Safe key track generation method and device for lane changing of autonomous vehicle

Abstract

The invention discloses a method and a device for generating a safe key track of lane change of an automatic driving automobile, and the method comprises the steps: 1, processing a highD public data set, obtaining an original data set of driving of the automobile, and carrying out the data preprocessing of the original data set of driving of the automobile; 2, constructing a pre-training module of a generator in the self-vehicle trajectory generation model; 3, constructing an adversarial training module of a generator in the self-vehicle trajectory generation model; the confrontation training module represents the interaction relationship between the vehicle and the background vehicle based on nodes and directed edges; 4, constructing a discriminator of the self-vehicle trajectory generation model; 5, constructing a trajectory completion module of a self-vehicle trajectory generation model; and 6, constructing a track generation module based on the lane changing danger degree of the vehicle, and optimizing the short-time initial state of the vehicle under the multi-vehicle background through a cost function.

Description

Technical Field

[0001] The present invention belongs to the application field of autonomous driving vehicle simulation testing, and specifically relates to a DWA-HEO-TransGAN generation method and device for a lane-changing trajectory of an ego vehicle including multiple vehicles. Background Art

[0002] At present, the commercialization of autonomous vehicles is gradually advancing, and the technology is progressing rapidly. However, in order to obtain a legal license for commercial use on the road, a large number of simulation tests and real vehicle tests are required to ensure the safety and reliability of the technology. Fully autonomous vehicles must travel hundreds of millions of miles, sometimes even hundreds of billions of miles, to prove their reliability in terms of casualties. Due to the high requirements of autonomous vehicle testing on test efficiency and scene richness, traditional road testing can no longer be used as the main testing method. Simulation testing, with its high safety, strong controllability and repeatability, rich test scenes, and convenience of data collection and analysis, occupies an increasingly important position in the autonomous vehicle testing process. The vehicle driving trajectory in the simulation test is usually referenced to the real driving vehicle. However, the equipment for collecting the driving trajectory of the real vehicle usually requires very strict debugging, and professionals are required to monitor in real time during the collection process. The data obtained also needs to be pre-processed and format converted according to the test environment. At the same time, in order to obtain sufficient data, the data collection vehicle needs to travel hundreds of thousands of kilometers in different traffic scenes, road conditions, and weather conditions, and the collection cost is extremely high. In addition, the trajectories with high risk confrontation between the self-driving car and the background car (safety-critical trajectories) with high test value for each module of the autonomous vehicle account for a very low proportion in the normal driving process of the data collection vehicle. In contrast, the generative trajectory data for simulation test scenarios not only saves costs but also solves the scarcity problem of safety-critical driving trajectories.

[0003] Guided by scenario testing, the trajectory generation of autonomous vehicles can be divided into three categories: data-driven generation methods, knowledge-based generation methods, and adversarial generation methods. The data-driven dangerous lane change trajectory generation method requires a large amount of high-quality real driving data containing safety-critical lane change trajectories as support, but safety-critical lane change trajectories are scarce in themselves, and the model performs poorly for unseen scenarios or extreme situations. The knowledge-based generation method generates trajectories that are highly interpretable, but relies on predefined rules and cannot adapt to complex and changing driving environments. Summary of the invention

[0004] The present invention aims to solve the above-mentioned defects in the prior art and proposes a method and device for generating safety-critical trajectories for lane changing of an autonomous driving vehicle.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A first aspect of the present invention provides a method for generating a safety-critical trajectory for lane changing of an autonomous driving vehicle, comprising:

[0007] Step 1: Obtain the highD data set, obtain the original data set of the self-driving vehicle, and perform data preprocessing on the original data set of the self-driving vehicle, including removing few-frame trajectories, wavelet decomposition filtering, lane change behavior annotation, and multi-vehicle trajectory extraction on the original data set of the self-driving vehicle;

[0008] Step 2: Construct a pre-training module of the generator in the ego-vehicle trajectory generation model; wherein the pre-training module of the generator in the ego-vehicle trajectory generation model adopts the Transformer model and designs a dynamic window mechanism:

