Automatic driving risk scene generation method and system suitable for strong interaction environment, and storage medium

Abstract

The invention relates to the technical field of automatic driving, in particular to an automatic driving risk scene generation method and system suitable for a strong interaction environment and a storage medium, and the method comprises the steps: firstly collecting the data of a real vehicle following or lane changing target scene, and then calculating the vehicle motion interaction feature F of the target scene; and a TimeGAN model is improved in combination with an attention mechanism to obtain an F-TimeGAN model, a target scene is input into the F-TimeGAN model for scene generation, and finally, a strong interaction risk scene is screened out from the generated scene for automatic driving test. The strong interaction risk scene data which is difficult to collect in reality can be efficiently generated based on the conventional target scene data, the interactivity and the risk of the generated scene are enhanced, the problem of rare key test scene data is effectively solved, and scene data support is provided for safety verification of automatic driving in a strong interaction environment.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a method, system, and storage medium for generating autonomous driving driving risk scenarios suitable for a highly interactive environment. Background Art

[0002] Existing methods for generating vehicle motion interaction scenarios mainly rely on two technical paths: deep learning and reinforcement learning. However, they generally have problems such as insufficient interaction feature modeling and limited scenario authenticity.

[0003] Existing scenario generation methods use static trajectory data collected offline to generate scenarios, which is unable to capture the dynamic temporal relationships of vehicle motion interactions. Second, existing technologies often assume that all vehicles have the same dynamic characteristics (such as braking performance and response delay), ignoring the significant differences between heavy trucks and cars in key parameters such as vehicle type and braking performance. This leads to deviations between the generated scenarios and actual traffic flow. Furthermore, for highly complex interactive scenarios such as lane changes that involve dynamic adjustment of heading angles, existing evaluation methods for vehicle motion interaction characteristics are simplistic and lack consideration of temporal changes in heading angles, making it difficult to quantify complex interactive behaviors. However, autonomous driving testing requires highly interactive and dangerous scenario data, but these critical scenarios are difficult to collect in reality.

[0004] In summary, there is an urgent need for a method, system and storage medium for generating risk scenarios for autonomous driving in a highly interactive environment to solve the problems existing in the existing technology. Summary of the Invention

[0005] The present invention aims to provide a method for generating risk scenarios for autonomous driving in highly interactive environments, aiming to address the problems of insufficient interactive feature modeling and limited scenario authenticity in existing vehicle motion interaction scenario generation methods. The specific technical solution is as follows: A method for generating risk scenarios for autonomous driving in a highly interactive environment, comprising: S1. Collect road vehicle trajectory data to build a basic database; S2. Extract the target scene from the basic database and calculate the vehicle motion interaction characteristics of the target scene ; S3, based on vehicle motion interaction features Build the F-TimeGAN model; S4. Set candidate hyperparameters of the F-TimeGAN model and input the extracted target scene into the F-TimeGAN model under different candidate hyperparameters to generate the scene; S5. Evaluate the similarity between the scene data generated by the F-TimeGAN model and the target scene data under different candidate hyperparameters, and select the one with the highest similarity. F-TimeGAN models are used as alternative models, among which is a natural number greater than or equal to 1; S6. Perform a risk assessment on the scenario library generated by the candidate models to obtain the number of risk scenarios for each candidate model; if there is a candidate model with a greater number of risk scenarios than the target number of scenarios, the F-TimeGAN model that can generate the most risk scenarios is selected as the optimal model; if there is no such model, return to step S4 to reset the candidate hyperparameters of the F-TimeGAN model; S7. Filter out high-interaction risk scenarios from the scenario data generated by the optimal model, and output the screening results to the autonomous driving test scenario library.

[0006] Preferably, the vehicle motion interaction feature Expressed as: in, is the vector connecting the center points of the two vehicles’ heads, The direction of the following vehicle and The angle of The direction of the vehicle ahead and The angle of express The speed of the car after time, express The speed of the car ahead at that moment, for Speed ​​of the vehicle behind In vector The weight on for Speed ​​of the preceding vehicle at the moment In vector The weight on Represents the static safety distance parameter, For safe headway, Indicates the comfortable deceleration of the following vehicle. is the maximum deceleration of the preceding vehicle, is the maximum deceleration of the following vehicle, To balance the distance, when two vehicles are in adjacent lanes, , when two vehicles are in the same lane , Indicates the actual distance between the two workshops, is the speed difference sensitivity coefficient, is the expected speed of the following vehicle, Indicates the reaction time of the following vehicle. is the depth of the potential well, is the dimensionless Morse potential parameter, is the base of the natural logarithm function.

[0007] Preferably, based on the vehicle motion interaction characteristics Determine the vehicle motion interaction state of the scene, specifically: in, express The vehicle is in a strong interactive state at all times. express The vehicle is in a non-strong interaction state at all times. for The vehicle motion interaction characteristics at each moment, is the threshold of strong interaction state and Less than 0, is the degree of interaction exposure, Indicates the total time in the strong interaction state in the scene, Indicates the total time of the scene, Indicates that the current scene is a strong interaction scene. Indicates that the current scene is a non-strong interaction scene.

