-changing stability constraints and vehicle actuator capability constraints are established. The vehicle lane-changing stability constraints include centroid sideslip angle constraints and yaw rate constraints. The method for constraining the yaw rate is as follows: Under the assumption of a small angle, the rate of change of the vehicle's center of gravity sideslip angle is calculated using the following expression: ； in, This indicates the rate of change of the vehicle's center of gravity sideslip angle; Indicates the lateral force on the rear wheel; Indicates the lateral force on the front wheel; Indicates the total mass of the vehicle; Indicates the longitudinal speed of the vehicle; Indicates yaw rate; Assuming the rate of change of the vehicle's center of gravity sideslip angle is zero, the steady-state yaw rate during vehicle operation is calculated as follows: ； Represents the steady-state yaw rate; When the vehicle has no longitudinal slip, the longitudinal force on the tires is zero, then the vehicle's lateral force limit is... Based on the aforementioned vehicle lateral force limit and stable yaw rate, the yaw rate limit boundary is calculated as follows: ； This represents the limit boundary value of the yaw rate; Indicates the road surface adhesion coefficient; Represents gravitational acceleration; The method for constructing the centroid sideslip angle constraint is as follows: The relationship function between the vehicle's center of gravity sideslip angle and the rear wheel sideslip angle is constructed as follows: ； in, Indicates the rear wheel slip angle; Indicates the lateral speed of the vehicle; Indicates the rear axle wheelbase; Indicates yaw rate; Indicates the longitudinal speed of the vehicle; Indicates the vehicle's sideslip angle; Based on the Pacejka tire model, under pure sideslip conditions, the maximum lateral force function of the tire is constructed as follows: ； This indicates the maximum lateral force of the tire; Indicates the road surface adhesion coefficient; Indicates the normal load on the tire; Linearizing the maximum lateral force function of the tire in the small slip angle region, we have: ； in, Indicates the lateral force of the tire; This indicates the equivalent sideslip stiffness of the rear wheel; the small sideslip angle region refers to the region where the sideslip angle of the vehicle's center of gravity is less than the sideslip angle threshold. The approximate limiting slip angle of the rear wheels when the vehicle generates maximum lateral force is: ； in, This indicates the approximate limit of the rear wheel's sideslip angle; Indicates the road surface adhesion coefficient; Represents gravitational acceleration; Indicates the total mass of the vehicle; Indicates the front axle wheelbase; Indicates the rear axle wheelbase; Based on the aforementioned approximate limiting sideslip angle of the rear wheel, the limiting boundary of the center-of-gravity sideslip angle is calculated as follows: ； in, Indicates the limiting boundary of the centroid sideslip angle; The vehicle actuator capability constraints include: The maximum front wheel steering angle constraint means that the absolute value of the front wheel steering angle is less than or equal to the maximum front wheel steering angle. Longitudinal speed constraint, which means that the longitudinal speed of the vehicle at any path point is a non-zero value that is less than or equal to the maximum longitudinal speed; Acceleration constraint, which means that the longitudinal acceleration value of the vehicle at any point on the path is between the maximum braking acceleration and the maximum driving acceleration; and the maximum braking acceleration is less than the maximum driving acceleration. The TimeGAN is trained using the aforementioned training dataset, and the model parameters are updated based on the comprehensive loss function during training to obtain a TimeGAN model with fused dynamic constraints. The comprehensive loss function includes a reconstruction loss function, an adversarial loss function, a supervision loss function, and a dynamic constraint loss function. The tracking loss term of the dynamic constraint loss function is constructed based on the aforementioned vehicle control dynamics model and the reconstructed vehicle lane-changing trajectory. The lower constraint set and the upper constraint set are constructed based on the aforementioned vehicle lane-changing stability constraints and vehicle actuator capability constraints. The lower constraint loss term and the upper constraint loss term of the dynamic constraint loss function are constructed based on the aforementioned lower constraint set and the upper constraint set. The aforementioned dynamic constraint loss function is: ； in, Represents the dynamic constraint loss function; Indicates the tracking loss term; Indicates the total duration; This represents the reconstructed vehicle's lane-changing trajectory at time t+1; Represents the vehicle control dynamics model; This represents the reconstructed vehicle's lane-changing trajectory at time t; This represents the control input for reconstructing the vehicle control dynamics model; This represents the lower bound constraint loss term; This represents the upper limit constraint loss term; Indicates the constraint penalty coefficient; and Both represent the lower bound constraint set and the upper bound constraint set, respectively. have: ； This represents the reconstructed front wheel steering angle at time t; This indicates the maximum front wheel steering angle; This indicates the lower limit of the front wheel steering angle; Indicates the maximum braking acceleration; This represents the reconstructed longitudinal acceleration at time t; This indicates the lower limit of longitudinal acceleration; This represents the longitudinal velocity of the reconstructed vehicle at time t; This indicates the lower limit of the vehicle's longitudinal speed; This represents the minimum lateral speed of the vehicle; This represents the lateral velocity of the reconstructed vehicle at time t; This indicates the lower limit of the vehicle's lateral speed; for have: ； This indicates the upper limit of the front wheel steering angle; Indicates the maximum driving acceleration; This indicates the upper limit of longitudinal acceleration; This indicates the upper limit of the vehicle's longitudinal speed; This indicates the maximum longitudinal speed of the vehicle; This indicates the maximum lateral speed of the vehicle; This indicates the maximum lateral speed limit for a vehicle.