[0009] Step 3: construct an adversarial training module of the generator in the ego-vehicle trajectory generation model; wherein the adversarial training module of the ego-vehicle trajectory generation model represents the interaction relationship between the ego-vehicle and the background vehicles based on nodes and directed edges;

[0010] Step 4: Construct a discriminator for the ego-vehicle trajectory generation model; the discriminator includes a generation comparison module to deal with mode collapse;

[0011] Step 5: construct a trajectory completion module of the ego-vehicle trajectory generation model; wherein the trajectory completion module of the ego-vehicle trajectory generation model adopts a DDPG network;

[0012] Step six: construct a trajectory generation module based on the lane-changing risk level of the ego vehicle; wherein the trajectory generation module based on the lane-changing risk level of the ego vehicle optimizes the initial state of the ego vehicle in the multi-vehicle background through a cost function.

[0013] Furthermore, the data of the original data set of the vehicle driving obtained in step 1 includes the driving behavior characteristics of the vehicle and the relative driving states of the vehicle and the front and rear vehicles in the original lane of the vehicle and the target lane of the vehicle changing lane.

[0014] The data preprocessing of the original data set of the self-driving vehicle includes removing the few-frame trajectory, wavelet decomposition filtering, lane change behavior annotation and multi-vehicle trajectory extraction of the original data set of the self-driving vehicle:

[0015] The low-frame trajectory deletes the trajectories below the preset frame number threshold and retains the trajectory data with sufficient time length. The wavelet decomposition filter performs multi-layer wavelet decomposition on the trajectory data to obtain the approximate coefficients and detail coefficients of different scales, and performs inverse wavelet transform according to the selected coefficients to obtain the smoothed trajectory data. The lane change behavior annotation is based on the ID change of the lane where the ego vehicle is located. The multi-vehicle trajectory extracts the trajectories of the ego vehicle and the surrounding background as a set of driving trajectories.

[0016] Furthermore, the dynamic window mechanism of the generator pre-training module in the vehicle trajectory generation model in step 2 specifically includes:

[0017] Determine the lane change state and the start and end of the window according to the set lateral acceleration threshold and heading angular velocity threshold. Update the window center and size in real time: set the initial window length σ0 and adjust the proportional constant k. According to the lateral acceleration a of the current time step y Determine whether the lane is changing. Calculate the angular velocity ω(t) of the current time step t, and update the window length according to the angular velocity. Get the lateral coordinate y(t) and angular velocity ω(t) of the current time step t and update the window center t center . Calculate the attention weights in the dynamic window: For each time step t′, calculate the Gaussian weight function. Embed the attention weights into the generator pre-training: In the Transformer model of the generator, use a separate attention head to embed the weights w(t′).

[0018] Furthermore, the adversarial training module of the generator in the vehicle trajectory generation model described in step 3 specifically includes:

[0019] In order to clarify the spatiotemporal interaction relationship of vehicles and reduce the modeling complexity, the adversarial training module of the generator uses a model based on nodes and directed edges to construct the spatiotemporal relationship between vehicles in the scene. Each vehicle in the scene is regarded as a node h i . Directed edges represent interactions between vehicles. The relationship between nodes and directed edges is captured using a graph neural network and modeled in the generator. In each layer of the graph neural network, the node feature h i Updates are made based on the features of adjacent nodes and directed edges.

[0020] Furthermore, the discriminator of the vehicle trajectory generation model in step 4 specifically includes:

[0021] In the discriminator D, τ is judged real With τ fake The contrastive learning loss function was added earlier to distinguish the differences within the generated data to avoid mode collapse of the generator.

[0022] Furthermore, the trajectory completion module of the self-vehicle trajectory generation model is constructed in step 5, specifically including:

[0023] Generator G θ Based on the current track status Select next action Generate trajectories and obtain the maximum expected reward. The discriminator gives reward values ​​to the vehicle lane changing sequence combined with the Deep Deterministic Policy Gradient (DDPG) to update the generator.

