Adas test scene automatic generation method and system based on parameterized preset risk field

Abstract

The application discloses an ADAS test scene automatic generation method and system based on a parameterized preset risk field, and particularly relates to the technical field of automatic driving simulation testing, and comprises the following steps: a mixed risk field model is constructed, wherein the model at least comprises a static risk field and a dynamic risk field; risk field physical parameters are extracted from real accident data through a physical feature extraction channel to construct a static risk field conforming to real risk distribution; a semantic large model analysis channel is used to map natural language instructions into risk field traffic parameters; a double-channel mixed mode of the physical feature extraction channel and the semantic large model analysis channel is used to generate a mixed GAN generative adversarial network for generating edge cases to construct a dynamic risk field; a scene element reverse generation module is designed, a three-dimensional engine is driven according to risk field physical parameters and risk field traffic parameters to generate an ADAS test scene containing a road network, traffic participants and environmental variables, and directional reinforcement generation of a high-risk scene is realized.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving simulation testing technology, specifically to a method and system for automatically generating ADAS test scenarios based on parameterized preset risk fields. Background Technology

[0002] As autonomous driving technology advances towards Level 3+, Expected Functional Safety (SOTIF, ISO 21448) has become a core challenge in safety testing. SOTIF requires that a system, even in a fault-free state, can still cope with unknown unsafe scenarios and foreseeable misuse. Traditional simulation testing methods face the following bottlenecks in meeting SOTIF requirements:

[0003] 1. The separation between scenario generation and risk perception:

[0004] Current mainstream tools (such as Prescan and CARLA) adopt a sequential process of "build the scenario first and then calculate the risk", which causes the risk field calculation to lag behind the scenario construction. This makes it impossible to establish a strong correlation between scenario parameters and risk intensity, and makes it difficult to systematically cover the potential risk triggering conditions required by ISO21448 (such as misjudgment caused by sudden changes in road curvature).

[0005] 2. Low efficiency in generating long-tail scenes:

[0006] Traditional scene generation methods based on real road survey data consume 99% of computing power in reconstructing regular scenes, with only 1% used for edge case generation (McKinsey 2023). Their coverage of low-probability, high-hazard scenarios (such as a child suddenly crossing a foggy road) is less than 0.03% (NHTSA 2022), failing to meet SOTIF's requirement for "exploration of unknown risks." Random scene generation methods lack risk orientation; only 0.7% of the generated 10^5 scenarios effectively trigger unconventional system responses (Waymo report).

[0007] 3. Lack of dynamic interaction risk modeling:

[0008] Existing risk field models (such as Gaussian mixture models) only consider static obstacles and ignore the spatiotemporal dynamic influences among traffic participants. They cannot simulate the irrational behavior of human drivers (such as aggressive lane changes), which are precisely the typical foreseeable misuse scenarios defined by ISO 21448.

[0009] 4. Limitations of known scene generation solutions:

[0010] Rule-based methods: Manually defining scenario logic trees, limited by the designer's professional knowledge, cannot exhaustively list parameter combinations, and suffer from a severe lack of case studies when dealing with long-tail problems.

[0011] Reinforcement learning generation methods: The black box nature of the generated scenes leads to a lack of interpretability, making it difficult to meet SOTIF's requirements for scene traceability and hindering the development of ADAS-related algorithms.

[0012] Digital twin reconstruction technology relies on high-precision maps and sensor data, but cannot generate fictional scenarios of unrecorded or future traffic environments. Summary of the Invention

[0013] To address this issue, the present invention provides a method and system for automatically generating ADAS test scenarios based on a parameterized preset risk field, thereby resolving the problems mentioned in the background art.

[0014] To achieve the above objectives, the present invention provides the following technical solution: an automatic generation method for ADAS test scenarios based on a parameterized preset risk field, which constructs a hybrid risk field model, wherein the model includes at least a static risk field and a dynamic risk field; extracts physical parameters of the risk field from real accident data through a physical feature extraction channel to construct a static risk field that conforms to the real risk distribution; maps natural language instructions to risk field traffic parameters through a semantic large model parsing channel; and generates a hybrid GAN generative adversarial network for generating edge cases using a dual-channel hybrid approach of physical feature extraction channel and semantic large model parsing channel to construct a dynamic risk field.

[0015] The design of the scene element reverse generation module drives the 3D engine to generate ADAS test scenes containing road networks, traffic participants and environmental variables based on the physical parameters and traffic parameters of the risk field, thereby realizing the targeted enhancement generation of high-risk scenarios.

[0016] Preferably, the extracted physical parameters of the risk field include road network, obstacle distribution, traffic flow dynamics, and environmental variables; environmental variables include weather parameters and lighting conditions.

[0017] Preferably, the static risk field describes the risk distribution of static elements, including road network generation and static obstacle generation.

[0018] Preferably, the static risk field modeling for the generated road network is as follows:

[0019] ;

[0020] in The total number of obstacles; For the first The coordinates of the center of each obstacle; This is the risk intensity coefficient, which represents the inherent danger level of an obstacle. For example, the value is 1.2±0.3 for small passenger cars and 0.5±0.1 for pedestrians (according to the NHTSA accident injury classification). The risk attenuation radius reflects the range of impact from obstacles and is related to the vehicle's condition and speed. This applies to small passenger vehicles in normal conditions (good environmental conditions and vehicle condition). The coefficient value is related to the braking sight distance; the smaller the value, the more likely the vehicle is to perceive the risk and complete the braking in a short time.

[0021] Preferably, static risk field modeling for static obstacles includes the following cases:

[0022] (1) The gradient risk generated by superimposing the non-point structure of the road, which includes road markings and edges, is modeled as follows:

[0023] ;

[0024] in, This represents the total number of road markings. This represents the risk intensity coefficient for different types of grading lines. This is a normalized baseline value, representing the average risk of the road network, which depends on the road type (e.g., the value for highways is higher than that for ordinary urban roads). Indicates the normal distance between the vehicle and the road markings;

[0025] (2) When multiple static risk fields are superimposed, the risk of multi-obstacle coupling is added. The modified static risk field is used to avoid misjudging the risk of gaps between multiple obstacles, generating local optimal solutions, and causing aggressive driving behaviors such as "squeezing through gaps" and "driving in a serpentine manner". The model is as follows:

[0026] ;

[0027] in, The coupling enhancement coefficient is recommended to have an initial value of 0.15 to 0.25. The larger the value, the more cautious the driving strategy tends to be. This represents the maximum risk intensity within the directional sensing range; This represents the potential energy field intensity gradient; The normalized baseline value is the average static risk intensity of the scenario.

[0028] (3) When there is a risk of interaction between traffic participants, the risk kinetic energy field of traffic participants is used for description, and the model is as follows:

[0029] ;

[0030] in, This is a longitudinal risk term, representing the risk arising from the expected longitudinal movement of traffic participants; This is a lateral risk term, representing the risk arising from the expected lateral movement of traffic participants; This is the lateral filtering coefficient, which is affected by vehicle type and lane risk filtering effect, reflecting the possibility of lateral movement of vehicles (a value of 0.3 is recommended for small vehicles in the middle lane of the dashed line).

[0031] The preferred dynamic risk field model is as follows:

[0032] In multi-vehicle interaction scenarios, the mutual influence between vehicles can significantly alter traffic dynamics. Therefore, a spatiotemporally coupled risk propagation equation is introduced, and a social force model is used to construct this hybrid risk field that conforms to real-world traffic flow. Based on the social force model, the time-varying dynamic expectations of vehicles are constructed as follows:

[0033] ;

[0034] in, This represents the expected speed of the k-th participant; Let this be the social interaction term between the k-th participant and the j-th participant; This represents the potential energy field influence coefficient.

[0035] The preferred spatiotemporal coupled risk propagation equation defines the spatiotemporal propagation law of the risk field to improve the accuracy of the risk field description mapping to scene elements.

[0036] ;

[0037] in, Let α be the spatiotemporal coupling tensor, β be the static risk diffusion coefficient, and β be the dynamic risk propagation coefficient. The term represents a sudden risk event, and the Dirac function is used to record significant risks that occur within a very short period of time to prevent unexpectedly large data points from arising from sudden changes in intensity.

[0038] ;

[0039] Where m represents the number of emergencies; γ is the risk weight of the emergencies; The spatiotemporal coordinates of the risk field measurement points; The coordinates of the sudden event are denoted by δ; the Dirac function represents the step risk caused by the spatiotemporal location jump of the sudden event. Indicates the duration of an emergency.

[0040] Preferably, the scene element reverse generation module includes:

[0041] The road network is generated based on the gradient distribution of the static risk field, and the road model is reconstructed by fitting B-spline curves. The static obstacle layout is generated by using a non-uniform Poisson point process according to the potential energy extremum and density distribution. The traffic flow and pedestrian behavior that conform to the characteristics of reality are generated by combining the vehicle dynamics model and the social force model with the dynamic risk field and the spatiotemporally coupled risk propagation equation as environmental variables.

[0042] Preferably, environment variable injection includes:

[0043] Weather parameters (rainfall intensity, fog concentration) are mapped to sensor noise model parameters; the visibility attenuation of the risk field is adjusted according to the lighting conditions to enhance the risk propagation effect when vehicles pass through water accumulation at high speed in rainy weather, causing water fog.

[0044] Preferably, the system also includes a closed-loop verification and optimization module: based on the test results of the autonomous driving system in the generated scenario and the generator training performance, it provides feedback to adjust the risk field parameters, including:

[0045] When the system fails to respond within the preset risk threshold, the potential energy field strength in the corresponding area is increased;

[0046] When the system's error generation rate is too high, optimize and adjust the key parameters of the risk propagation equation; when the traffic elements are incorrectly mapped to the risk field parameters, reduce the confidence of the erroneous elements corresponding to the training parameter intervals of the offspring.

[0047] This invention also discloses an automatic generation system for ADAS test scenarios based on a parameterized preset risk field, used to implement the above-mentioned method. The system includes:

[0048] a) Dual-channel input interface, including a physical feature extraction channel and a semantic large model parsing channel connected to the accident database, and a hybrid GAN generative adversarial network for generating edge cases;

[0049] b) Hybrid risk field model, which constructs a static risk field and a dynamic risk field that incorporates spatiotemporal coupling equations;

[0050] c) Scene element reverse generation module, which maps known traffic elements through a risk field, including:

[0051] The road topology generation unit generates a road network that conforms to the ISO 34502 standard based on the potential energy field gradient distribution.

[0052] The obstacle layout unit generates obstacles constrained by a static risk field through a Poisson point sampling process.

[0053] The traffic flow simulation unit integrates vehicle dynamics models and social force models to generate interactive behaviors of traffic participants.

[0054] The 3D rendering adapter converts the generated traffic element data into a scene description file that can be recognized by CARLA / UE (Python).

[0055] The present invention has the following advantages:

[0056] 1. Improved test coverage: Through parametric modeling of risk fields, the efficiency of scene generation in high-risk areas is significantly improved, effectively covering long-tail scenarios such as emergency lane changes that violate traffic regulations and pedestrians suddenly appearing out of nowhere.

[0057] 2. Enhanced safety verification: Significantly increases the number of successful emergency braking / avoidance cases that were not anticipated by the tested autonomous driving algorithm, effectively shortening the construction time for complex scenarios such as "road construction zones";

[0058] 3. Enhanced interpretability: The mapping relationship between risk field parameters and scene elements supports traceable analysis of tests and supports the construction of multi-level risk fields from micro-vehicle interactions to macro-road network planning. Attached Figure Description

[0059] Figure 1 Overview of the technical approach provided by this invention;

[0060] Figure 2 The risk field modeling roadmap provided for this invention;

[0061] Figure 3 The implementation path of the generator module provided by this invention;

[0062] Figure 4 The spatiotemporal coupling risk propagation process provided by this invention. Detailed Implementation

[0063] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] like Figures 2-4 As shown, the process from risk field constructor to scene fusion technology is as follows:

[0065] 1. Static risk field parameters

[0066] Input source: The accident database (NHTSA) processed by the physical feature extraction channel, i.e., the real accident database;

[0067] Key parameter: Potential field gradient Extreme coordinates of obstacle potential energy distribution and risk intensity Risk radius (Considering environmental / visibility attenuation factors);

[0068] Output format: Spatial discretized risk matrix ;

[0069] 2. Dynamic risk field parameters

[0070] Input source: Semantic instruction input;

[0071] Key parameters: Spatiotemporal propagation coefficients (diffusion coefficient α, propagation coefficient β), Dirac function weights for sudden events (risk weights) );

[0072] Output format: Spatiotemporal coupling tensor ;

[0073] 3. Road topology generation

[0074] Fusion Logic: Potential Field Gradient and static risk field High-risk areas are marked;

[0075] Based on the ISO 34502 standard, B-spline curves are used to fit equipotential lines, and the coordinates of the lane centerline and the point cloud of the road boundary are output.

[0076] 4. Static obstacle layout

[0077] Implement non-uniform Poisson point sampling, sampling density :

[0078]

[0079] in These are extreme value coordinates. The baseline sampling density;

[0080] Obstacle types are determined by Mapping, with positions taken from extreme coordinates ;

[0081] 5. Traffic flow dynamic behavior injection

[0082] Social force model constructs vehicle dynamic expectations:

[0083]

[0084] Inject risk pulses at specified coordinates using the Dirac function:

[0085] ;

[0086] 6. Risk Correction for Environmental Parameter Integration

[0087] Core logic: Environmental parameters (weather / light) simultaneously affect the propagation of traffic risks and the performance of vehicle sensors.

[0088] Visibility attenuation correction: fog concentration Expanding the scope of risk impact: Where k is the visibility attenuation factor, and for a standard binocular camera vision sensor (prescan standard), k is 0.5 to 0.75;

[0089] Road surface adhesion coefficient correction: Slippery road surfaces reduce the ability to filter lateral risks;

[0090] ;

[0091] The lateral risk filtering coefficient for slippery road surfaces. The lateral risk filtering coefficient for dry road surfaces. The coefficient of adhesion for wet and slippery road surfaces. The coefficient of adhesion for dry road surfaces;

[0092] For example, for ordinary icy asphalt roads , The increase of 21% will significantly impact horizontal risk assessment due to the large base value of the horizontal risk gradient.

[0093] Sensor degradation (simulation test for non-ideal sensors; the example uses default sensor values ​​from Prescan 8.5.0):

[0094] Millimeter-wave radar: Rainfall intensity Increase noise:

[0095] Signal-to-noise ratio loss = ;

[0096] LiDAR: Fog concentration reduces point cloud density.

[0097] Effective point cloud ratio = ;

[0098] Camera vision: Light intensity Affected dynamic range:

[0099] Image contrast = ;

[0100] Scene 1: Heavy rain at night in a highway construction area

[0101] 1. Construction of a mixed risk field

[0102] 1) Static field input source: physical feature extraction (NHTSA accident database)

[0103] Extracted parameter: Cone risk intensity Risk decay radius ;

[0104] Potential field gradient , where ;

[0105] 2) Dynamic field input source: Semantic instruction input (heavy rain at night on highway construction area)

[0106] Generation parameters: Static risk diffusion coefficient , dynamic risk propagation coefficient , weight of emergency , spatio-temporal coordinates Duration ;

[0107] Define spatio-temporal coupling tensor:

[0108] ;

[0109] ;

[0110] 2. Static risk field analysis

[0111] 1) Road network generation: Generate curved diversion lanes based on potential field gradient (radius of curvature = 200m);

[0112] The gradient direction indicates the normal direction of the curve, and the gradient mutation area , the road shoulder is narrowed to 2.0m; Road point fitting, fitting the center line of the lane along the high-risk equipotential line ( );

[0113] 2) Static obstacle generation: Extract the positions of cone barrels from the extreme points of the static field;

[0114] Define the distribution range R: The area of the equipotential platform boundary ;

[0115] R < 0.3 is mapped to a standard cone barrel ;

[0116] 0.3 < R < 0.5 is mapped to a cone barrel in the splash risk area, corrected ;

[0117] 3. Dynamic risk field analysis

[0118] 1) Semantic instruction analysis:

[0119] Visibility attenuation factor k = 0.5 - 0.75 (affecting );

[0120] The frequency of sudden risk events is greater than 5 times / minute (Dirac function weight );

[0121] 2) Dynamic event injection: A truck splashing water event is generated every 30 seconds, with a water mist radius of 8m. Construction vehicles have a 15% probability of illegally crossing the solid line to change lanes (social force model adjusts expected speed). );

[0122] 4. Environmental variable control

[0123] Environment parameter settings:

[0124] Visibility: Equivalent fog concentration during heavy rain (Values ​​range from 0 to 1, with 1 representing complete occlusion).

[0125] Road surface adhesion: (Reference values ​​for dry road surfaces) );

[0126] Light intensity: Nighttime (Dusk reference value) );

[0127] Rainfall intensity: (Short-term heavy rainfall);

[0128] Static risk radius expanded ;

[0129] Horizontal risk filtering weakened (for );

[0130] 5. Scenario Verification Metrics: Adjust the bucket mechanism based on the AEB trigger failure rate. :

[0131] ;

[0132] The learning rate is set to 0.05, K is the actual collision loss function, and N... cone Let N be the number of collisions related to the cone. total This represents the total number of collisions.

[0133] Scene 2: Pedestrians suddenly dart out from the school zone

[0134] 1. Construction of a mixed risk field

[0135] 1) Static field input source: physical feature extraction (morning rush hour accident data)

[0136] Extracted parameters: Buses picking up and dropping off passengers risk radius Risk benchmark coefficient for pedestrian crossing markings ;

[0137] 2) Dynamic field input source: Hybrid GAN generation

[0138] Generation parameters: pedestrian trajectory variance pedestrian crossing incident , ;

[0139] Define the spatiotemporal coupling tensor:

[0140] ;;

[0141] 2. