A method and system for generating a vehicle test boundary scenario based on a pre-boundary scenario

Abstract

The application discloses a vehicle test boundary scene generation method and system based on a pre-boundary scene, comprising the following steps: obtaining a first feature data set; obtaining a real boundary scene and a real pre-boundary scene based on the first feature data set and a preset risk discrimination criterion; constructing a second feature data set based on real pre-boundary scene data and the real boundary scene data; obtaining generated pre-boundary scene data by using a generation model based on the real pre-boundary scene data; training a prediction model based on the real pre-boundary scene data and the real boundary scene data; and obtaining test prediction boundary scene data based on the generated pre-boundary scene data and the prediction model. The method provided by the application obtains test prediction boundary scene data based on real pre-boundary scene data, and uses a generation model and a prediction model to expand real boundary scene data, thereby solving the technical problem that existing test scene data is sparse due to the limited space of generated scene data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent vehicle testing, and in particular to a vehicle test boundary scenario generation method and system based on a pre-boundary scenario. BACKGROUND

[0002] With the rapid development of autonomous driving technology, the importance of its intelligent testing link is increasingly prominent, and the scene-based safety testing method is becoming a key way to replace the traditional long-distance, long-time global testing. Among them, the boundary scenario located in the logical scene parameter space, between the collision risk and the safety boundary, has the dual characteristics of low occurrence probability and high safety risk, which can accelerate the verification of the system's ability boundary under extreme conditions, thereby comprehensively evaluating its safety and robustness. Therefore, in the process of autonomous driving testing and verification, how to efficiently generate sparse boundary scenarios and provide high-quality scene data for subsequent systematic testing has become a problem to be solved.

[0003] In the prior art, the methods for scene generation mainly include deep generation model, search optimization algorithm and generative adversarial network model. Among them, the patent CN119783568A discloses an intelligent vehicle extreme test scenario generation method and system based on deep learning. The patent is based on deep learning, and a hybrid adversarial generation network scene is built to generate an extreme test scenario with high risk controllability. The patent CN119808597A discloses an automatic driving simulation test scenario generation method, system, device and storage medium. The method generates a diversified traffic scene by using a differential evolution algorithm based on a random test scenario.

[0004] At present, the scene generation method highly depends on limited real scene data for training and optimization, resulting in insufficient sample diversity and limited coverage, making it difficult to effectively simulate extreme or rare boundary conditions. In addition, most generation processes often rely on deep models or search algorithms alone, lack of prediction and verification of the generated results, and the generation efficiency is low and it is difficult to control the quality of the sampling results in advance. Moreover, the existing scene generation scheme only focuses on the direct reconstruction of known scenes, lacks system description and utilization of their generalization predecessors, resulting in limited generation space and difficulty in covering more potential extreme situations.

[0005] Therefore, it is necessary to provide a vehicle test boundary scenario generation method based on a pre-boundary scenario to solve or at least partially alleviate the above technical problems. SUMMARY

[0006] The technical problem to be solved by the present invention is to provide a method for generating vehicle test boundary scenarios based on pre-boundary scenarios, which aims to solve the technical problem in the prior art that the generated test scenario data is sparse and difficult to cover potential extreme situations due to the limited space of scenario generation data.

[0007] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows:

[0008] A method for generating vehicle test boundary scenarios based on pre-boundary scenarios includes the following steps: S10, acquiring current vehicle trajectory data and vehicle-related feature parameters, wherein the current vehicle trajectory data is the current target vehicle at various time points. The corresponding trajectory frame data set includes vehicle-related feature parameters, including the current vehicle feature parameters and environmental vehicle feature parameters corresponding to each trajectory frame. The current vehicle feature parameters are the vehicle operation parameters of the current target vehicle corresponding to the trajectory frame, and the environmental vehicle feature parameters are the vehicle operation parameters of the environmental vehicles related to the current target vehicle position. After cleaning, denoising and format standardization of the current vehicle trajectory data and vehicle-related feature parameters, the first feature dataset is obtained.

[0009] S20, based on the first feature dataset, calculate and obtain the current target vehicle at each time step. The corresponding driving risk level is determined based on preset risk assessment criteria for each time period. The corresponding driving risk level is judged and the true boundary scene is extracted. The true boundary scene is the trajectory frame corresponding to the driving risk level not greater than the risk judgment standard. The true boundary scene has corresponding true boundary scene data.

[0010] S30, obtain the length of the historical window before determining the real boundary scene. corresponding The frame is a real pre-boundary scene, and the real pre-boundary scene has corresponding real pre-boundary scene data. A second feature dataset is constructed based on the real pre-boundary scene data and the real boundary scene discrimination data.

[0011] S41, Generate pre-boundary scene data by using a generative model based on real pre-boundary scene data;

[0012] S42, a prediction model is trained and obtained based on real pre-boundary scene data and real boundary scene discrimination data;

[0013] S50 obtains test predicted boundary scene data based on generated pre-boundary scene data and prediction model.

