An automatic driving vehicle risk scene generation method and system suitable for a strong interaction environment and a storage medium

Abstract

The present application relates to the technical field of automatic driving, in particular to an automatic driving driving risk scene generation method and system suitable for strong interaction environment and a storage medium, the method comprising: firstly collecting data of a target scene such as following or lane changing, then calculating vehicle motion interaction features F of the target scene, and combining the attention mechanism to improve the TimeGAN model to obtain an F-TimeGAN model, inputting the target scene into the F-TimeGAN model for scene generation, and finally screening strong interaction risk scenes from the generated scenes for automatic driving test. The present application can efficiently generate strong interaction risk scene data which is difficult to collect in reality based on conventional target scene data, enhance the interaction and risk of the generated scene, effectively solve the problem of lack of key test scene data, and provide scene data support for safety verification of automatic driving in strong interaction environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, in particular to an automatic driving driving risk scene generation method and system suitable for strong interaction environment and a storage medium. BACKGROUND

[0002] The existing vehicle motion interaction scene generation method mainly depends on two technical paths of deep learning and reinforcement learning, but it generally has the problems of insufficient interaction feature modeling and limited scene authenticity.

[0003] The existing scene generation method generates scenes using off-line collected static trajectory data, which cannot capture the dynamic timing relationship of vehicle motion interaction. Secondly, the existing technology assumes that all vehicles have the same dynamic characteristics (such as braking performance, response delay), ignoring the significant differences in key parameters such as vehicle type and braking performance between heavy trucks and small cars, resulting in deviations between the generated scenes and the actual traffic flow. In addition, for high-complexity interaction scenes such as lane changing involving dynamic adjustment of heading angle, the evaluation method of existing vehicle motion interaction features is single, lacking consideration of the timing changes of the heading angle, and it is difficult to quantify complex interaction behaviors. However, automatic driving testing requires strong interaction and high-risk scene data, and such key scenes are difficult to collect in reality.

[0004] In summary, there is an urgent need for an automatic driving driving risk scene generation method and system suitable for strong interaction environment to solve the problems existing in the prior art. SUMMARY

[0005] The present application aims to provide an automatic driving driving risk scene generation method suitable for strong interaction environment, which aims to solve the problems of insufficient interaction feature modeling and limited scene authenticity of the existing vehicle motion interaction scene generation method. The specific technical solution is as follows:

[0006] An automatic driving driving risk scene generation method suitable for strong interaction environment, comprising:

[0007] S1, collecting road vehicle trajectory data to construct a basic database;

[0008] S2, extracting a target scene from the basic database, and calculating the vehicle motion interaction features of the target scene ;

[0009] S3, constructing a F-TimeGAN model based on the vehicle motion interaction features ;

[0010] S4, setting candidate hyperparameters of the F-TimeGAN model, and inputting the extracted target scene into the F-TimeGAN model under different candidate hyperparameters for scene generation;

[0011] S5, evaluating the similarity of the scene data generated by the F-TimeGAN model under different candidate hyperparameters and the target scene data, and selecting the top similar F-TimeGAN model as a candidate model, wherein is a natural number greater than or equal to 1;

[0012] S6, performing risk assessment on the scene library generated by the candidate model to obtain the number of risk scenes of each candidate model; if the number of risk scenes of the candidate model is greater than the target scene number, the F-TimeGAN model that can generate the most risk scenes is selected as the optimal model; if not, return to step S4 to reset the candidate hyperparameters of the F-TimeGAN model;

[0013] S7, screening strong interaction risk scenes in the scene data generated by the optimal model, and outputting the screening result to the automatic driving test scene library.

[0014] Preferably, the vehicle motion interaction feature is represented as:

[0015]

[0016]

[0017]

[0018]

[0019]

[0020] wherein, is the vector of the line connecting the head center points of the two vehicles, is the angle between the direction of the rear vehicle and is the angle between the direction of the front vehicle and represents the speed of the rear vehicle at time t, represents the speed of the front vehicle at time t, is the component of the speed of the rear vehicle at time t on the vector is the component of the speed of the front vehicle at time t on the vector represents the static safety distance parameter,​​​​​​​​​​​​​ is the safe headway, denotes the comfortable deceleration of the rear vehicle, is the maximum deceleration of the front vehicle, is the maximum deceleration of the rear vehicle, is the equilibrium distance, taken when the two vehicles are in adjacent lanes , taken when the two vehicles are in the same lane , denotes the actual distance between the two vehicles, is the speed difference sensitivity coefficient, is the desired speed of the rear vehicle, denotes the reaction time of the rear vehicle, is the depth of the potential well, is the dimensionless Mors potential parameter, is the base of the natural logarithm function.

[0021] Preferably, the vehicle motion interaction characteristics based on are used to determine the vehicle motion interaction state of the scene, specifically:

[0022]

[0023]

[0024]

[0025] wherein, denotes that the vehicle is in a strong interaction state at the moment, denotes that the vehicle is in a non-strong interaction state at the moment, is the vehicle motion interaction characteristic at the moment, is the threshold value of the strong interaction state and is less than 0, is the interaction exposure degree, denotes the total time in the strong interaction state in the scene, denotes the total time of the scene, denotes that the current scene is a strong interaction scene, denotes that the current scene is a non-strong interaction scene.