[0008] Preferably, the generator is introduced based on the vehicle motion interaction feature The attention mechanism is used to build the F-TimeGAN model, specifically: in, Indicates the calculation of attention weights, represents the hidden state of the trajectory mapping of the main vehicle and other vehicles in the original time series data, is the query vector in the attention mechanism module, is the key vector in the attention mechanism module, is the value vector in the attention mechanism module, Query vector in the attention mechanism module The linear mapping parameters of is the key vector in the attention mechanism module The linear mapping parameters of is the median vector of the attention mechanism module The linear mapping parameters of Represents the key vector Dimensions, is the key vector The transposed matrix of represents scaled dot product attention; is the sensitivity adjustment parameter; The weighted hidden state, Vehicle motion interaction features The tanh function value of The generator of the F-TimeGAN model is expressed as: in, represents the static feature latent code, A generative network representing static features, represents a vector space that defines a known distribution, represents the current time feature potential code, represents the generative network for temporal features, represents a random vector, Represents the potential code of the previous time feature.

[0009] Preferably, the target scene data set is input into the F-TimeGAN model Expressed as: in, is the total number of samples, Contains the trajectory information of the main vehicle Ego and other vehicles OV and the vehicle motion interaction characteristics between the two , represents the trajectory of the main vehicle, represents the trajectory of other vehicles, express The host vehicle's position at the moment, express The positions of other vehicles at the moment, express The vehicle motion interaction characteristics between the host vehicle and other vehicles at the moment, set Each sample in has the same time step , ; The scene data set generated by the F-TimeGAN model is represented as: in, The total number of samples generated for the F-TimeGAN model, The first samples, The main vehicle trajectory generated by the F-TimeGAN model, Generated for the F-TimeGAN model The host vehicle's position at the moment, Other vehicle trajectories generated by the F-TimeGAN model, Generated for the F-TimeGAN model m The position of other vehicles at the moment, the collection Each generated sample has the same time step , .

[0010] Preferably, the method for confirming the alternative model in step S5 is: Filter out generated data that meets < and and The F-TimeGAN model is then used to calculate the comprehensive score of each selected F-TimeGAN model. , the selected F-TimeGAN models are ranked according to their comprehensive scores Sort from low to high, select the one at the front An F-TimeGAN model was used as an alternative model; Among them, the comprehensive score Expressed as: in, is the FID score, is the prediction score, is the discriminant score, is the JS divergence value, 、 and All are set thresholds.

[0011] Preferably, in step S6, risk assessment is performed on the scenario library generated by the candidate model, specifically: The driving risk of the scene is assessed using the stopping sight distance SSD and the stopping distance index SDI. SSD and SDI are expressed as: in, is the vehicle braking distance, Indicates the driver's reaction distance, Indicates vehicle exist The speed of time, Indicates the driver's reaction time, is the deceleration; in, express The distance between the front of the vehicle in front and the front of the vehicle behind at the moment, Indicates that the vehicle ahead is Stopping sight distance at all times, Indicates that the following vehicle is Stopping sight distance at all times, Indicates the vehicle length of the preceding vehicle; Among them, when 1 means Time Vehicle At risk, 1 means that the scenario is considered a risk scenario. 0 means Time Vehicle In a safe state.

[0012] Preferably, by collision risk factor Determine the risk level of the scene and the collision risk coefficient Greater than threshold The scene is considered to be a high-risk scene; the collision risk coefficient Expressed as: in, is the degree of risk exposure; Indicates the total time in the risk state in the scenario; Indicates the total driving time of the scene; is the severity of the risk; is the observed maximum stopping distance index; is the theoretical maximum stopping distance index.

[0013] The present invention also provides an autonomous driving driving risk scenario generation system suitable for a highly interactive environment, the system comprising a processor and a memory, wherein the memory stores a computer program, and the processor executes the described method when running the computer program.

[0014] The present invention also provides a storage medium, wherein a computer program is stored in the storage medium, and the method described is executed when the computer program is executed.

[0015] The application of the technical solution of the present invention has the following beneficial effects: The generation method of the present invention first collects data of target scenarios such as real car-following or lane changing, then uses the vehicle motion interaction feature F of the target scenario in combination with the attention mechanism to improve the TimeGAN model to obtain the F-TimeGAN model, inputs the target scenario into the F-TimeGAN model for scenario generation, and finally screens out the strong interaction risk scenarios from the generated scenarios for autonomous driving testing. The method of the present invention effectively improves the interactivity and risk of the generated scenarios. The method of the present invention can efficiently generate strong interaction risk scenario data that is difficult to collect in reality based on conventional target scenario data, effectively solving the problem of scarce key test scenario data, and providing scenario data support for the safety verification of autonomous driving in a strong interaction environment.

[0016] The method of the present invention solves the problem that existing methods rely solely on basic parameters such as vehicle spacing, relative speed, and acceleration changes to describe the interaction between the main vehicle and surrounding vehicles, while ignoring the impact of key factors such as dynamic adjustment of heading angle and vehicle model dynamics differences on the intensity of interaction. In addition, existing methods mainly generate trajectories or motion parameters statically, lacking consideration of the vehicle motion interaction process in the scene. The method of the present invention defines a vehicle motion interaction intensity evaluation index by considering vehicle speed, heading angle, and vehicle model differences, and further integrates vehicle motion interaction features with an attention mechanism to improve the generation model. The method of the present invention is more suitable for strong interaction scenarios, improving the interactivity, authenticity, and danger of the scene.