[0024] Furthermore, step 6 constructs a trajectory generation module based on the lane-changing risk level of the vehicle, specifically including:

[0025] The initial y coordinates of the five vehicles in the scene are located at the center of the lane, and the x coordinate of the ego vehicle is set to 0. The initial velocity of the ego vehicle is obtained by randomly selecting values ​​within a certain interval, and the initial v of the five vehicles is set y =0,a y = 0. Then randomly select values ​​in different safety-critical TTCs and relative speeds or relative distances, and infer the corresponding relative speeds or relative distances between the self-vehicle and the other four vehicles. Thus, the initial state of the five vehicles is determined. In order to characterize the movement trend of the vehicle, the cost function f ttc,total The optimization of the initial state is performed to determine the feasible initial state within a certain time stamp. In order to accurately evaluate the comprehensive safety factor of the ego vehicle in the multi-vehicle background, the TTC constraint between the ego vehicle and the four background vehicles in the initial state time period [t0, t0+Δt] is considered. The initial state is optimized by minimizing the total cost function of the initial state variables (including the speed and acceleration of the ego vehicle, the distance difference and speed difference of the four background vehicles relative to the ego vehicle).

[0026] A second aspect of the present invention relates to a device for generating lane-changing trajectories for a self-driving vehicle including multiple vehicles, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the safety-critical trajectory generation method for lane changing of an autonomous driving vehicle of the present invention.

[0027] A third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the method for generating safety-critical trajectories for lane changing of an autonomous driving vehicle of the present invention.

[0028] The present invention has at least the following beneficial effects:

[0029] The present invention proposes a method for generating safety-critical lane-changing trajectories in simulation tests of autonomous vehicles. Based on the public highD dataset, DWA-HEO-TransGAN is proposed. According to the dangerous lane-changing level of the ego vehicle, the initial states of five vehicles including the ego vehicle are calculated, and a safety-critical lane-changing scenario is constructed. The safety-critical lane-changing scenario library is generated in batches using random sampling of the initial states. The batch-generated lane-changing trajectories of the ego vehicle have sufficient similarity with the real dataset in terms of the distance characteristics between multiple vehicles and the ego vehicle; under the TTC level divided by different relative speeds and relative distances between the background vehicle and the ego vehicle, the model accurately generates trajectories that meet the requirements; there are certain differences between the generated trajectories and the comparison with the real dataset. It shows that the method proposed in this paper can meet the construction requirements of the safety-critical trajectory scenario library for lane-changing of autonomous vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is the DWA-HEO-TransGAN flow chart of the present invention.

[0031] Figure 2 This is a structural diagram of the DWA-HEO-TransGAN model of the present invention.

[0032] Figure 3 and Figure 4 This is a comparison of the DWA-HEO-TransGAN model and the real dataset in the safety-critical trajectory range. Figure 3 The intervals are divided according to the relative speed. Figure 4 The intervals are divided according to relative distance.

[0033] Figure 5 Schematic diagram of the device of the present invention. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] Example 1

[0036] refer to Figure 1 , this embodiment provides a DWA-HEO-TransGAN generation method of a lane-changing trajectory of an ego vehicle including multiple vehicles, comprising the following steps:

[0037] S1: In step 1, the original data set already has the lane number information of the vehicle, so the vehicle whose lane number changes is the lane-changing vehicle. The state sequence of the vehicle is combined with the state sequence of the vehicles in the original lane before and after the vehicle, and the state sequence of the vehicles in the target lane before and after the vehicle to obtain a lane-changing trajectory sample. Finally, the lane-changing trajectory samples of the ego vehicle that include the vehicles in front and behind the dual lanes are selected. Sample data include: vehicle x speed, vehicle y speed, vehicle x acceleration, vehicle y acceleration, vehicle distance from lane change line, vehicle density in original lane, vehicle density in target lane, vehicle distance from vehicle in front of original lane, x speed difference between vehicle in front of original lane, y speed difference between vehicle in front of original lane, distance from vehicle in back of original lane, x speed difference between vehicle in back of original lane, y speed difference between vehicle in back of original lane, distance from vehicle in front of target lane, x speed difference between vehicle in front of target lane, y speed difference between vehicle in front of target lane, distance from vehicle in back of target lane, x speed difference between vehicle in front of target lane, y speed difference between vehicle in front of target lane, distance from vehicle in back of target lane, x speed difference between vehicle in back of target lane, y speed difference between vehicle in back of target lane, and lane change duration.