[0014] Furthermore, in step S10,

[0015] If the current target vehicle is a car-following vehicle, then at any time Corresponding associated feature parameter data = ;

[0016] If the current target vehicle is a lane-changing vehicle, then at any time... Corresponding associated feature parameter data = ;

[0017] in, For the current target vehicle instantaneous speed, For the current target vehicle Instantaneous acceleration, For the current target vehicle Environmental vehicles Current distance, For environmental vehicles instantaneous speed, For environmental vehicles Instantaneous acceleration, The current time frame. , and They are positive integers, The total number of vehicles in the first feature dataset;

[0018] The time indicates the vehicle preceding the current target vehicle before it changes lanes. The time indicates the vehicles following the current target vehicle before it changes lanes. The time indicates the vehicle in front of the target vehicle after it changes lanes. This indicates the vehicle following the current target vehicle after it changes lanes.

[0019] Furthermore, in step S20, the formula is used.

[0020] Perform calculations.

[0021] in,

[0022]

[0023] in, for Current vehicle in time frame Environmental vehicles MTTC value, for Current vehicle in time frame The minimum MTTC value.

[0024] Furthermore, in step S20, the formula is used. Perform calculations.

[0025] wherein, is a real boundary scene, denotes the time point corresponding to the real boundary scene, denotes the time frame in the trajectory sequence, is a preset MTTC risk threshold, when the MTTC value at a certain time point is less than or equal to for the first time, the corresponding trajectory frame is determined as the real boundary scene, is the real boundary scene data at time .

[0026] Further, in step S30, the formula is used for calculation,

[0027] wherein, is a real pre-boundary scene, denotes the time point corresponding to the real boundary scene, denotes the number of frames before the occurrence of the real boundary scene, is the vehicle-related feature parameters and the environmental vehicle feature parameters at time (real pre-boundary scene data).

[0028] Further, in step S41, in the generation training process of the generation model, the bulldozer distance algorithm and the JS divergence algorithm are used as evaluation indexes for measuring the similarity between the generated data distribution and the real data distribution, and the generation model is optimized by minimizing the bulldozer distance algorithm and the JS divergence algorithm as evaluation indexes.

[0029] Further, the generation model adopts a deep learning model or a search algorithm model.

[0030] Further, the formula

[0031]

[0032] ,

[0033] ,

[0034] is used for calculation,

[0035] wherein, denotes a set of joint probability distributions, denotes a joint measure on the product space , and is the real pre-boundary scene data in the second feature data set under the time frame . to generate pre-boundary scene data, to generate a model, and respectively represent the distribution of generated data and real data, is the average distribution of generated data and real data, represents the KL divergence, represents the bulldozer distance algorithm, represents the JS divergence algorithm.

[0036] Further, in step S42,

[0037] using the formula

[0038]

[0039] the calculation is performed,

[0040] wherein, is the discriminative real boundary scene data in the time frame of the second feature data set, is the real pre-boundary scene data in the time frame of the second feature data set, is the prediction model, is the model prediction boundary scene data of the model in the training process , represents the scene , is the total number of scenes, is the root mean square error, is the mean absolute error.

[0041] The present application also provides a vehicle test boundary scene generation system based on pre-boundary scene, comprising

[0042] a data acquisition module for acquiring current vehicle trajectory data and vehicle associated feature parameters;

[0043] a scene discrimination module for extracting discriminative real boundary scenes;

[0044] a generation-prediction module for obtaining test prediction boundary scene data based on real pre-boundary scene data;

[0045] a data processing module for implementing the steps of the vehicle test boundary scene generation method based on pre-boundary scene described above.

[0046] Compared with the prior art, the present application has the following advantages:

[0047] The application provides a vehicle test boundary scene generation method based on a pre-boundary scene, which comprises the following steps: obtaining first feature data sets by processing original data after obtaining current vehicle trajectory data and vehicle associated feature parameters, wherein the first feature data sets comprise current vehicle trajectory data, current vehicle feature parameters and environmental vehicle feature parameters after data processing; performing risk discrimination based on preset risk discrimination criteria and extracting a discrimination real boundary scene based on current vehicle feature parameters and environmental vehicle feature parameters corresponding to each frame (each time point) , that is, each scene; performing scene generalization based on the discrimination real boundary scene, and obtaining a historical window length of corresponding to the discrimination real boundary scene before extraction frame real pre-boundary scene, wherein the real pre-boundary scene has corresponding real pre-boundary scene data, a second feature data set is constructed based on the real pre-boundary scene data and the discrimination real boundary scene data, and scene generalization is used to expand the scene number and scene data; then, real pre-boundary scene data is used to obtain generated pre-boundary scene data by using a generation model, and a prediction model is obtained by training based on the real pre-boundary scene data and the discrimination real boundary scene data; finally, test prediction boundary scene data is obtained based on the generated pre-boundary scene data and the prediction model, and the generated pre-boundary scene data is expanded by using the generation model and the prediction model. The vehicle test boundary scene generation method based on the pre-boundary scene of the application uses scene generalization to expand the scene number and scene data to obtain real pre-boundary scenes and real pre-boundary scene data, and reasonably expands the generated pre-boundary scene data to obtain test prediction boundary scene data. The test prediction boundary scene data is obtained based on the real pre-boundary scene data, the real boundary scene data is expanded by using the generation model and the prediction model, and the technical problem that the generated test scene data is difficult to cover potential extreme situations due to the limited scene generation data space is solved. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 is a flowchart of the vehicle test boundary scene generation method based on the pre-boundary scene of the application;

[0049] Figure 2 is a driving state schematic diagram in the vehicle test boundary scene generation method based on the pre-boundary scene of the application, wherein a is a lane changing state schematic diagram, and b is a following state schematic diagram. DETAILED DESCRIPTION

[0050] It should be understood that the specific embodiments described herein are only used to explain the application, and are not used to limit the application.