[0017] In addition to the above-described objects, features and advantages, the present invention has other objects, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings: Figure 1is a flow chart of the method for generating risk scenarios for autonomous driving according to the present invention; Figure 2 It is a schematic diagram of the balanced distance between vehicles in adjacent lanes and the same lane; Figure 3 This is a PCA visualization diagram of the data generated by the F-TimeGAN model; Figure 4 This is a t-SNE visualization of the data generated by the F-TimeGAN model; Figure 5 It is a schematic diagram of the collision risk coefficient distribution of the entire sample; Figure 6 yes Figure 5 Schematic diagram of the sample distribution with a medium collision risk coefficient greater than or equal to 0.1; Figure 7 It is a schematic diagram of the distribution of interaction exposure of samples. DETAILED DESCRIPTION

[0019] To facilitate understanding of the present invention, the present invention will be described more fully below, along with preferred embodiments thereof. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the disclosure of the present invention.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0021] Example 1: See also Figure 1 This embodiment provides a method for generating risk scenarios for autonomous driving in a highly interactive environment, as follows: S1. Collect road vehicle trajectory data to build a basic database; Preferably, in this embodiment, vehicle trajectory data is obtained based on drone aerial photography, vehicle-mounted sensors (such as GPS / IMU integrated navigation system, visual camera) and real vehicle experimental platform. The fields for constructing the basic database must include the following core fields: timestamp , the two-dimensional plane position of the vehicle trajectory , Driver , vehicle width and the lane the vehicle is in .

[0022] Of course, in some embodiments, other methods may be used to collect vehicle trajectory data, and this embodiment does not list all possible methods.

[0023] S2. Extract the target scene from the basic database and calculate the vehicle motion interaction characteristics of the target scene ; Preferably, in this embodiment, the data in the basic database is first cleaned of noise, and then the target scene is extracted by analyzing the vehicle motion characteristics. In this embodiment, noise cleaning is performed by removing abnormal trajectory points through a statistical filtering algorithm based on a sliding window, and spatial offset is eliminated through coordinate transformation, thereby ensuring data continuity and reliability.

[0024] Furthermore, the vehicle motion characteristics specifically include: static attributes, dynamic parameters, the relative distance between the main vehicle and other vehicles, and the relative speed between the main vehicle and other vehicles. Among them, the static attributes include vehicle type, length and width, and the dynamic parameters include vehicle speed, vehicle acceleration and vehicle heading angle (i.e., vehicle driving direction).

[0025] Furthermore, in this embodiment, the target scene includes a car-following scene and / or a lane-changing scene. Specifically, the target scene is identified through the spatial and temporal interaction between the host vehicle trajectory and the trajectories of other vehicles.

[0026] In this embodiment, lane-changing scenario samples must meet the following requirements: a) the longitudinal relative distance between the following vehicle and the lane-changing vehicle must not exceed 120 meters; b) in the extracted samples, only the lane-changing vehicle must change lanes, while the following vehicle that is cut in must be traveling in its own lane; and c) the lane-changing time step must be greater than 3 seconds.

[0027] In this embodiment, the following scenario samples must meet the following requirements: a) the two vehicles in the same lane must be identified, and neither vehicle changes lanes during the following process; b) the distance between the following vehicles must be less than 125 meters to ensure that they are not in a free-flow state; c) the duration of the following process must be greater than 10 seconds.

[0028] Of course, in some embodiments, the requirements for the car-following scenario and the lane-changing scenario may be different. Those skilled in the art can flexibly adjust and modify the judgment requirements for the car-following scenario and the lane-changing scenario.

[0029] The following will describe the vehicle motion interaction features in this embodiment. For detailed explanation: Morse potential function about The interaction force can be obtained by differentiation , the specific calculation is as follows: in, represents the distance between atoms, is the depth of the potential well, is the dimensionless Morse potential parameter, represents the equilibrium distance, is the base of the natural logarithm function.

[0030] See also Figure 2 In this embodiment, the vehicle heading angle, speed and vehicle model differences are taken into consideration, the Morse potential function is modified, and a vehicle motion interaction feature suitable for strong interaction scenarios is constructed. , where when the vehicle and the interacting vehicle are in adjacent lanes, the safe distance model of the Gipps model is used to take the minimum safe lane change distance as the expected distance When the vehicle and the interacting vehicle are in the same lane, the safe distance is calculated according to the IDM model. .when 0 means there is repulsion between the two vehicles. 0 indicates that there is attraction between the two vehicles, and the vehicle motion interaction characteristics Specifically expressed as: in, is the vector connecting the center points of the two vehicles’ heads, The direction of the following vehicle and The angle of The direction of the vehicle ahead and The angle of express The speed of the car after time, express The speed of the car ahead at that moment, for Speed ​​of the vehicle behind In vector The weight on for Speed ​​of the preceding vehicle at the moment In vector The weight on Represents the static safety distance parameter, For safe headway, Indicates the comfortable deceleration of the following vehicle. is the maximum deceleration of the preceding vehicle, is the maximum deceleration of the following vehicle, To balance the distance, when two vehicles are in adjacent lanes, the Gipps model is used to obtain When two vehicles are in the same lane, the IDM model is used to obtain , Indicates the actual distance between the two workshops, is the speed difference sensitivity coefficient, is the expected speed of the following vehicle, Indicates the reaction time of the following vehicle. is the depth of the potential well, is the dimensionless Morse potential parameter, is the base of the natural logarithm function.