[0038] Specifically, a total of 4,300 lane-changing trajectory samples of the vehicle that meet the requirements were extracted from the highD dataset. The adversarial generation of multiple vehicle trajectories was then performed. The model structure diagram is shown in the figure below. Figure 2 shown.

[0039] S2: The dynamic window mechanism of the generator pre-training module in the ego-vehicle trajectory generation model in step 2 specifically includes:

[0040] Initialize dynamic window parameters: set lateral acceleration threshold To determine the lane change status. Set the threshold ω threshold To determine the start and end of the window. , the calculation method of the heading angular velocity ω(t) is:

[0041]

[0042] Update the window center and size in real time: set the initial window length σ0 and adjust the proportional constant k. According to the lateral acceleration a of the current time step y Determine whether the vehicle is changing lanes. If the vehicle is changing lanes, the window width is the width of two lanes, otherwise it is the width of a single lane. Calculate the angular velocity ω(t) of the current time step t, and update the window length based on the angular velocity:

[0043] σ=σ0+k·|ω(t)| (2)

[0044] Get the lateral coordinate y(t) and heading angular velocity ω(t) of the current time step t and update the window center t center :

[0045] t center =t+α·y(t)+β·ω(t) (3)

[0046] Calculate the attention weights within the dynamic window: For each time step t′, calculate the Gaussian weight function:

[0047]

[0048] Embed attention weights into generator pre-training: In the generator’s Transformer model, the attention weights w(t′) at each time step are embedded.

[0049] S3: In the generative model, the adversarial training module of the generator builds a multi-vehicle model based on nodes and directed edges, including:

[0050] In order to clarify the spatiotemporal interaction relationship of vehicles and reduce the modeling complexity, the adversarial training module of the generator uses a traffic graph based on nodes and directed edges to construct the spatiotemporal relationship between vehicles in the scene. Five vehicles are set in the scene, each of which is a node h i, where i can be 0, 1, 2, 3, 4, representing the ego vehicle, the vehicle in front of the original lane, the vehicle behind the original lane, the vehicle in front of the target lane, and the vehicle behind the target lane, respectively. The feature vector h of the node i Including the vehicle's position, speed and acceleration information:

[0051]

[0052] Among them, the x-axis is parallel to the vehicle's driving direction, and the y-axis is perpendicular to the vehicle's driving direction. i ,y i ) is the position coordinate of the i-th vehicle, are the velocity components on the x and y axes respectively, are the acceleration components on the x and y axes respectively.

[0053] Directed edges represent interactions between vehicles. For each pair of vehicles (v i ,v j ), directed edge e ij Indicates that from v i to v j The relationship between the eigenvector e ij Including information about relative position and relative speed. The directed edge feature is expressed as:

[0054]

[0055] A graph neural network (GNN) is used to capture the relationship between nodes and directed edges and model them in the generator. In each layer of the graph neural network, the node feature h i It can be updated based on the characteristics of adjacent nodes and directed edges:

[0056]

[0057] in, represents the node features of the (k)th layer, N(i) represents the node v i The set of adjacent nodes, W v and W e is a learnable weight matrix, b is a bias term, and μ is a nonlinear activation function. The update formula for directed edge features is:

[0058]

[0059] in, represents the edge feature of the (k)th layer, which transforms the node v i and v j Features For splicing, W ve is the learnable weight matrix, b e is the bias term, is a non-linear activation function.

[0060] S4: Generate a comparison module for the generative model discriminator, including:

[0061] In the discriminator D, τ is judged real With τ fake The contrastive learning loss function added earlier is used to distinguish the differences within the generated data to avoid mode collapse of the generator:

[0062]

[0063] Here, m is the margin for contrastive learning.

[0064] S5: Construct the trajectory completion module of the ego-vehicle trajectory generation model, which includes:

[0065] Generator G θ The goal is to be based on the current trajectory state Select next action Generate trajectories and obtain the maximum expected reward. The discriminator gives a reward value for the lane-changing sequence of vehicles combined with the Deep Deterministic Policy Gradient (DDPG) network to update the generator. The expected reward formula can be expressed as:

[0066]

[0067] in represents the DDPG experience replay buffer, R N is the reward of the discriminator for a trajectory, θ is the strategy of the generator, Indicates that from state z n-1 Start by following the generation strategy G θ Actions taken Rewards received.