[0051] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative effort are within the protection scope of the present application.

[0052] In addition, the descriptions involving “first”, “second” and the like in the present application are only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features defined with “first”, “second” can explicitly or implicitly include at least one of the features. In addition, the technical solutions of the various embodiments can be combined with each other, but it must be based on the fact that a person of ordinary skill in the art can realize it. When the combination of technical solutions contradicts each other or cannot be realized, it should be considered that the combination of technical solutions does not exist and is not within the protection scope of the present application.

[0053] Please refer to Figure 1 and Figure 2 An embodiment of the present application provides a vehicle test boundary scene generation method based on a pre-boundary scene, comprising the steps of:

[0054] S10, obtaining current vehicle trajectory data and vehicle associated feature parameters, the current vehicle trajectory data being a data set of a trajectory frame corresponding to each time t of a current target vehicle, the vehicle associated feature parameters including current vehicle feature parameters and environment vehicle feature parameters corresponding to each trajectory frame, the current vehicle feature parameters being vehicle operation parameters of the current target vehicle corresponding to the trajectory frame, and the environment vehicle feature parameters being vehicle operation parameters of an environment vehicle corresponding to a position of the current target vehicle, the first feature data set being obtained after data processing of cleaning, noise reduction and format standardization on the current vehicle trajectory data and the vehicle associated feature parameters;

[0055] S20, based on the first feature data set, calculating and obtaining a driving risk level of the current target vehicle corresponding to each time (tracking frame), and discriminating the driving risk level corresponding to each time and extracting a discrimination real boundary scene according to a preset risk discrimination criterion, the discrimination real boundary scene being a trajectory frame with a driving risk level not greater than the risk discrimination criterion, and the discrimination real boundary scene corresponding to discrimination real boundary scene data;

[0056] S30, obtaining a historical window length of corresponding to the discrimination real boundary scene; The frame is a real pre-boundary scene, and the real pre-boundary scene has corresponding real pre-boundary scene data. A second feature dataset is constructed based on the real pre-boundary scene data and the real boundary scene discrimination data.

[0057] S41, Generate pre-boundary scene data by using a generative model based on the real pre-boundary scene data in the second feature dataset;

[0058] S42, a prediction model is trained and obtained based on the real pre-boundary scene data and the real boundary scene data in the second feature dataset;

[0059] S50 obtains test predicted boundary scene data based on generated pre-boundary scene data and prediction model.

[0060] The vehicle test boundary scene generation method based on pre-boundary scene provided by this invention processes the raw data after acquiring the current vehicle trajectory data and vehicle associated feature parameters to obtain a first feature dataset. The first feature dataset includes the processed current vehicle trajectory data, current vehicle feature parameters, and environmental vehicle feature parameters. Then, based on each frame (each time point)... This involves analyzing the current vehicle feature parameters and environmental vehicle feature parameters for each scenario, performing risk assessment based on preset risk assessment criteria, and extracting the true boundary scenarios. Each true boundary scenario corresponds to true boundary scenario data. Then, scenario generalization is performed based on the true boundary scenarios to obtain the historical window length before the extraction of the true boundary scenarios. corresponding The present invention provides a vehicle test boundary scene generation method based on pre-boundary scenes. This method utilizes scene generalization to expand the number of scenes and scene data for obtaining real pre-boundary scenes and their corresponding real pre-boundary scene data. A second feature dataset is constructed based on the real pre-boundary scene data and the discriminant real boundary scene data. Scene generalization is used to expand the number of scenes and scene data, thereby solving the technical problem that existing methods struggle to cover potential extreme scenarios due to limited scene generation data space. A second feature dataset is constructed based on the real pre-boundary scene data and the discriminant real boundary scene data. A second feature dataset is then constructed by expanding the number of scenes and scene data using the real pre-boundary scene data and the discriminant real boundary scene data. Finally, test prediction boundary scene data is obtained based on the generated pre-boundary scene data and the prediction model, and the generated pre-boundary scene data is further expanded using the generative and predictive models.

[0061] Optionally, compared with the method of generating directly based on real boundary scene data using a single generation model, the present invention innovatively proposes a two-stage combined strategy of "generating pre-boundary scene data and predicting and obtaining test prediction boundary scene data (generation + prediction)", which realizes the efficient generation of vehicle test boundary scenes based on real pre-boundary scenes and effectively improves the relevance and efficiency of generation.

[0062] Optionally, the history window length can be 5, or 3, 7, 8, or other values, and the duration of a single window can be 30 seconds, or 1 minute, 2 minutes, or other values; if the frame rate is 60, the window duration is 1 second.