[0031] Furthermore, based on the vehicle motion interaction characteristics You can determine whether the scene is a strong interaction scene, specifically: In this embodiment, the high interaction state (HI) and the interaction exposure level (IEL) are defined to judge the interaction state of the scene. The specific calculation is as follows: in, express Whether the vehicle is in a strong interactive state at the moment, Indicates that the vehicle is in a strong interactive state. Indicates that the vehicle is in a non-strong interaction state. for The vehicle motion interaction characteristics at each moment, is the threshold of strong interaction state and Less than 0, drawn using kernel density estimation method The distribution curve is then used to perform peak detection to determine the threshold; The value range is from 0 to 1. Indicates the total time in the strong interaction state in the scene, Indicates the total time of the scene. hour, It means that the vehicle is in a high interaction state for half of the driving time, and this scenario is considered a high interaction scenario (HIS).

[0032] So far, the vehicle trajectory information and vehicle motion interaction characteristics The target scene data set It can be expressed as: in, is the total number of samples, Contains the trajectory information of the main vehicle Ego and other vehicles OV and the vehicle motion interaction characteristics between the two , represents the trajectory of the main vehicle, represents the trajectory of other vehicles, express The host vehicle's position at the moment, express The positions of other vehicles at the moment, express The vehicle motion interaction characteristics between the main vehicle and other vehicles at the same time, each sample has the same time step , .

[0033] S3, based on vehicle motion interaction features Build the F-TimeGAN model; The F-TimeGAN model in this embodiment has a similar network structure to the TimeGAN model, and mainly includes the following modules: encoder, restorer, supervisor, generator, and discriminator.

[0034] This embodiment introduces the vehicle motion interaction feature into the generator based on the TimeGAN model (i.e., time series generative adversarial network). The attention mechanism of the generator is improved. In order to reduce the long tail interaction force value The difference in numerical scale improves the stability of model training and uses the tanh function to analyze the vehicle motion interaction characteristics. Perform nonlinear mapping and compress it into Between. When The closer it is to 1, the stronger the attraction interaction is. The closer to -1, the stronger the repulsive interaction. This method enhances the vehicle's own trajectory dependency in time series modeling. The attention weight is The value is linearly biased, making the generator more concerned with Larger time steps to exploit vehicle motion interaction features Effectively improve the spatial interaction modeling capabilities between vehicles. The specific mathematical expression is as follows: in, Indicates the calculation of attention weights, represents the hidden state of the trajectory mapping of the main vehicle and other vehicles in the original time series data, is the query vector in the attention mechanism module, is the key vector in the attention mechanism module, is the value vector in the attention mechanism module, Query vector in the attention mechanism module The linear mapping parameters of is the key vector in the attention mechanism module The linear mapping parameters of is the median vector of the attention mechanism module The linear mapping parameters of Represents the key vector Dimensions, is the key vector The transposed matrix of represents scaled dot product attention; To adjust the sensitivity parameter, adjust the weight of the interaction feature F; The weighted hidden state is finally Mapped to the output of the generator through a fully connected layer.

[0035] The original generator of the TimeGAN model can be expressed as 、 , the generator of the F-TimeGAN model after the improvement of the attention mechanism in this embodiment can be expressed as: in, represents the static feature latent code, A generative network representing static features, represents a vector space that defines a known distribution, represents the current time feature potential code, represents the generative network for temporal features, represents a random vector, Represents the potential code of the previous time feature.

[0036] S4. Set candidate hyperparameters of the F-TimeGAN model and input the extracted target scene into the F-TimeGAN model under different candidate hyperparameters to generate the scene; Specifically, those skilled in the art can set multiple sets of candidate hyperparameters according to actual conditions and collect the extracted target scene data. The input is then fed into the F-TimeGAN model under different candidate hyperparameters for scene generation, thereby achieving the goal of selecting the optimal hyperparameter F-TimeGAN model.

[0037] The setting of candidate hyperparameters is common knowledge in the art, so step S4 is not described in detail in this implementation.

[0038] Specifically, the scene data set generated by the F-TimeGAN model Expressed as: in, The total number of samples generated for the F-TimeGAN model, The first samples, The main vehicle trajectory generated by the F-TimeGAN model, Generated for the F-TimeGAN model The host vehicle's position at the moment, Other vehicle position trajectories generated by the F-TimeGAN model, Generated for the F-TimeGAN model m The positions of other vehicles at the moment, each generated sample has the same time step , .

[0039] It should be noted that As the output data of the F-TimeGAN model, it only includes the trajectories of the main vehicle and other vehicles, and does not include the vehicle motion interaction features between the two vehicles.