[0026] Preferably, an attention mechanism based on the vehicle motion interaction characteristics is introduced in the generator to construct the F-TimeGAN model, specifically:

[0027]

[0028] ​

[0029]

[0030]

[0031]

[0032]

[0033]

[0034] wherein, denotes the computation of attention weights, denotes the hidden state of the trajectory mapping of the host vehicle and other vehicles of 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, is the linear mapping parameter of the query vector in the attention mechanism module, is the linear mapping parameter of the key vector in the attention mechanism module, is the linear mapping parameter of the value vector in the attention mechanism module, denotes the dimension of the key vector , is the transpose matrix of the key vector , denotes the scaled dot-product attention; is the sensitivity adjustment parameter; weighted hidden state, is the tanh function value of the vehicle motion interaction feature ,

[0035] The generator of the F-TimeGAN model is represented as:

[0036]

[0037]

[0038] wherein, denotes the static feature latent encoding, denotes the generation network of the static feature, denotes the vector space defined by the known distribution, denotes the current time feature latent encoding, denotes the generation network for the time feature, denotes a random vector, This represents the latent encoding of the previous time feature.

[0039] Preferably, the target scene dataset is input into the F-TimeGAN model. Represented as:

[0040]

[0041]

[0042] in, The total number of samples, the sample Includes trajectory information of the main vehicle (Ego) and other vehicles (OV) and the vehicle motion interaction characteristics between them. , Indicates the trajectory of the main vehicle. Indicates the trajectories of other vehicles. express The location of the main vehicle at any given time. express The positions of other vehicles at that time express Real-time vehicle motion interaction characteristics between the main vehicle and other vehicles, set Each sample in the dataset has the same time step. , ;

[0043] The scene data set generated by the F-TimeGAN model is represented as follows:

[0044]

[0045]

[0046] in, The total number of samples generated by the F-TimeGAN model. The first generation of F-TimeGAN model One sample, The main vehicle trajectory generated by the F-TimeGAN model. Generated for F-TimeGAN model The location of the main vehicle at any given time. Other vehicle trajectories generated by the F-TimeGAN model, Generated for F-TimeGAN model m Locations of other vehicles at any given time, set Each generated sample has the same time step. , .

[0047] Preferably, the way of confirming the candidate model in step S5 is:

[0048] The F-TimeGAN model that generates data satisfying < and and is screened out, and then the comprehensive score of each screened F-TimeGAN model is calculated The screened F-TimeGAN models are arranged from low to high according to the comprehensive score , and the top F-TimeGAN models are selected as candidate models.

[0049] The comprehensive score is expressed as:

[0050]

[0051] Among them, is the FID score, is the prediction score, is the discrimination score, is the JS divergence value, , and are all set threshold values.

[0052] Preferably, the risk assessment of the scene library generated by the candidate model in step S6 is:

[0053] The driving risk of the scene is assessed by the stopping sight distance SSD and the stopping distance index SDI, and SSD and SDI are respectively expressed as:

[0054]

[0055] Among them, is the braking distance of the vehicle, represents the driver reaction distance, represents the speed of the vehicle at moment, represents the driver reaction time, is the deceleration;

[0056]

[0057] Among them, represents the vehicle distance between the front vehicle and the rear vehicle at moment, represents the stopping sight distance of the front vehicle Indicates the following vehicle is Sight distance while parking at any time Indicates the length of the vehicle in front;

[0058]

[0059] Among them, when 1 represents Time vehicle In a state of risk, 1 means that the scenario is considered a risky scenario, when 0 represents Time vehicle It is in a safe state.

[0060] Preferably, based on the collision risk coefficient Assess the risk level of the scenario, including the collision risk coefficient. Greater than the threshold The scenario is then considered a high-risk scenario; the collision risk coefficient is... Represented as:

[0061]

[0062]

[0063]

[0064] in, To assess the degree of risk exposure; This represents the total time the scenario is in a risky state. Indicates the total travel time for the scene; To assess the severity of the risk; To observe the maximum parking distance index; This is the theoretical maximum parking distance index.

[0065] The present invention also provides an autonomous driving risk scenario generation system suitable for highly interactive environments. The system includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the method described thereon when running the computer program.

[0066] The present invention also provides a storage medium storing a computer program, which, when run, executes the method.

[0067] The application of the technical solution of the present invention has the following beneficial effects:

[0068] The generation method of the application first collects data of a target scene such as following or lane changing, and then improves the TimeGAN model by combining the vehicle motion interaction feature F of the target scene with the attention mechanism to obtain an F-TimeGAN model, inputs the target scene into the F-TimeGAN model for scene generation, and finally selects a strong interaction risk scene from the generated scene for automatic driving test. The method of the application effectively improves the interaction and risk of the generated scene. The method of the application can efficiently generate strong interaction risk scene data that is difficult to collect in reality based on conventional target scene data, effectively solves the problem of lack of key test scene data, and provides scene data support for safety verification of automatic driving in a strong interaction environment.

[0069] The method of the application solves the problem that the existing method only relies on basic parameters such as vehicle distance, relative speed, and acceleration change to describe the interaction relationship between the host vehicle and the surrounding vehicles, ignores the influence of key factors such as heading angle dynamic adjustment and vehicle dynamics difference on the interaction intensity, and the existing method mainly generates a static trajectory or motion parameter, and lacks consideration of the vehicle motion interaction process in the scene. The method of the application defines a vehicle motion interaction intensity evaluation index by considering the vehicle driving speed, heading angle, and vehicle type difference, and further fuses the vehicle motion interaction feature into the attention mechanism to improve the generation model. The method of the application is more suitable for strong interaction scenes, and improves the interaction, reality, and danger of the scene.