[0063] Furthermore, in step S10,

[0064] If the current target vehicle is a car-following vehicle, then at any time Corresponding associated feature parameter data = If the current target vehicle is a lane-changing vehicle, then at any time... Corresponding associated feature parameter data = ;in, For the current target vehicle instantaneous speed, For the current target vehicle Instantaneous acceleration, For the current target vehicle Environmental vehicles Current distance, For environmental vehicles instantaneous speed, For environmental vehicles Instantaneous acceleration, The current time frame. , and They are positive integers, The total number of vehicles in the first feature dataset; The time indicates the vehicle preceding the current target vehicle before it changes lanes. The time indicates the vehicles following the current target vehicle before it changes lanes. The time indicates the vehicle in front of the target vehicle after it changes lanes. The time indicates the vehicle following the current target vehicle after it changes lanes. Optionally, in a specific embodiment of the present invention, a [missing information] can be defined. This represents the instantaneous acceleration of the target vehicle.

[0065] Understandably, in the solution of this invention, current vehicle trajectory data and current vehicle feature parameters are obtained from roadside or vehicle-mounted sensing devices, and environmental vehicle feature parameters are obtained based on vehicle behavior. Environmental vehicles are defined as the vehicles in front and behind the current target vehicle in the same lane. If the current target vehicle changes lanes, the vehicles in front and behind in the lane after the lane change are added. Missing values ​​in the data are imputed using the mean, and outliers in the data are smoothed using the Loess method. A first feature dataset is constructed to provide data support for risk level assessment. At a certain moment... The first feature of the dataset can be represented as:

[0066] Following behavior vehicles

[0067] Lane changing vehicles

[0068] in, For the current target vehicle instantaneous speed, For environmental vehicles instantaneous speed, For the current target vehicle Instantaneous acceleration, For environmental vehicles Instantaneous acceleration, For the current target vehicle Environmental vehicles distance, For the given time frame, among the above feature parameters , This represents the number of vehicles in the initial dataset.

[0069] Furthermore, in step S10, the formula is used.

[0070]

[0071] Calculations are performed, in which,

[0072]

[0073] in, for Current vehicle in time frame Environmental vehicles MTTC value, for Current vehicle in time frame The minimum MTTC value.

[0074] Furthermore, in step S20, the formula is used.

[0075]

[0076] Calculations are performed, in which, To determine the true boundary scene, This indicates the moment corresponding to the determination of the real boundary scene. Represents a time frame in the trajectory sequence. The preset MTTC risk threshold is defined as the threshold at which the MTTC value first becomes less than or equal to a certain value. At that time, the corresponding trajectory frame is determined to identify the real boundary scene. For at any time The vehicle-related feature parameters and environmental vehicle feature parameters at that time. Understandably, the data for determining the true boundary scene includes time-related features. Vehicle-related feature parameters and environmental vehicle feature parameters at the time.

[0077] Optionally, in one specific embodiment, during the extraction and identification of the true boundary scenario, a quantitative assessment of the current target vehicle's risk level is performed. This is achieved by introducing Surrogate Safety Measures (SSMs) to quantify the risk level of the current target vehicle in the dynamic traffic environment, thereby obtaining the minimum risk level between the current target vehicle and surrounding vehicles, providing support for the accurate determination of the boundary scenario (identifying the true boundary scenario). In the scheme of this invention, the Modified Time-to-Collision (MTTC) indicator in the SSM system is preferably used for risk level assessment. The specific calculation method can be expressed as follows:

[0078]

[0079]

[0080] in, for Current vehicle in time frame Environmental vehicles MTTC value, for The minimum MTTC value of the current vehicle n in the time frame.

[0081] Furthermore, the criteria for identifying true boundary scenarios are further defined. A true boundary scenario is defined as the moment when the risk level first falls below the set MTTC collision threshold in the current vehicle trajectory data. Specifically, in this scheme, the improved collision time is used as the risk level evaluation index, and the MTTC threshold is set accordingly. As a criterion for determining collision risk, the criteria for determining boundary scenarios can be expressed as:

[0082]

[0083] wherein, to determine the real boundary scene, represents the time corresponding to the determination of the real boundary scene, represents the time frame in the trajectory sequence, is a preset MTTC risk threshold, when the MTTC value at a certain time is less than or equal to for the first time, the corresponding trajectory frame is determined as the determination of the real boundary scene, is the real boundary scene data at time Through the above method, the automatic determination of the real boundary scene based on the dynamic risk evolution process in the trajectory data (trajectory sequence) is realized, and an accurate and reasonable scene basis is provided for subsequent generation and prediction modeling.

[0084] Further, in step S30, the formula is used for calculation,

[0085] wherein, is the real pre-boundary scene, represents the time corresponding to the determination of the real boundary scene, represents the number of frames before the occurrence of the determination of the real boundary scene, is the vehicle-related feature parameter and the environmental vehicle feature parameter at time , that is, the real pre-boundary scene data at time .