[0040] S5. Evaluate the similarity between the scene data generated by the F-TimeGAN model and the target scene data under different candidate hyperparameters, and select the one with the highest similarity. F-TimeGAN models are used as alternative models, among which is a natural number greater than or equal to 1; In this embodiment, select Four metrics, namely, Fréchet Inception Distance (FID), Prediction Score (PS), Discriminative Score (DS), and Jensen–Shannon divergence (JS), evaluate the quality of data generated by each F-TimeGAN model from multiple perspectives (i.e., judging the similarity between the generated data and the real data). Specifically, First, filter out the generated data that meets < and and The F-TimeGAN model is then used to calculate the comprehensive score of each selected F-TimeGAN model. , the selected F-TimeGAN models are ranked according to their comprehensive scores Sort from low to high, select the one at the front An F-TimeGAN model is used as an alternative model.

[0041] Among them, the comprehensive score Used to express the similarity between the generated scene data and the target scene data, The lower the score, the higher the similarity between the generated scene data and the target scene data (ie, real data). Expressed as: (twenty four), in, is the FID score, is the prediction score, is the discriminant score, is the JS divergence value. The above four indicators are used to compare and generate data. and real data similarity, and generate data Timing is assessed; 、 and All are set thresholds. In this embodiment, 、 、 .

[0042] S6. Perform a risk assessment on the scenario library generated by the candidate models to obtain the number of risk scenarios for each candidate model; if there is a candidate model with a greater number of risk scenarios than the target number of scenarios, the F-TimeGAN model that can generate the most risk scenarios is selected as the optimal model; if there is no such model, return to step S4 to reset the candidate hyperparameters of the F-TimeGAN model; Specifically, in this embodiment, risk assessment is performed on the scenario library generated by the candidate model to achieve the purpose of identifying risk scenarios. The risk assessment method is as follows: The driving risk assessment of the scene is performed using the Stopping Sight Distance (SSD) (unit: m) and the Stopping Distance Index (SDI) (unit: m). The specific calculation methods of SSD and SDI are as follows: in, The braking distance of the vehicle (unit: m), that is, the distance from the driver taking braking measures to the vehicle coming to a complete stop. Indicates the driver's reaction distance (unit: m), Indicates vehicle exist Speed ​​at the moment (unit: km / h), vehicle Refers to the main vehicle or other vehicles (i.e. the leading vehicle or the following vehicle); Indicates the driver's reaction time (unit: s), usually 2.5 seconds. is the deceleration (unit: ), the deceleration of the truck is 2.4 , the car's deceleration is 3.4 .

[0043] in, express The distance between the front of the vehicle and the front of the vehicle behind at the moment (unit: m), Indicates that the vehicle ahead is Stopping sight distance at the moment (unit: m), Indicates that the following vehicle is Stopping sight distance at the moment (unit: m), Indicates the vehicle length of the preceding vehicle (unit: m), i.e. It represents the distance from the rear of the leading vehicle to the front of the trailing vehicle (unit: m); when 0 means that when the leading vehicle stops suddenly, the following vehicle cannot brake safely, the stopping sight distance is insufficient, and there is danger. Otherwise, it is safe: in, for Time Vehicle Is it in danger? 1 represents vehicle In a risky state, that is, the scenario is a risky scenario. 0 means vehicle be in a safe state; The risk scenarios in the scenario library generated by the alternative model can be screened out by using formula (25)-formula (27). This embodiment further uses the collision risk coefficient The risk coefficient comprehensively reflects the risk exposure time and risk severity in the risk scenario to determine the risk level of the risk scenario: in, is the risk exposure level, which ranges from 0 to 1; Indicates the total time of unsafe state in the scene (i.e. total time); Indicates the total driving time of the scene; is the severity of the risk, which ranges from 0 to 1; is the observed maximum stopping distance index; is the theoretical maximum stopping distance index, The value is 565m.

[0044] In this embodiment, the collision risk coefficient Greater than threshold As a high-risk scenario, the threshold The general value is 0.3; of course, those skilled in the art can adjust the threshold The value of can be adjusted flexibly.

[0045] Specifically, when confirming the optimal model in step S6, the F-TimeGAN model that generates the most risk scenarios can be selected as the optimal model only when the number of risk scenarios of the alternative model is greater than the target number of scenarios. This is because only when the number of risk scenarios contained in the data generated by the alternative model is greater than the target number of scenarios, can the scenario generation of the alternative model be considered to be of practical significance.

[0046] Specifically, the scene library generated by the optimal model is recorded as : in, The total number of samples generated for the optimal model, The optimal model generated samples.

[0047] S7. Filter out high-interaction risk scenarios from the scenario data generated by the optimal model, and output the screening results to the autonomous driving test scenario library.

[0048] Furthermore, according to formula (25)-formula (27), the risk scenarios in the optimal model generation scenario library can be screened out, and according to formula (3)-formula (10), the interaction state of the optimal model generation risk scenario can be determined. Specifically, if the risk index , it is marked as a risk scenario; if the strong interaction scenario indicator , it is marked as a strong interaction scene; if a scene and , it is marked as a strong interaction risk scenario. Filter out the scenario data generated by the optimal model to meet the and The scenarios are regarded as strong interaction risk scenarios, and all the screened strong interaction risk scenarios are output to the autonomous driving test scenario library for autonomous driving testing.