[0068] S6: Construct a trajectory generation module based on the lane-changing danger level of the ego vehicle, generate trajectories in batches, and compare the real trajectories of each safety criticality level with the key sample points in the generated trajectory. Figure 3 , Figure 4 As shown, specifically including:

[0069] The initial y coordinates of the five vehicles in the scene are located at the center of the lane, and the x coordinate of the ego vehicle is set to 0. The initial velocity of the ego vehicle is obtained by randomly selecting values ​​within a certain interval, and the initial v of the five vehicles is set y =0,a y= 0. Then, randomly select values ​​within different safety-critical TTCs and relative speeds or relative distances, and inversely deduce the corresponding relative speeds or relative distances between the self-vehicle and the other four vehicles. Thus, the initial states of the five vehicles are determined.

[0070]

[0071] In order to describe the movement trend of the vehicle, the cost function f ttc,total To accurately evaluate the comprehensive safety factor of the ego vehicle in the multi-vehicle background, the TTC constraint between the ego vehicle and the four background vehicles in the initial state time period [t0, t0+Δt] is considered, and its cost function is shown as follows:

[0072]

[0073] The initial state is optimized by minimizing the total cost function through the initial state variables (including the speed and acceleration of the ego vehicle, the distance difference and speed difference of the four background vehicles relative to the ego vehicle):

[0074]

[0075] Example 2

[0076] like Figure 5 This embodiment provides a device for generating lane-changing trajectories of a self-vehicle including multiple vehicles, including a memory and one or more processors, wherein the memory stores executable codes, and when the one or more processors execute the executable codes, they are used to implement a method for generating lane-changing trajectories of a self-vehicle including multiple vehicles in embodiment 1.

[0077] Example 3

[0078] The present embodiment provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, a method for generating lane-changing trajectories of a vehicle including multiple vehicles is implemented in the present embodiment.

[0079] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A safety-critical trajectory generation method for lane changing of an autonomous vehicle, including: Step 1: Obtain the highD data set, obtain the original data set of the self-driving vehicle, and perform data preprocessing on the original data set of the self-driving vehicle, including removing few-frame trajectories, wavelet decomposition filtering, lane change behavior annotation, and multi-vehicle trajectory extraction on the original data set of the self-driving vehicle; Step 2: Construct a pre-training module of the generator in the ego-vehicle trajectory generation model; wherein the pre-training module of the generator in the ego-vehicle trajectory generation model adopts the Transformer model and designs a dynamic window mechanism: Step 3: construct an adversarial training module of the generator in the ego-vehicle trajectory generation model; wherein the adversarial training module of the ego-vehicle trajectory generation model represents the interaction relationship between the ego-vehicle and the background vehicles based on nodes and directed edges; Step 4: Construct a discriminator for the ego-vehicle trajectory generation model; the discriminator includes a generation comparison module to deal with mode collapse; Step 5: construct a trajectory completion module of the ego-vehicle trajectory generation model; wherein the trajectory completion module of the ego-vehicle trajectory generation model adopts a DDPG network; Step six: construct a trajectory generation module based on the lane-changing risk level of the ego vehicle; wherein the trajectory generation module based on the lane-changing risk level of the ego vehicle optimizes the initial state of the ego vehicle in the multi-vehicle background through a cost function.

2. The safety-critical trajectory generation method for lane changing of an autonomous driving vehicle according to claim 1, characterized in that: The data of the original data set of the vehicle driving in step 1 includes the driving behavior characteristics of the vehicle and the relative driving states of the vehicle and the front and rear vehicles in the original lane and the target lane of the vehicle changing lane. The data preprocessing of the original data set of the self-driving vehicle includes removing the few-frame trajectory, wavelet decomposition filtering, lane change behavior annotation and multi-vehicle trajectory extraction of the original data set of the self-driving vehicle: The few-frame trajectory deletes the trajectories below the preset frame number threshold and retains the trajectory data with sufficient time length; the wavelet decomposition filter performs multi-layer wavelet decomposition on the trajectory data to obtain approximate coefficients and detail coefficients of different scales, and performs inverse wavelet transform according to the selected coefficients to obtain the smoothed trajectory data; the lane change behavior is labeled based on the ID change of the lane where the ego vehicle is located; the multi-vehicle trajectory extracts the trajectories of the ego vehicle and the surrounding background as a group of driving trajectories.