[0086] In a specific embodiment of the present application, based on the determination of the real boundary scene, the definition of the real pre-boundary scene is innovatively proposed. The real pre-boundary scene refers to the current vehicle trajectory data and the current vehicle feature parameter and the environmental vehicle feature parameter within the continuous frame time window before the determination frame of the real boundary scene in the current vehicle trajectory data (trajectory sequence). By introducing the real pre-boundary scene, the present scheme can effectively depict the dynamic evolution process of the current target vehicle before reaching the real boundary scene, supplement and enrich the historical state information of the real boundary scene, and improve the accuracy of subsequent models in scene generation and risk evolution modeling. Assuming that the time frame corresponding to the determination of the real boundary scene is , the corresponding pre-boundary scene is the feature data sequence of the current target vehicle and the environmental vehicle within the time interval , wherein is the set historical window length.

[0087] Specifically, the extraction method of the real pre-boundary scene is as follows:

[0088]

[0089] wherein, is a real pre-boundary scenario, represents the time frame corresponding to the real boundary scenario, that is, the moment when the risk level is first lower than the set MTTC collision threshold , is a predefined historical time window size, representing the continuous trajectory data from frame before the boundary scenario occurs, is the current vehicle feature data and the environment vehicle feature data at the moment .

[0090] The scheme of the present application extracts real pre-boundary scenarios, real pre-boundary scenario data, and discriminative real boundary scenario data, and combines the real pre-boundary scenario data and the discriminative real boundary scenario data to construct a unified second scene feature data set, providing comprehensive and systematic feature support for subsequent model training, generation, and prediction processes.

[0091] Further, in the generation training process of the generation model in step S41, the bulldozer distance algorithm and the JS divergence algorithm are used as evaluation indexes for measuring the similarity between the generated data distribution and the real data distribution, and the generation model is optimized by minimizing the bulldozer distance algorithm and the JS divergence algorithm as evaluation indexes.

[0092] Further, the generation model adopts a deep learning model or a search algorithm model.

[0093] Further, the formula

[0094] ,

[0095] ,

[0096] ,

[0097]

[0098] is calculated, wherein, represents a set of joint probability distributions, represents a joint measure on the product space ( ), is real pre-boundary scenario data in the second feature data set at time frame , is generated pre-boundary scenario data, is a generation model, and represent the distribution of generated data and real data, respectively, is the average distribution of the generated data and the real data, represents the KL divergence, represents the Earth Mover’s Distance algorithm, represents the JS divergence algorithm.

[0099] In a specific embodiment of the present application, the generated pre-boundary scene data is obtained based on the real pre-boundary scene through a generative model, and in the generative training process, a corresponding performance evaluation and optimization strategy is designed in combination with the stage characteristics. Specifically, for the generative model, the present scheme adopts the Earth Mover’s Distance (EMD) and the Jensen-Shannon Divergence (JSD) as evaluation indexes to measure the similarity between the distribution of the generated data and the distribution of the real data, and optimizes the generative model by minimizing the EMD and JSD indexes to ensure that the generated pre-boundary scene is reasonably and diversely distributed in the feature space. The specific calculation method can be represented as:

[0100]

[0101]

[0102]

[0103]

[0104] In the formula, is the real pre-boundary scene data in the time frame of the second feature data set, is the generated pre-boundary scene data, is the generative model (Time GAN) adopted by the present scheme. and respectively represent the distribution of the generated data and the real data, is the average distribution of the two, represents the KL divergence (Kullback-Leibler Divergence).

[0105] Further, in step S42,

[0106] the formula

[0107]

[0108] is used for calculation,

[0109] wherein,​​​ discriminative real boundary scenario data in the second feature data set, real pre-boundary scenario data in the second feature data set, for the prediction model, for the prediction model in the training process, representing the scenario ,

[0110] Further, the formula is used to obtain test prediction boundary scenario data, wherein,

[0111] In a specific embodiment of the present application, test prediction boundary scenario data is obtained based on generated pre-boundary scenario data and a prediction model. Specifically, for the prediction model, first, the model is trained based on real pre-boundary scenario data and discriminative real boundary scenario data in the second feature data set. In this scheme, root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators to quantify the closeness of the predicted boundary scenario and the real boundary scenario in the feature value during the training process, and to guide the performance optimization of the prediction model. Through targeted and multi-index evaluation and optimization mechanism, this scheme effectively improves the distribution consistency and numerical accuracy of discriminative real pre-boundary scenario data and generated boundary scenario data, thereby ensuring the quality and application reliability of the boundary scenario.

[0112] In the scheme of the present application, some specific feature names are described as follows:

[0113] Discriminative real boundary scenario: refers to the time frame when the risk level is lower than the set SSM collision threshold for the first time in the current vehicle trajectory data.

[0114] Real pre-boundary scenario: refers to the trajectory and feature data within the continuous frame time window located before the discriminative real boundary scenario (discriminative real boundary frame) in the current vehicle trajectory data.

[0115] ​​​​​​​​Generalization: The process of expanding from specific, individual instances to general cases. In this context, it refers to expanding from emergency, dangerous scenarios that distinguish real boundary cases to non-emergency, real pre-boundary scenarios.

[0116] SSM (Surrogate Safety Measures): Surrogate safety measures refer to indicators found in traffic data that approximate near-miss or potential collision events (but not actual collisions).