2. The method for generalizing and generating risky lane-changing scenarios by incorporating vehicle dynamics constraints as described in claim 1, characterized in that, The clustering analysis methods include: For each lane change trajectory in the lane change scenario trajectory data, key features are extracted, including average longitudinal speed and lane change duration. The average longitudinal speed is calculated based on the longitudinal speed of each sampling point between the starting point and the ending point of the lane change, and the lane change duration is calculated based on the time it takes for the vehicle to arrive at the starting point and the ending point of the lane change. The key features are transformed into two-dimensional feature vectors, and the two-dimensional feature vectors are standardized to obtain standard feature vectors; The lane-changing scenario is divided into multiple regions. For each region, DBSCAN is used to perform lane-changing style clustering analysis based on the standard feature vectors. The lane-changing styles include aggressive, normal, and conservative.

3. The method for generalizing and generating risky lane-changing scenarios by incorporating vehicle dynamics constraints as described in claim 1, characterized in that, The method for constructing the vehicle control dynamics model is as follows: The initial model in the Frenet coordinate system is constructed as follows: ； in, This indicates the rate at which the lateral displacement error changes over time. Indicates the longitudinal speed of the vehicle; This indicates the yaw angle error between the vehicle's current position and the nearest reference path point; Indicates the lateral speed of the vehicle; This indicates the rate at which the heading error changes over time; Indicates yaw rate; Indicates the curvature at the reference path point; This represents the projected speed of the vehicle along the reference path. This represents the lateral displacement error between the vehicle's current position and the nearest reference path point; This represents the speed at which the vehicle's longitudinal velocity changes over time. Indicates the longitudinal acceleration of the vehicle; This represents the speed at which the vehicle's lateral velocity changes over time. Indicates the lateral force on the front wheel; Indicates the lateral force on the rear wheel; Indicates the total mass of the vehicle; This represents the rate at which the yaw rate changes with time. This indicates the longitudinal distance from the vehicle's center of gravity to the front axle; This indicates the longitudinal distance from the vehicle's center of gravity to the rear axle; Indicates the moment of inertia about the vertical direction; Based on the initial model, the nonlinear function form of vehicle dynamics is constructed, which is the vehicle control dynamics model, and its expression is: , Represents state variables The first derivative with respect to time, and we have ; Indicates control input, and has , This indicates the rate of change of the front wheel steering angle.

4. The method for generalizing and generating risky lane-changing scenarios by incorporating vehicle dynamics constraints according to claim 1, characterized in that, The reconstruction loss function is: ； in, This represents the expected value of the vehicle's lane-changing trajectory. The actual data distribution representing the vehicle lane-changing trajectory; This represents the vehicle's lane-changing trajectory at time t; This indicates the reconstructed vehicle lane-changing trajectory.

5. The method for generalizing and generating risky lane-changing scenarios by incorporating vehicle dynamics constraints according to claim 1, characterized in that, The adversarial loss function is as follows: ； in, Represents the adversarial loss function; This represents the expected value of the vehicle's lane-changing trajectory; The actual data distribution representing the vehicle lane-changing trajectory; This represents the probability output by TimeGAN that the lane-changing trajectory of a real vehicle is a real sample. This represents the mathematical expectation of the vehicle lane-changing trajectory under the probability distribution obtained from training. This represents the probability output by the discriminant model that the reconstructed vehicle lane-changing trajectory is a real sample; This represents the probability distribution of vehicle lane-changing trajectories obtained from TimeGAN training.

6. The method for generalizing and generating risky lane-changing scenarios by incorporating vehicle dynamics constraints according to claim 1, characterized in that, The supervised loss function is: ； Indicates monitoring losses; This represents the expected value of all static features and vehicle lane-changing trajectories; This represents the low-dimensional latent vector obtained at time t based on the input data; The approximate expectation calculation function is expressed as: , This represents random noise at time t. Represents a random noise distribution. This represents the probability distribution of vehicle lane-changing trajectories obtained from TimeGAN training. This represents the latent variable generated by TimeGAN at time t. This indicates that the values ​​generated by TimeGAN range from 1 to... Historical latent variable sequence, This represents the low-dimensional latent state vector generated by TimeGAN at time t-1.

Images