3. The safety-critical trajectory generation method for lane changing of an autonomous driving vehicle according to claim 1, characterized in that: The dynamic window mechanism of the generator pre-training module in the vehicle trajectory generation model described in step 2 specifically includes: Determine the lane change state and the start and end of the window according to the set lateral acceleration threshold and heading angular velocity threshold; update the window center and size in real time: set the initial window length σ0 and adjust the proportional constant k; according to the lateral acceleration a of the current time step y Determine whether the lane is changing; calculate the angular velocity ω(t) of the current time step t, and update the window length according to the angular velocity; obtain the lateral coordinate y(t) and angular velocity ω(t) of the current time step t to update the window center t center ; Calculate the attention weights within the dynamic window: For each time step t′, calculate the Gaussian weight function; Embed the attention weights into the generator pre-training: In the generator's Transformer model, use a separate attention head to embed the weights w(t′).

4. The safety-critical trajectory generation method for lane changing of an autonomous driving vehicle according to claim 1, characterized in that: The adversarial training module of the generator in the vehicle trajectory generation model described in step 3 specifically includes: In order to clarify the spatiotemporal interaction relationship of vehicles and reduce the modeling complexity, the adversarial training module of the generator uses a model based on nodes and directed edges to construct the spatiotemporal relationship between vehicles in the scene; each vehicle in the scene is regarded as a node h i ; Directed edges represent interactions between vehicles; Use graph neural networks to capture the relationship between nodes and directed edges and model them in the generator; In each layer of the graph neural network, the node feature h i Updates are made based on the features of adjacent nodes and directed edges.

5. The safety-critical trajectory generation method for lane changing of an autonomous driving vehicle according to claim 1, characterized in that: The discriminator of the vehicle trajectory generation model described in step 4 specifically includes: In the discriminator D, τ real With τ fake The contrastive learning loss function was added earlier to distinguish the differences within the generated data to avoid mode collapse of the generator.

6. The safety-critical trajectory generation method for lane changing of an autonomous driving vehicle according to claim 1, characterized in that: Step 5 constructs a trajectory completion module for the ego-vehicle trajectory generation model, specifically including: Generator G θ Based on the current track status Select next action Generate a trajectory and obtain the maximum expected reward. The discriminator gives a reward value for the vehicle lane changing sequence combined with the deep deterministic policy network DDPG to update the generator.

7. The method for generating a safety-critical trajectory for lane changing of an autonomous driving vehicle according to claim 1, wherein: Step 6 constructs a trajectory generation module based on the lane-changing risk level of the ego vehicle, specifically including: The initial y coordinates of the five vehicles in the scene are located at the center line of the lane, and the x coordinate of the ego vehicle is set to 0. The initial velocity of the ego vehicle is obtained by randomly selecting values ​​within a certain interval, and the initial v of the five vehicles is set. y =0,a y = 0; then randomly select values ​​within different safety-critical TTCs and relative speeds or relative distances, and infer the corresponding relative speeds or relative distances between the self-vehicle and the other four vehicles; thus determining the initial state of the five vehicles; in order to characterize the movement trend of the vehicle, the cost function f ttc,total The feasible initial state is determined within the timestamp by optimizing the ego vehicle. To accurately evaluate the comprehensive safety factor of the ego vehicle in the multi-vehicle background, the TTC constraint between the ego vehicle and the four background vehicles in the initial state time period [t0, t0+Δt] is considered. The initial state is optimized by minimizing the total cost function by the initial state variables. The initial state variables include the speed and acceleration of the ego vehicle, and the distance difference and speed difference of the four background vehicles relative to the ego vehicle.

8. A lane-changing trajectory generating device for a vehicle including multiple vehicles, characterized in that: It includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the safety-critical trajectory generation method for lane changing of an autonomous driving vehicle according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the method for generating a safety-critical trajectory for lane changing of an autonomous driving vehicle as described in any one of claims 1-7 is implemented.

Images