[0117] MTTC (Modified Time-to-Collision): Modified Time-to-Collision is an improved version of the traditional Time-to-Collision (TTC) metric. It incorporates additional factors such as speed changes and acceleration to improve the accuracy of potential conflict prediction.

[0118] Time GAN (Time-series Generative Adversarial Networks): Time-series Generative Adversarial Networks is a generative model used to generate high-quality time series data. It combines the GAN framework and time series modeling characteristics, commonly used for data augmentation, simulating future sequences, and other tasks.

[0119] SOFTS (Series-core Fused Time Series forecaster): Series-core Fused Time Series forecaster is a multi-element time series prediction method based on sequence core fusion. It extracts the core dynamic information in the sequence and performs fusion modeling to improve the accuracy of complex time series prediction.

[0120] EMD (Earth Mover’s Distance): Earth Mover's Distance is a measure of the difference between two distributions, which is metaphorically described as the minimum "work" required to move a pile of "soil" (probability distribution) into another pile.

[0121] JSD (Jensen–Shannon Divergence): Jensen-Shannon Divergence is a symmetric, finite measure of the difference between two probability distributions, commonly used to evaluate the similarity between two probability distributions. It is an improvement over the KL divergence (Kullback-Leibler Divergence).

[0122] RMSE (Root Mean Square Error): Root Mean Square Error is a commonly used evaluation metric in regression tasks, representing the square of the mean deviation between predicted and actual values. It is sensitive to large errors.

[0123] MAE (Mean Absolute Error): Mean Absolute Error, another commonly used evaluation metric in regression tasks, represents the average of the absolute errors between predicted and true values, less sensitive to outliers than RMSE.

[0124] PCA (Principal Component Analysis): Principal Component Analysis, a commonly used data dimensionality reduction technique, projects the original data onto the direction with the largest variance through linear transformation, extracting the most important information features.

[0125] t-SNE (t-Distributed Stochastic Neighbor Embedding): t-Distributed Stochastic Neighbor Embedding, a nonlinear dimensionality reduction method, especially suitable for mapping high-dimensional data to low-dimensional (usually two or three-dimensional) for visualization, can well preserve the local data structure.

[0126] The present application provides a preferred example method as follows:

[0127] CitySim dataset, specifically using its Freeway B (Basic Segment) Highway Asia data subset. The dataset is collected through unmanned aerial vehicle aerial photography technology, with a flight height of 320 meters, using a 30Hz sampling frequency to record continuously for 60 minutes. The data collection section is located in the basic segment of the typical Asian highway, with an effective observation road length of 620 meters, using the right-hand traffic rule, with a six-lane configuration (single three-lane structure). The original data statistics show that a total of 9,288 effective passing vehicles were recorded during the observation period, of which 7,576 were small passenger cars (accounting for 81.56%), and 1,712 were freight vehicles (accounting for 18.44%), with an average traffic flow rate of 2.58 vehicles per second, containing 34 feature dimensions, after feature screening, 11 key parameters are retained for analysis, as shown in Table 1 (real data features).

[0128] Car lane changing dataset, derived from field data collected on a straight highway section. The data collection section is a two-way 4-lane straight road, with an effective observation road length of 300 meters, and the legal speed limit range is 60-120 km / h. After standardization processing, a total of 7,155 complete current vehicle trajectory data were captured, with an effective record duration of 116.47 minutes, of which 5,467 were small passenger cars (accounting for 76.4%), and 1,688 were freight vehicles (accounting for 23.6%), with an average traffic flow rate of 0.99 vehicles per second. The data feature extraction standard refers to the CitySim dataset specification to ensure the comparability of data structure with existing research.

[0129] Table 1

[0130]

[0131] For the large truck lane changing behavior, simulation data is used for experimental verification. In the simulation data of the automatic driving large truck lane changing in this scheme, the simulation scene is set as a five-lane highway with good road conditions for mixed passenger and freight traffic. The leftmost lane is an automatic driving large truck dedicated lane. The mixed traffic C-HDV (Car-Human-Driven Vehicle), T-HDV (Truck-Human-Driven Vehicle), C-CAV (Car-Connected-Automated Vehicle) and T-CAV (Truck-Connected-Automated Vehicle) are mixed. T-CAV merges from the ramp to the main road and changes lanes to the leftmost dedicated lane.

[0132] For this simulation environment, the value range of each key parameter in the simulation is calibrated based on real data, and the remaining non-key parameters are the default values of SUMO, as shown in Table 2 (simulation parameters). To distinguish the behavior characteristics of automatic driving and manual driving, this study uses the Krauss model to simulate the driving behavior of manually driven vehicles, while the driving behavior of automatic driving vehicles is controlled by the IDM model (Intelligent Driver Model).

[0133] Table 2

[0134]

[0135] The simulation frequency is set to 0.1 seconds, i.e. 10Hz. The simulation is terminated when one of the following conditions is met: a) collision occurs; b) T-CAV successfully changes lanes to the dedicated lane. In each simulation round, the key scene parameters of each simulation step are recorded, and the parameters are shown in Table 3 (simulation output parameters).