[0070] In addition to the purposes, features and advantages described above, the application has other purposes, features and advantages. The application will be further described below with reference to the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0071] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application, illustrate the preferred embodiments of the application, and assist in the explanation of the application. In the drawings:

[0072] Figure 1 is a flowchart of the automatic driving risk scene generation method of the application;

[0073] Figure 2 is a balance distance schematic diagram of a vehicle adjacent lane and a same lane;

[0074] Figure 3 is a PCA visualization schematic diagram of F-TimeGAN model generated data;

[0075] Figure 4 is a t-SNE visualization schematic diagram of F-TimeGAN model generated data;

[0076] Figure 5is a collision risk coefficient distribution diagram of the whole sample;

[0077] Figure 6 is Figure 5 is a sample distribution diagram of the collision risk coefficient greater than or equal to 0.1;

[0078] Figure 7 is a sample interaction exposure degree distribution diagram. DETAILED DESCRIPTION

[0079] In order to facilitate the understanding of the present application, the present application will be described more fully below, and preferred embodiments of the present application will be given. However, the present application can be realized in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0080] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used in the specification of the present application herein are only for the purpose of describing specific embodiments and are not intended to limit the present application.

[0081] Embodiment 1:

[0082] Referring to Figure 1 , the present embodiment provides an automatic driving driving risk scene generation method suitable for a strong interaction environment, which is specifically as follows:

[0083] S1, collecting road vehicle trajectory data to construct a basic database;

[0084] Preferably, in the present embodiment, the vehicle trajectory data is obtained based on unmanned aerial vehicle aerial photography, vehicle-mounted sensors (such as GPS / IMU combined navigation system, visual camera) and real vehicle experiment platform, and the fields of the basic database constructed must include the following core fields: time stamp , two-dimensional plane position of vehicle trajectory , vehicle length , vehicle width and lane where the vehicle is located .

[0085] Of course, in some embodiments, other ways of collecting vehicle trajectory data can also be used, and the possible ways used in the present embodiment are not listed one by one.

[0086] S2, extracting a target scene from the basic database and calculating the vehicle motion interaction features of the target scene ;

[0087] Preferably, the data in the base database is first cleaned of noise in this embodiment, and then the target scene is extracted by analyzing the vehicle motion characteristics. In this embodiment, the noise cleaning is performed by a statistical filtering algorithm based on a sliding window to eliminate abnormal trajectory points, and by coordinate conversion to eliminate spatial bias, thereby ensuring data continuity and reliability.

[0088] Further, the vehicle motion characteristics specifically include: static attributes, dynamic parameters, relative distance of the host vehicle and other vehicles, and relative speed of the host vehicle and other vehicles, wherein the static attributes include vehicle type, vehicle length and vehicle width, and the dynamic parameters include vehicle speed, vehicle acceleration and vehicle heading angle (i.e. vehicle driving direction).

[0089] Further, the target scene in this embodiment includes a following scene and / or a lane changing scene, which is specifically identified by the interaction relationship between the host vehicle trajectory and the trajectories of other vehicles in space and time.

[0090] In this embodiment, the lane changing scene sample is required to meet: a) the longitudinal relative distance between the rear vehicle and the lane changing vehicle is not more than 120m; b) only the lane changing vehicle has a lane changing behavior in the extracted sample, and the rear vehicle cut in is driving in the lane; c) the lane changing time step needs to be greater than 3 seconds.

[0091] In this embodiment, the following scene sample is required to meet: a) determine the front and rear two vehicles in the same lane, and neither of the two vehicles has lane changing behavior in the following process; b) the distance between the following vehicles needs to be less than 125 meters to ensure that they are not in a free flow state; c) the duration of the following needs to be greater than 10s.

[0092] Of course, the requirements for the following scene and the lane changing scene may be different in some embodiments, and those skilled in the art can flexibly adjust and modify the judgment requirements for the following scene and the lane changing scene.

[0093] The vehicle motion interaction characteristics in this embodiment will be described in detail as follows:

[0094] The Morse potential function The derivative The interaction force is calculated as follows:

[0095]

[0096]

[0097] wherein, represents the distance between atoms, is the depth of the potential well, is the dimensionless Morse potential parameter, ​Indicates the equilibrium distance. It is the base of the natural logarithm function.

[0098] See Figure 2 In this embodiment, considering vehicle heading angle, speed, and vehicle model differences, the Morse potential function is modified to construct vehicle motion interaction features suitable for highly interactive scenarios. In this context, when vehicles and interacting vehicles are in adjacent lanes, the Gipps model is used as the safety distance model, which takes the minimum safe lane-changing distance as the expected spacing. When vehicles and interacting vehicles are in the same lane, the safe distance is calculated based on the IDM model. .when 0 indicates that there is a repulsive force between the two vehicles. A value of 0 indicates that there is attraction between the two vehicles, and the vehicle motion interaction characteristics are as follows. Specifically, it is expressed as follows:

[0099]

[0100]

[0101]

[0102]

[0103]