[0049] The generation method of this embodiment first collects data from real target scenarios such as car-following and lane changing. It then uses the target scenario's vehicle motion interaction feature F, combined with the attention mechanism, to improve the TimeGAN model to generate the F-TimeGAN model. The target scenario is then input into the F-TimeGAN model for scenario generation. Finally, strong interaction risk scenarios are screened from the generated scenarios. This method effectively improves the interactivity and risk of the generated scenarios. By generating strong interaction risk scenario data based on conventional target scenario data, it addresses the practical difficulty of collecting strong interaction risk scenario data and provides scenario data support for the safety verification of autonomous driving in highly interactive environments.

[0050] Example 2: This example obtains a public dataset, taking the HighD highway open dataset as an example. The data is collected from a two-way six-lane scene (files 25-60), and the method in Example 1 is used to generate the scene. The basic information of the HighD highway open dataset's two-way six-lane scene data is shown in Table 1: Table 1 Basic database information In this example, the lane-changing scene is used as the target scene. The lane-changing scene extraction rule is to extract 49 frames forward and 50 frames backward, centered at the lane change point, to form a 100-frame lane-changing sequence. In addition, the following conditions must be met: a) The longitudinal relative distance between the following vehicle and the lane-changing vehicle shall not exceed 120m; b) Among the extracted samples, only the lane-changing vehicle changes lanes, and the following vehicle that is cut in is driving in its own lane; c) The lane-changing time step needs to be unified, with each sample being 100 frames, i.e. =100.

[0051] A statistical analysis of the lane-changing scenarios in the basic database was conducted, and the interactions of different vehicle types were divided into car-to-car (C-cut-in-C), car-to-truck (C-cut-in-T), truck-to-car (T-cut-in-C), and truck-to-truck (T-cut-in-T). The driving risk of the scenarios was assessed using the Stopping Sight Distance (SSD) (unit: m) and the Stopping Distance Index (SDI) (unit: m). =1, the scenario is a risk scenario, and the statistical results are shown in Table 2: Table 2 Statistics of risk sample size for different scenarios Although scenarios involving truck lane changes, such as T-cut in-C and T-cut in-T, have a small sample size in natural driving data, they have a high risk ratio, indicating that truck lane changes are more likely to cause potential danger. Table 2 shows that T-cut in-C and T-cut in-T risk scenarios account for 80.69% and 47.90%, respectively, reflecting the greater risk potential of trucks when changing lanes. Among them, the C-cut in-T scenario has the second-largest sample size and the lowest risk ratio. This may be because, under natural driving conditions, when a car cuts in front of a truck, the driver consciously maintains a large longitudinal safety distance from the truck, thereby reducing the risk level. Driving experience shows a certain correlation between vehicle motion interaction and driving risk. Therefore, it is inferred that the interaction of a car cutting into a large truck is also weak. Clearly, it is difficult to capture highly interactive and dangerous C-cut in-T scenarios on a large scale in real-world scenarios. Autonomous driving lacks the risk perception of human drivers, making it even more important to test these rare risk scenarios. Therefore, in this example, the C-cut in-T scenario is selected as an example for scenario generation to enrich the scenario library for autonomous driving testing.

[0052] The parameters to be determined were calibrated based on the relevant calculation of the vehicle motion interaction feature F in Example 1, and the calibration results of the parameters to be determined are shown in Table 3. The vehicle motion feature F was calculated using the calibrated parameters, and the distribution curve of the vehicle motion interaction feature F was drawn using the kernel density estimation method. Then, the peak value was detected to determine the threshold value. is -0.68, that is When , it is a strong interaction state, otherwise it is a weak interaction. Finally, the vehicle trajectory information and vehicle motion interaction characteristics are obtained. The target scene data set .

[0053] Table 3 Parameters after calibration An F-TimeGAN model was constructed according to the method of Example 1, and candidate hyperparameters of the F-TimeGAN model were set. The extracted lane-changing scene data was input into the F-TimeGAN model under different candidate hyperparameters for scene generation. Each lane-changing sample had the same time step of 100 frames. The candidate hyperparameters set for the F-TimeGAN model are shown in Table 4: Table 4 F-TimeGAN model parameter settings The scene library generated by the optimal model is obtained through steps S5 and S6 in Example 1 As shown in Table 5, a comparative analysis of quantitative evaluation indicators (including discrimination score, prediction score, JS divergence, and FID score) shows that the F-TimeGAN model significantly outperforms the baseline TimeGAN model in terms of the similarity between generated data and real data. Judging from the evaluation results of the prediction score, the data generated by the method in Example 1 is more temporal.

[0054] Table 5 Comparison of results between F-TimeGAN model and TimeGAN model The generated data of the F-TimeGAN model is reduced in dimension and visualized using PCA and t-SNE algorithms. Figure 3 and Figure 4 As shown in the figure, the red points represent real data, and the blue points represent generated data. Figure 3 and Figure 4 It can be found that the data generated by the F-TimeGAN model and the original data have a large overlap, which shows that the F-TimeGAN model can better cover the previous real scene. Figure 3 and Figure 4 As shown, it can be found that there is a moderate distribution shift between the data generated by the F-TimeGAN model and the original features, indicating that this method enhances the diversity of generated data.

[0055] Furthermore, the generated data of the optimal F-TimeGAN model and the TimeGAN model were screened for strong interaction scenarios, risk scenarios, and strong interaction risk scenarios, and the results are shown in Table 6. In terms of the number of risk samples, compared with the real data and the unimproved TimeGAN model, the F-TimeGAN model generated a maximum of 341 risk scenarios, which is 3.79 times that of the real data and 2.66 times that of the TimeGAN model. Further comparison of the risk coefficient, such as Figure 5 As shown in Figure 2, the risk coefficient of a large number of data samples is less than 0.1. After extracting samples with risk coefficients greater than or equal to 0.1, Figure 6 As can be seen, the data generated by the F-TimeGAN model has a higher risk factor. In the real data, there are only two scenarios with a risk factor greater than 0.3, while the TimeGAN model generates three scenarios with a risk factor greater than 0.3. In contrast, the F-TimeGAN model generates 11 scenarios with a risk factor greater than 0.3, a 5.5-fold increase compared to the real data and a 3.7-fold increase compared to the unmodified TimeGAN model.

[0056] Table 6 Risk and interaction sample analysis By calculating the lane-changing interaction level (IEL), from Table 6 and Figure 7 As can be seen from the data, the F-TimeGAN model generates samples with a higher degree of interaction. Specifically, there are only 183 samples with an IEL greater than 0.5 in the real data, 230 in the TimeGAN-generated data, and 444 in the F-TimeGAN-generated data. The F-TimeGAN model generates 2.42 times more highly interactive lane-changing samples than the real data, and 1.93 times more than the unmodified TimeGAN model.

[0057] Furthermore, to avoid the contingency of analyzing a single experiment, this example conducted a large number of experiments on models with different parameters. The results are shown in Tables 7 to 9. In comparison, the F-TimeGAN model has higher coverage. The different numerical values ​​in the data source represent "model - hidden units - number of layers - batch size - sensitivity coefficient", where 1 indicates coverage and 0 indicates non-coverage.

[0058] Table 7 Initial relative positions of two vehicles of different models Table 8 Horizontal displacement of lane-changing vehicles in different models Table 9 Horizontal displacement of the rear vehicle of different models Tables 7-9 show that the data generated by the TimeGAN model for some parameter combinations is not reasonable, reflecting the large jitter in the generated trajectories. This requires extensive training and parameter tuning to accurately predict the TimeGAN model. In contrast, the F-TimeGAN model better learns the vehicle's motion characteristics, resulting in less jitter, smoother trajectories, and a closer resemblance to real data.

[0059] Example 3: This embodiment provides a system for generating risk scenarios for autonomous driving in a highly interactive environment. The system includes a processor and a memory. The memory stores a computer program. When the processor runs the computer program, the method in Example 1 is executed.

[0060] The system of this embodiment can efficiently generate strong interaction risk scenario data that is difficult to collect in reality based on conventional target scenario data, effectively solving the problem of scarce key test scenario data and providing scenario data support for safety verification of autonomous driving in a strong interaction environment.

[0061] Example 4: This embodiment provides a storage medium, in which a computer program is stored. When the computer program is executed, the method in embodiment 1 is executed.

[0062] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for generating risk scenarios for autonomous driving in a highly interactive environment, characterized in that: include: S1. Collect road vehicle trajectory data to build a basic database; S2. Extract the target scene from the basic database and calculate the vehicle motion interaction characteristics of the target scene ; S3, based on vehicle motion interaction features Build the F-TimeGAN model; S4. Set candidate hyperparameters of the F-TimeGAN model and input the extracted target scene into the F-TimeGAN model under different candidate hyperparameters to generate the scene; S5. Evaluate the similarity between the scene data generated by the F-TimeGAN model and the target scene data under different candidate hyperparameters, and select the one with the highest similarity. F-TimeGAN models are used as alternative models, among which is a natural number greater than or equal to 1; S6. Perform risk assessment on the scenario library generated by the alternative models to obtain the number of risk scenarios for each alternative model; If the number of risk scenarios of the candidate model is greater than the number of target scenarios, the F-TimeGAN model that can generate the most risk scenarios is selected as the optimal model; if no alternative model exists, return to step S4 to reset the candidate hyperparameters of the F-TimeGAN model; S7. Filter out high-interaction risk scenarios from the scenario data generated by the optimal model, and output the screening results to the autonomous driving test scenario library.

2. The method for generating risk scenarios for autonomous driving in a highly interactive environment according to claim 1, characterized in that: The vehicle motion interaction feature Expressed as: in, is the vector connecting the center points of the two vehicles’ heads, The direction of the following vehicle and The angle of The direction of the vehicle ahead and The angle of express The speed of the car after time, express The speed of the car ahead at that moment, for Speed ​​of the vehicle behind In vector The weight on for Speed ​​of the preceding vehicle at the moment In vector The weight on Represents the static safety distance parameter, For safe headway, Indicates the comfortable deceleration of the following vehicle. is the maximum deceleration of the preceding vehicle, is the maximum deceleration of the following vehicle, To balance the distance, when two vehicles are in adjacent lanes, , when two vehicles are in the same lane , Indicates the actual distance between the two workshops, is the speed difference sensitivity coefficient, is the expected speed of the following vehicle, Indicates the reaction time of the following vehicle. is the depth of the potential well, is the dimensionless Morse potential parameter, is the base of the natural logarithm function.

3. The method for generating risk scenarios for autonomous driving in a highly interactive environment according to claim 2, characterized in that: Based on vehicle motion interaction features Determine the vehicle motion interaction state of the scene, specifically: in, express The vehicle is in a strong interactive state at all times. express The vehicle is in a non-strong interaction state at all times. for The vehicle motion interaction characteristics at each moment, is the threshold of strong interaction state and Less than 0, is the degree of interaction exposure, Indicates the total time in the strong interaction state in the scene, Indicates the total time of the scene, Indicates that the current scene is a strong interaction scene. Indicates that the current scene is a non-strong interaction scene.

4. The method for generating risk scenarios for autonomous driving in a highly interactive environment according to claim 1, characterized in that: Introducing vehicle motion interaction features into the generator The attention mechanism is used to build the F-TimeGAN model, specifically: in, Indicates the calculation of attention weights, represents the hidden state of the trajectory mapping of the main vehicle and other vehicles in the original time series data, is the query vector in the attention mechanism module, is the key vector in the attention mechanism module, is the value vector in the attention mechanism module, Query vector in the attention mechanism module The linear mapping parameters of is the key vector in the attention mechanism module The linear mapping parameters of is the median vector of the attention mechanism module The linear mapping parameters of Represents the key vector Dimensions, is the key vector The transposed matrix of represents scaled dot product attention; is the sensitivity adjustment parameter; The weighted hidden state, Vehicle motion interaction features The tanh function value of The generator of the F-TimeGAN model is expressed as: in, represents the static feature latent code, A generative network representing static features, represents a vector space that defines a known distribution, represents the current time feature potential code, represents the generative network for temporal features, represents a random vector, Represents the potential code of the previous time feature.

5. The method for generating risk scenarios for autonomous driving in a highly interactive environment according to claim 4, characterized in that: Input the target scene data set into the F-TimeGAN model Expressed as: in, is the total number of samples, Contains the trajectory information of the main vehicle Ego and other vehicles OV and the vehicle motion interaction characteristics between the two , represents the trajectory of the main vehicle, represents the trajectory of other vehicles, express The host vehicle's position at the moment, express The positions of other vehicles at the moment, express The vehicle motion interaction characteristics between the host vehicle and other vehicles at the moment, set Each sample in has the same time step , ; The scene data set generated by the F-TimeGAN model is represented as: in, The total number of samples generated for the F-TimeGAN model, The first samples, The main vehicle trajectory generated by the F-TimeGAN model, Generated for the F-TimeGAN model The host vehicle's position at the moment, Other vehicle trajectories generated by the F-TimeGAN model, Generated for the F-TimeGAN model m The position of other vehicles at the moment, the collection Each generated sample has the same time step , .

6. The method for generating risk scenarios for autonomous driving in a highly interactive environment according to claim 1, characterized in that: The method for confirming the alternative model in step S5 is: Filter out generated data that meets < and and The F-TimeGAN model is then used to calculate the comprehensive score of each selected F-TimeGAN model. , the selected F-TimeGAN models are ranked according to their comprehensive scores Sort from low to high, select the one at the front An F-TimeGAN model was used as an alternative model; Among them, the comprehensive score Expressed as: in, is the FID score, is the prediction score, is the discriminant score, is the JS divergence value, 、 and All are set thresholds.

7. The method for generating risk scenarios for autonomous driving in a highly interactive environment according to claim 1, characterized in that: In step S6, risk assessment is performed on the scenario library generated by the candidate model, specifically: The driving risk of the scene is assessed using the stopping sight distance SSD and the stopping distance index SDI. SSD and SDI are expressed as: in, is the vehicle braking distance, Indicates the driver's reaction distance, Indicates vehicle exist The speed of time, Indicates the driver's reaction time, is the deceleration; in, express The distance between the front of the vehicle in front and the front of the vehicle behind at the moment, Indicates that the vehicle ahead is Stopping sight distance at all times, Indicates that the following vehicle is Stopping sight distance at all times, Indicates the vehicle length of the preceding vehicle; Among them, when 1 means Time Vehicle At risk, 1 means that the scenario is considered a risk scenario. 0 means Time Vehicle In a safe state.

8. The method for generating risk scenarios for autonomous driving in a highly interactive environment according to claim 7, characterized in that: By collision risk factor Determine the risk level of the scene and the collision risk coefficient Greater than threshold The scene is considered to be a high-risk scene; the collision risk coefficient Expressed as: in, is the degree of risk exposure; Indicates the total time in the risk state in the scenario; Indicates the total driving time of the scene; is the severity of the risk; is the observed maximum stopping distance index; is the theoretical maximum stopping distance index.

9. A system for generating risk scenarios for autonomous driving in a highly interactive environment, characterized in that: The system includes a processor and a memory, wherein a computer program is stored in the memory, and when the processor runs the computer program, the method according to any one of claims 1 to 8 is executed.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed, executes the method according to any one of claims 1 to 8.

Images