[0136]

[0137] After processing the simulation data and real data, the number of real boundary scene data and real pre-boundary scene data obtained is shown in Table 4 (number of real boundary scene data and real pre-boundary scene data), and a unified second scene feature data set is constructed based on the obtained real boundary scene data and real pre-boundary scene data.

[0138] Table 4

[0139]

[0140] The training of the generation model Time GAN model and the prediction model SOFTS model based on the second scene feature data set, the performance evaluation results of the corresponding generation model after training of the vehicle interaction behaviors of the two vehicle types are shown in Table 5 (performance evaluation results of the generation model Time GAN), and in Table 5, is the current distance of the current target vehicle n and the environment vehicle, is the instantaneous speed of the environment vehicle, is the instantaneous acceleration of the environment vehicle, avg is the average value, respectively represent the current vehicle feature information of the current target vehicle, and from the evaluation results, the JSD and EMD values corresponding to each data set are in a relatively low range, reflecting that the generated data is highly consistent with the original data in terms of feature distribution. These feature distributions not only approach the original data, but also maintain a good consistency range. In addition, whether it is a speed-related feature, an acceleration-related feature, or a distance-related feature, the corresponding EMD and JSD values further indicate that the generated data and the original data have extremely high similarity in terms of distribution. In summary, the Time GAN model has excellent performance in reproducing the feature distribution of the original data. The generated data not only effectively preserves the feature structure of the original data, but also achieves highly consistent distribution matching on multiple key features.

[0141]

[0142] The performance evaluation results MAE and RMSE of the corresponding prediction model after training of the vehicle interaction behaviors of the two vehicle types are shown in Table 6 (performance evaluation results of the prediction model SOFTS), and it can be concluded that the prediction performance of the SOFTS model on multiple features is excellent, exhibiting high prediction accuracy and stability of the prediction results. In the lane changing behavior scene, the feature prediction accuracy is overall better than that in the following scene, especially in the prediction of speed features and distance features, which shows more significant advantages compared to acceleration features. Comprehensive analysis shows that the SOFTS model has significant advantages in overall prediction ability on various features, not only can accurately capture the data distribution and change trend of different features, but also can maintain low prediction error and high consistency in the overall range.

[0143]

[0144] ​​​After obtaining the generated pre-boundary scene data based on the real pre-boundary scene in the second scene feature data set using the trained generation model, the scheme uses PCA (principal component analysis) and t-SNE (nonlinear dimension reduction) method for dimension reduction comparison analysis of the real pre-boundary scene and the generated pre-boundary scene. Through the PCA method, the main variation characteristics of the data are extracted, and the consistency and diversity of the expanded scene in the feature space with the original scene are evaluated; at the same time, the t-SNE algorithm is used for nonlinear dimension reduction of the high-dimensional feature space, and the expansion and coverage of the generated scene in the feature distribution are intuitively displayed. It can be seen that the generalization data has high overlap with the original data in the reduced space, and the generated data is closely distributed around the original data, while the generated data appears a slight clustering separation phenomenon in the local area, which shows that the generated data has reasonable difference on the basis of maintaining the authenticity.

[0145] After obtaining the predicted boundary scene data based on the generated pre-boundary scene using the trained prediction model, the performance of the two-stage generation method (the "generation + prediction" method of the scheme) and the traditional single generation model (using the Time GAN model) in the boundary scene generation is compared and analyzed. Specifically, the proportion of the number of boundary scenes in the total number of scenes generated by the two methods is counted and compared, which is used as an evaluation index of the generation effect. It can be seen that with the increase of the length of the historical window, the proportion of the number of boundary scenes generally increases, and the longer historical window provides more rich time sequence information, which is helpful for the generation of complex scenes, especially in the lane changing scene. This result shows that sufficient historical time sequence information plays a key role in accurately modeling the boundary scene. Overall, in most scenes, the two-stage method is better than the single model generation method, especially in the lane changing scene. This shows that the two-stage method has obvious advantages in capturing the dynamic interaction between vehicles and adapting to complex scenes through pre-boundary scenes.

[0146] The application also provides a vehicle test boundary scene generation system based on pre-boundary scenes, comprising

[0147] A data acquisition module is configured to acquire current vehicle trajectory data and vehicle associated feature parameters.

[0148] A data discrimination module is configured to extract and discriminate real boundary scenes.

[0149] A generation-prediction module is configured to obtain test prediction boundary scene data based on the second feature data set real pre-boundary scene data.

[0150] A data processing module is configured to implement the steps of the vehicle test boundary scene generation method based on pre-boundary scenes.

[0151] The above is the preferred embodiment of the present application, the protection scope of the present application is not limited to the above-mentioned examples, all technical solutions under the idea of the present application belong to the protection scope of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the principle of the present application under the premise of several improvements and refinements, should be considered as the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for generating vehicle test boundary scenarios based on pre-boundary scenarios, characterized in that, Including the following steps: S10, acquire current vehicle trajectory data and vehicle-related feature parameters, wherein the current vehicle trajectory data is the current target vehicle at various times. The data set of the corresponding trajectory frames, the vehicle associated feature parameters include the current vehicle feature parameters and the environmental vehicle feature parameters corresponding to each trajectory frame, the current vehicle feature parameters are the vehicle operation parameters of the current target vehicle corresponding to the trajectory frame, and the environmental vehicle feature parameters are the vehicle operation parameters of the environmental vehicle related to the position of the current target vehicle. After cleaning, denoising and format standardization of the current vehicle trajectory data and the vehicle associated feature parameters, the first feature dataset is obtained. S20, based on the first feature dataset, calculate and obtain the current target vehicle at each time. The corresponding driving risk level is determined based on preset risk assessment criteria for each time period. The corresponding driving risk level is judged and the true boundary scene is extracted. The true boundary scene is the trajectory frame corresponding to the driving risk level not greater than the risk judgment criterion. The true boundary scene corresponds to the true boundary scene data. S30, obtain the length of the historical window before the determination of the real boundary scene. corresponding A frame of real pre-boundary scene, wherein the real pre-boundary scene has corresponding real pre-boundary scene data, and a second feature dataset is constructed based on the real pre-boundary scene data and the discriminated real boundary scene data; S41, Based on the real pre-boundary scene data, generate pre-boundary scene data is obtained using a generative model; During the generative training process of the generative model, the Bulldozer Distance algorithm and the JS divergence algorithm are used as evaluation metrics to measure the similarity between the generated data distribution and the real data distribution. The generative model is optimized by minimizing the Bulldozer Distance algorithm and the JS divergence algorithm as evaluation metrics; the formula is used. ， ， ， Perform calculations. in, express The set of joint probability distributions Indicates in the product space ( Joint measure on ) For the second feature dataset, time frames Real-world pre-boundary scenario data, To generate pre-boundary scene data, To generate models, and These represent the distributions of generated data and real data, respectively. It is the average distribution of generated data and real data. Denotes KL divergence, This represents the algorithm for bulldozer distance. This represents the JavaScript divergence algorithm; S42, a prediction model is trained and obtained based on the real pre-boundary scene data and the discriminated real boundary scene data; S50, based on the generated pre-boundary scene data and the prediction model, obtain test predicted boundary scene data.

2. The method for generating vehicle test boundary scenarios based on pre-boundary scenarios according to claim 1, characterized in that, In step S10, If the current target vehicle is a car-following vehicle, then at time... Corresponding associated feature parameter data ; If the current target vehicle is a lane-changing vehicle, then at time... Corresponding associated feature parameter data ; in, For the current target vehicle instantaneous speed, For the current target vehicle Instantaneous acceleration, For the current target vehicle Environmental vehicles Current distance, For environmental vehicles instantaneous speed, For environmental vehicles Instantaneous acceleration, The current time frame. , and They are positive integers, The total number of vehicles in the first feature dataset; The time indicates the vehicle preceding the current target vehicle before it changes lanes. The time indicates the vehicle following the current target vehicle before it changes lanes. The time indicates the vehicle preceding the current target vehicle after it changes lanes. The time indicates the vehicle following the current target vehicle after it changes lanes.

3. The method for generating vehicle test boundary scenarios based on pre-boundary scenarios according to claim 2, characterized in that, In step S20, Using formula Perform calculations. in, in, for Current vehicle in time frame Environmental vehicles MTTC value, for Current vehicle in time frame The minimum MTTC value.

4. The method for generating vehicle test boundary scenarios based on pre-boundary scenarios according to claim 3, characterized in that, In step S20, the formula is used. Perform calculations. in, To determine the true boundary scene, This indicates the moment corresponding to the determination of the real boundary scene. Represents a time frame in the trajectory sequence. The preset MTTC risk threshold is defined as the threshold at which the MTTC value first becomes less than or equal to a certain value. At that time, the corresponding trajectory frame is determined to identify the real boundary scene. For at any time Real-time boundary scene data for judgment.

5. The method for generating vehicle test boundary scenarios based on pre-boundary scenarios according to claim 2, characterized in that, In step S30, the formula is used. Perform calculations. in, For realistic pre-boundary scenarios, This indicates the moment corresponding to the determination of the real boundary scene. This indicates the number of frames before the actual boundary scene is determined to have occurred. For at any time Vehicle-related feature parameters and environmental vehicle feature parameters at the time.

6. The method for generating vehicle test boundary scenarios based on pre-boundary scenarios according to claim 1, characterized in that, The generative model employs a deep learning model or a search algorithm model.

7. The method for generating vehicle test boundary scenarios based on pre-boundary scenarios according to claim 6, characterized in that, In step S42, Using formula Perform calculations. in, For the second feature dataset, time frames The following data is used to determine the true boundary scene. For the second feature dataset, time frames Real-world pre-boundary scenario data, For prediction models, For the training process The model predicts boundary scene data. Represented as a scene , This represents the total number of scenes. The root mean square error, This represents the mean absolute error.

8. A vehicle test boundary scene generation system based on pre-boundary scenarios, characterized in that, include The data acquisition module is used to acquire current vehicle trajectory data and vehicle-related feature parameters; The scene discrimination module is used to extract and discriminate the real boundary scene; The generation-prediction module is used to obtain test prediction boundary scene data based on the real pre-boundary scene data; The data processing module is used to implement the steps of the vehicle test boundary scene generation method based on the pre-boundary scene as described in any one of claims 1 to 7.

Images