[0104] in, Let be the vector of the line connecting the center points of the front ends of the two vehicles. The direction the following vehicle is traveling and The included angle, The direction the vehicle in front is traveling and The included angle, express The speed of the car behind at all times. express Always keep track of the speed of the car in front. for Speed ​​of the car behind In vector The amount on, for Speed ​​of the vehicle in front In vector The amount on, This represents the stationary safety distance parameter. For safe headway, Indicates the comfortable deceleration of the following vehicle. The maximum deceleration of the vehicle in front. The maximum deceleration of the following vehicle. To balance the distance, when two vehicles are in adjacent lanes, the Gipps model is used to determine the distance. When two vehicles are in the same lane, the IDM model is used to determine the position. , This indicates the actual distance between the two workshops. For speed difference sensitivity coefficient, For the expected speed of the following vehicle, Indicates the reaction time of the following vehicle. Let the depth of the potential well be . It is a dimensionless Morse potential parameter. It is the base of the natural logarithm function.

[0105] Furthermore, based on vehicle motion interaction features This allows us to determine whether a scene is a highly interactive scene, specifically:

[0106] In this embodiment, High Interaction (HI) and Interaction Exposure Level (IEL) are defined to determine the interaction state of the scene, and the specific calculation is as follows:

[0107]

[0108]

[0109]

[0110] in, express Whether the vehicle is in a highly interactive state at any given time. This indicates that the vehicle is in a highly interactive state. This indicates that the vehicle is in a non-strong interaction state. for Real-time vehicle motion interaction characteristics, The threshold for strong interaction state and If the value is less than 0, the kernel density estimation method is used to plot the graph. The distribution curve is then used to determine the threshold through peak detection; The value range is from 0 to 1. This represents the total time spent in a highly interactive state within the scene. Indicates the total time of the scene. When hour, This indicates that the vehicle is in a highly interactive state for half of the driving time, and the scenario is considered a High Interaction Scenario (HIS).

[0111] So far, the target scene data set composed of vehicle trajectory information and vehicle motion interaction features of vehicles Can be expressed as:

[0112]

[0113]

[0114] Wherein, is the total number of samples, and the sample contains the trajectory information of the ego vehicle and other vehicles OV and the vehicle motion interaction features between the two , represents the trajectory of the ego vehicle, represents the trajectory of the other vehicle, represents the ego vehicle position at time t, represents the other vehicle position at time t, represents the vehicle motion interaction features between the ego vehicle and the other vehicle at time t, and each sample has the same time step , .

[0115] S3, based on vehicle motion interaction features Construct F-TimeGAN model;

[0116] The F-TimeGAN model in this embodiment is similar in network structure to the TimeGAN model, and mainly includes the following modules: encoder Embedder, restorer Recovery, supervisor Supervisor, generator Generator and discriminator Discriminator.

[0117] This embodiment is based on the TimeGAN model (i.e. time series generative adversarial network), introduces an attention mechanism based on vehicle motion interaction features in the generator, and improves the RNN model of the generator. In order to reduce the difference in numerical scale of part of the long-tail interaction force value and improve the stability of model training, the vehicle motion interaction features are nonlinearly mapped by using the tanh function to compress them between . When is closer to 1, it means that the attractive interaction is stronger, and when is closer to -1, it means that the repulsive interaction is stronger. This method realizes the enhancement of the dependence of the vehicle's own trajectory in time series modeling, and the attention weight is Value linear bias, so that the generator pays more attention to larger time steps, thus exploiting the vehicle motion interaction features effectively improve the spatial interaction modeling capability between vehicles, which can be mathematically expressed as follows:

[0118]

[0119]

[0120]

[0121]

[0122]

[0123]

[0124]

[0125] wherein, represents the calculation of attention weight, represents the hidden state of the trajectory mapping of the host vehicle and other vehicles of 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, is the linear mapping parameter of the query vector in the attention mechanism module, is the linear mapping parameter of the key vector in the attention mechanism module, is the linear mapping parameter of the value vector in the attention mechanism module, represents the dimension of the key vector , is the transpose matrix of the key vector , represents the scaled dot-product attention; is a sensitivity adjustment parameter that adjusts the weight of the interaction feature F; weighted hidden state, finally is mapped to the output of the generator through a fully connected layer.

[0126] The original generator of the TimeGAN model can be represented as , The generator of the F-TimeGAN model improved by the attention mechanism in the embodiment can be represented as:

[0127]

[0128]

[0129] wherein, represents a static feature latent code, represents a generation network of the static feature, represents a vector space defined with a known distribution, represents a current time feature latent code, represents a generation network for the time feature, represents a random vector, represents a previous time feature latent code.

[0130] S4, set F-TimeGAN model candidate hyperparameters, input the extracted target scene to the F-TimeGAN model under different candidate hyperparameters for scene generation;

[0131] Specifically, a person skilled in the art can set multiple groups of candidate hyperparameters according to actual conditions, and input the extracted target scene data set to the F-TimeGAN model under different candidate hyperparameters for scene generation, so as to achieve the purpose of screening the optimal hyperparameter F-TimeGAN model.

[0132] The setting of the candidate hyperparameters is common sense in the art, so step S4 will not be described in detail in this embodiment.

[0133] Specifically, the scene data set generated by the F-TimeGAN model is represented as:

[0134]

[0135]

[0136] wherein, is the total number of samples generated by the F-TimeGAN model, is the th sample generated by the F-TimeGAN model, is the main vehicle trajectory generated by the F-TimeGAN model, is the main vehicle position at the time generated by the F-TimeGAN model, is the other vehicle position trajectory generated by the F-TimeGAN model, is the other vehicle position at the time generated by the F-TimeGAN model, each generated sample has the same time step m ​​, .

[0137] It should be noted that, As the output data of the F-TimeGAN model, it only includes the trajectories of the host vehicle and other vehicles, and does not include the vehicle motion interaction features between the two vehicles.

[0138] S5, evaluate the similarity of the scene data generated by the F-TimeGAN model under different candidate hyperparameters and the target scene data, and select the top F-TimeGAN model as a candidate model, wherein is a natural number greater than or equal to 1;

[0139] In this embodiment, four indicators, namely Fréchet Inception Distance (FID), Prediction Score (PS), Discriminative Score (DS), and Jensen–Shannon divergence (JS), are selected to evaluate the quality of the data generated by each F-TimeGAN model from multiple angles (that is, to judge the similarity between the generated data and the real data), and specifically:

[0140] First, filter out the F-TimeGAN model whose generated data satisfies < and and Then, calculate the comprehensive score of each F-TimeGAN model filtered out, arrange the filtered F-TimeGAN models in ascending order of comprehensive score , and select the top F-TimeGAN model as a candidate model.

[0141] Among them, the comprehensive score is used to represent 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 (i.e., the real data), and the comprehensive score is expressed as:

[0142] (24),

[0143] Among them, is the FID score, is the prediction score, is the discriminative score, For JS divergence value, the above four indexes are used to compare the similarity of generated data and real data , and evaluate the timing of generated data ; , and are set thresholds, and in the embodiment, the values are , , .

[0144] S6, risk assessment is performed on the scene library generated by the candidate model to obtain the number of risk scenes of each candidate model; if there is a candidate model whose number of risk scenes is greater than the target number of scenes, the F-TimeGAN model that can generate the most risk scenes is selected as the optimal model; if not, return to step S4 to reset the candidate hyperparameters of the F-TimeGAN model;

[0145] Specifically, in the embodiment, risk assessment is performed on the scene library generated by the candidate model to achieve the purpose of identifying risk scenes, and the risk assessment method is as follows:

[0146] Stopping Sight Distance (SSD) (unit: m) and stopping distance index (SDI) (unit: m) are used to evaluate the driving risk of the scene, and the specific calculation methods of SSD and SDI are as follows:

[0147]

[0148] wherein, is the vehicle braking distance (unit: m), that is, the distance traveled by the driver from taking braking measures to the vehicle completely stopping, represents the driver reaction distance (unit: m), represents the speed (unit: km / h) of the vehicle at time , the vehicle refers to the host vehicle or other vehicles (i.e. the front or rear vehicle); represents the driver reaction time (unit: s), which is generally 2.5 seconds. is the deceleration (unit: ), the deceleration of a truck is 2.4 , and the deceleration of a car is 3.4 .

[0149]

[0150] wherein, represents The distance between the front and rear of the vehicle at any given time (unit: m). Indicates that the car in front is Sight distance for parking at any given time (unit: m). Indicates the following vehicle is Sight distance for parking at any given time (unit: m). This indicates the length of the vehicle in front (unit: m), i.e. It indicates the distance between the rear of the vehicle in front and the front of the vehicle behind (unit: m);

[0151] when 0 indicates that when the vehicle in front suddenly stops, the vehicle behind cannot brake safely, resulting in insufficient visibility and a potential danger; conversely, a higher number indicates safety.

[0152]

[0153] in, for Time vehicle Is it in a dangerous situation? 1 indicates vehicle Being in a risky state, meaning the scenario is a risky scenario, when 0 indicates a vehicle It is in a safe state;

[0154] Risk scenarios in the candidate model generation scenario library can be filtered out using formulas (25) and (27). In this embodiment, a collision risk coefficient is further used. The risk coefficient comprehensively reflects the duration and severity of risk exposure in a risk scenario, thus determining the level of risk in that scenario.

[0155]

[0156]

[0157]

[0158] in, The level of risk exposure is represented by a value ranging from 0 to 1. This represents the total time of unsafe states in the scenario (i.e. (total time) Indicates the total travel time for the scene; The value ranges from 0 to 1, representing the severity of the risk. To observe the maximum parking distance index; This is the theoretical maximum parking distance index. The value is 565m.

[0159] In this embodiment, the collision risk coefficient greater than the threshold value is regarded as a high-risk scenario, and the threshold value is generally 0.3; of course, those skilled in the art can flexibly adjust the value of the threshold value .

[0160] Specifically, in step S6, when confirming the optimal model, it is required that the number of risk scenarios of the candidate model is greater than the number of target scenarios, and only then can the F-TimeGAN model generating the most risk scenarios be regarded as the optimal model. The reason is that only when the number of risk scenarios contained in the generated data of the candidate model is greater than the number of target scenarios, it means that the scene generation of the candidate model has practical significance.

[0161] Specifically, the scene library generated by the optimal model is denoted as :

[0162]

[0163] wherein, N is the total number of samples generated by the optimal model, and xi is the i-th sample generated by the optimal model.

[0164] S7, screening strong interaction risk scenarios in the scene data generated by the optimal model, and outputting the screening results to the automatic driving test scenario library.

[0165] Further, according to formulas (25)-(27), the risk scenarios in the scene library generated by the optimal model can be screened, and according to formulas (3)-(10), the interaction state of the risk scenarios generated by the optimal model can be judged. Specifically, if the danger index , it is marked as a risk scenario; if the strong interaction scenario index , it is marked as a strong interaction scenario; if a certain scenario and , it is marked as a strong interaction risk scenario. In the scene data generated by the optimal model, the scenarios satisfying and are screened as strong interaction risk scenarios, and all the screened strong interaction risk scenarios are output to the automatic driving test scenario library for automatic driving test.

[0166] ​​​The generation method of the embodiment first collects data of target scenes such as following and lane changing, then improves the TimeGAN model by combining the vehicle motion interaction feature F of the target scene with the attention mechanism to obtain an F-TimeGAN model, inputs the target scene into the F-TimeGAN model for scene generation, and finally screens strong interaction risk scenes from the generated scenes. The method of the embodiment effectively improves the interaction and risk of the generated scenes, generates strong interaction risk scene data based on conventional target scene data, solves the problem that strong interaction risk scene data is difficult to collect in practice, and provides scene data support for safety verification of autonomous driving in a strong interaction environment.

[0167] Embodiment 2

[0168] In this embodiment, the public data set is obtained, and the HighD open data set is taken as an example. The data is collected in a bidirectional six-lane scene (25th-60th file), and the method in embodiment 1 is used for scene generation. The basic situation of the bidirectional six-lane scene data of the HighD open data set is shown in Table 1.

[0169] Table 1 Basic database information

[0170]

[0171] In this example, the lane changing scene is taken as the target scene. The lane changing scene extraction rule is to take 49 frames forward and 50 frames backward as the center of the lane change point to form a 100-frame lane changing sequence. In addition, the following conditions also need to be met:

[0172] a) The longitudinal relative distance between the rear vehicle and the lane changing vehicle is not more than 120m;

[0173] b) Only the lane changing vehicle has a lane changing behavior in the extracted sample, and the rear vehicle cut into the lane is driving in the lane;

[0174] c) The lane changing time step needs to be unified, and each sample is 100 frames, that is, =100.

[0175] Statistical analysis is performed on the risk of the lane changing scene in the basic database. The interaction of different vehicle types is divided into car-cut-in-car scene (C-cut in-C), car-cut-in-truck scene (C-cut in-T), truck-cut-in-car scene (T-cut in-C), and truck-cut-in-truck scene (T-cut in-T). The stopping sight distance (SSD) (unit: m) and the stopping distance index (SDI) (unit: m) are used to evaluate the driving risk of the scene. When =1, the scenario is a risk scenario, and the statistical results are shown in Table 2:

[0176] Table 2 Risk sample size statistics of different categories of scenarios

[0177]

[0178] Although the T-cut in-C, T-cut in-T and other scenarios involving truck lane changing have a small number of samples in natural driving data, but their risk proportion is high, indicating that truck lane changing is more likely to cause potential danger. As can be seen from Table 2, the risk proportion of T-cut in-C and T-cut in-T scenarios is 80.69% and 47.90% respectively, reflecting that trucks have stronger risk potential when changing lanes. Among them, the total number of samples of the C-cut in-T scenario ranks second, and the risk sample proportion is the lowest. This may be because under the condition of natural driving, when a car cuts into the front of a truck, the driver will consciously maintain a larger longitudinal safety distance from the truck, thereby reducing the risk level. From the driving experience, there is a certain correlation between vehicle motion interaction and driving risk, and it is inferred that the interaction of the car cutting into the truck scenario is also weak. Obviously, it is difficult to obtain a strong interaction and dangerous C-cut in-T scenario on a large scale in real scenarios, and automatic driving lacks the risk perception of human drivers, and therefore needs to test this kind of rare risk scenario, so the C-cut in-T scenario is selected as an example in this embodiment to generate a scenario to enrich the scenario library of automatic driving test.

[0179] The undetermined parameters are calibrated according to the related calculation of the vehicle motion interaction feature F in Embodiment 1, wherein the calibration results of the undetermined parameters are shown in Table 3. The vehicle motion feature F is calculated using the calibrated parameters, the kernel density estimation method is used to draw the distribution curve of the vehicle motion interaction feature F, and then the threshold value is determined by peak detection is -0.68, that is is a strong interaction state, otherwise it is a weak interaction. Finally, the target scenario data set composed of vehicle trajectory information and vehicle motion interaction feature is obtained. .

[0180] Table 3 Calibrated parameters

[0181]

[0182] The F-TimeGAN model is constructed according to the method of embodiment 1, candidate hyperparameters of the F-TimeGAN model are set, and the extracted lane changing scene data is input into the F-TimeGAN model under different candidate hyperparameters for scene generation; wherein each lane changing sample has the same time step of 100 frames, and the candidate hyperparameters set by the F-TimeGAN model are as shown in Table 4:

[0183] Table 4 F-TimeGAN model parameter settings

[0184]

[0185] The scene library generated by the optimal model is obtained through steps S5 and S6 in embodiment 1 . As shown in Table 5, through comparative analysis of quantitative evaluation indicators (including discrimination score, prediction score, JS divergence and FID score), it is shown that the F-TimeGAN model is significantly better than the benchmark TimeGAN model in terms of the similarity of generated data and real data, and from the evaluation result of the prediction score, the data generated by the method in embodiment 1 has stronger time sequence.

[0186] Table 5 F-TimeGAN model and TimeGAN model result comparison table

[0187]

[0188] The generated data of the F-TimeGAN model is reduced and visualized through PCA and t-SNE algorithms, as shown in Figure 3 and Figure 4 , the red points in the figure represent real data, and the blue points represent generated data, it can be found from Figure 3 and Figure 4 that the generated data of 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. As shown in Figure 3 and Figure 4 , it can be found that the generated data of the F-TimeGAN model and the original features have a moderate distribution offset, which shows that the method enhances the diversity of the generated data.

[0189] Further, the strong interaction scene, the risk scene and the strong interaction risk scene are respectively screened out from the generated data of the optimal F-TimeGAN model and the TimeGAN model, and the results are shown in Table 6. From the number of risk samples, compared with the real data and the unimproved TimeGAN model, the F-TimeGAN model generates the most risk scenes of 341, which is 3.79 times of the real data and 2.66 times of the TimeGAN model. Further comparison of the risk coefficient, Figure 5As shown, the risk coefficients of a large number of data samples are less than 0.1. After extracting the samples with risk coefficients greater than or equal to 0.1, it is found that Figure 6 It can be seen that the risk coefficients of the data generated by the F-TimeGAN model are higher. In the real data, there are only 2 scenarios with risk coefficients greater than 0.3, while in the data generated by the TimeGAN model, there are 3 scenarios with risk coefficients greater than 0.3. In contrast, the F-TimeGAN model generates 11 scenarios with risk coefficients greater than 0.3, which is 5.5 times higher than the real data and 3.7 times higher than the unimproved TimeGAN model.

[0190] Table 6 Risk and interaction sample analysis

[0191]

[0192] By calculating the lane-changing interaction degree (IEL), it can be seen from Table 6 and Figure 7 that the interaction degree of the samples generated by the F-TimeGAN model is stronger. Specifically, there are only 183 samples with IEL>0.5 in the real data, 230 in the data generated by the TimeGAN model, and 444 in the data generated by the F-TimeGAN model. The F-TimeGAN model generates strong interaction degree lane-changing samples, which is 2.42 times higher than the real data and 1.93 times higher than the unimproved TimeGAN model.

[0193] Further, in order to avoid the randomness of analyzing a single experiment, a large number of experiments were conducted on models with different parameters in this embodiment, and the results are shown in Tables 7 to 9. In contrast, the F-TimeGAN model has a higher coverage, and the different numerical meanings in the data source are "model-hidden unit-number of layers-batch size-sensitivity coefficient", where 1 represents coverage and 0 represents non-coverage.

[0194] Table 7 Two-car initial relative position of different models

[0195]

[0196] Table 8 Lane-changing vehicle horizontal displacement of different models

[0197]

[0198] Table 9 Rear vehicle horizontal displacement of different models

[0199]

[0200] It can be found from Tables 7-9 that the data generated by the TimeGAN model under some parameter combinations is not reasonable, which reflects that the trajectory jitter generated by the model is relatively large, and a large number of training and parameter adjustment times are needed to search for the TimeGAN model. In contrast, the F-TimeGAN model can better learn the vehicle motion characteristics, the trajectory jitter is smaller and more stable, and is more similar to the real data.

[0201] Embodiment 3

[0202] The embodiment provides an automatic driving driving risk scene generation system suitable for a strong interaction environment, the system comprising a processor and a memory, the memory storing a computer program, and the processor executing the computer program to perform the method in embodiment 1.

[0203] The system of the embodiment can efficiently generate strong interaction risk scene data which is difficult to collect in reality based on conventional target scene data, effectively solves the problem of lack of key test scene data, and provides scene data support for safety verification of automatic driving in a strong interaction environment.

[0204] Embodiment 4

[0205] The embodiment provides a storage medium, the storage medium storing a computer program, and the computer program being executed to perform the method in embodiment 1.

[0206] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for generating autonomous driving risk scenarios suitable for highly interactive environments, 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 features of the target scene. ; S3, based on vehicle motion interaction features Construct an F-TimeGAN model; S4. Set candidate hyperparameters for 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 model with the highest similarity. Several F-TimeGAN models were selected as candidate models, among which It is a natural number greater than or equal to 1; S6. Conduct a risk assessment on the scenario library generated by the candidate models to obtain the number of risk scenarios for each candidate model; If the number of risky scenarios for candidate models exceeds the number of target scenarios, the F-TimeGAN model that generates the most risky scenarios will be selected as the optimal model; otherwise, 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 filtering results to the autonomous driving test scenario library; Introducing vehicle motion interaction features into the generator The attention mechanism used to construct the F-TimeGAN model is as follows: in, This indicates the calculation of attention weights. This represents the hidden state of the trajectory mapping between the master vehicle and other vehicles in the original time series data. This is the query vector in the attention mechanism module. This refers to the key vector in the attention mechanism module. This is the value vector in the attention mechanism module. For the query vector in the attention mechanism module The linear mapping parameters, Key vectors in the attention mechanism module The linear mapping parameters, The value vector in the attention mechanism module The linear mapping parameters, Represents the key vector Dimensions Key vector The transpose of the matrix, This indicates scaling dot product attention; This is a sensitivity adjustment parameter; Weighted hidden states, Vehicle motion interaction features The value of the tanh function; The generator of the F-TimeGAN model is represented as follows: in, Represents the latent encoding of static features. Generative networks representing static features This represents a vector space with a known distribution. This represents the latent encoding of current time features. This represents the generative network used for temporal features. Represents a random vector. This represents the latent encoding of the previous time feature.

2. The method for generating autonomous driving risk scenarios suitable for highly interactive environments according to claim 1, characterized in that, The vehicle motion interaction features Represented as: in, Let be the vector of the line connecting the center points of the front ends of the two vehicles. The direction the following vehicle is traveling and The included angle, The direction the vehicle in front is traveling and The included angle, express The speed of the car behind at all times. express Always keep track of the speed of the car in front. for Speed ​​of the car behind In vector The amount on, for Speed ​​of the vehicle in front In vector The amount on, This represents the stationary safety distance parameter. For safe headway, Indicates the comfortable deceleration of the following vehicle. The maximum deceleration of the vehicle in front. The maximum deceleration of the following vehicle. To maintain a balanced distance, when two vehicles are in adjacent lanes, take... When two vehicles are in the same lane, take , This indicates the actual distance between the two workshops. For speed difference sensitivity coefficient, For the expected speed of the following vehicle, Indicates the reaction time of the following vehicle. Let be the depth of the potential well. It is a dimensionless Morse potential parameter. It is the base of the natural logarithm function.

3. The method for generating autonomous driving risk scenarios suitable for highly interactive environments according to claim 2, characterized in that, Based on vehicle motion interaction features Determine the vehicle's motion interaction state in the scene, specifically as follows: in, express The vehicle is always in a highly interactive state. express The vehicle is in a non-interactive state at all times. for Real-time vehicle motion interaction characteristics, The threshold for strong interaction state and Less than 0, For the degree of interactive exposure, This represents the total time spent in a highly interactive state within the scene. Indicates the total time of the scene. This indicates that the current scenario is a highly interactive scenario. This indicates that the current scenario is a non-strong interaction scenario.

4. The method for generating autonomous driving risk scenarios suitable for highly interactive environments according to claim 1, characterized in that, Input the target scene dataset into the F-TimeGAN model Represented as: in, The total number of samples, the sample Includes trajectory information of the main vehicle (Ego) and other vehicles (OV) and the vehicle motion interaction characteristics between them. , Indicates the trajectory of the main vehicle. Indicates the trajectories of other vehicles. express The location of the main vehicle at any given time. express The positions of other vehicles at that time express Real-time vehicle motion interaction characteristics between the main vehicle and other vehicles, set Each sample in the dataset has the same time step. , ; The scene data set generated by the F-TimeGAN model is represented as follows: in, The total number of samples generated by the F-TimeGAN model. The first generation of F-TimeGAN model One sample, The main vehicle trajectory generated by the F-TimeGAN model. Generated for F-TimeGAN model The location of the main vehicle at any given time. Other vehicle trajectories generated by the F-TimeGAN model, Generated for F-TimeGAN model m Locations of other vehicles at any given time, set Each generated sample has the same time step. , .

5. The method for generating autonomous driving risk scenarios suitable for highly interactive environments according to claim 1, characterized in that, The method for confirming the candidate model in step S5 is as follows: Filter out the generated data that meet the requirements < and and The F-TimeGAN model was selected, and then its comprehensive score was calculated for each selected F-TimeGAN model. The selected F-TimeGAN models are ranked according to their comprehensive scores. Arrange from lowest to highest, and select the ones at the top. One F-TimeGAN model was selected as the alternative model; Among them, the overall score Represented as: in, For FID score, To predict scores, To determine the score, The JS divergence value. , and All of these are set thresholds.

6. The method for generating autonomous driving risk scenarios suitable for highly interactive environments according to claim 1, characterized in that, Step S6 involves a risk assessment of the scenario library generated from the candidate models, specifically: The driving risk assessment of the scenario is performed using the parking sight distance (SSD) and parking distance index (SDI). SSD and SDI are respectively represented as: in, This refers to the vehicle's braking distance. Indicates the driver's reaction distance. Indicates vehicle exist The speed of time, Indicates the driver's reaction time. For deceleration; in, express The distance between the front and rear of the vehicle at all times. Indicates that the car in front is Sight distance while parking at any time Indicates the following vehicle is Sight distance while parking at any time Indicates the length of the vehicle in front; Among them, when express Time vehicle In a state of risk, That is, the scenario is considered a risk scenario, when express Time vehicle It is in a safe state.

7. The method for generating autonomous driving risk scenarios suitable for highly interactive environments according to claim 6, characterized in that, Collision risk coefficient Assess the risk level of the scenario, collision risk coefficient Greater than the threshold The scenario is then considered a high-risk scenario; the collision risk coefficient is... Represented as: in, To assess the degree of risk exposure; This represents the total time the scenario is in a risky state. Indicates the total travel time for the scene; To assess the severity of the risk; To observe the maximum parking distance index; This is the theoretical maximum parking distance index.

8. A system for generating autonomous driving risk scenarios suitable for highly interactive environments, characterized in that, The system includes a processor and a memory, the memory storing a computer program, which the processor executes when running the computer program, performing the method as described in any one of claims 1-7.

9. A storage medium, characterized in that, The storage medium stores a computer program, which, when run, executes the method as described in any one of claims 1-7.

Images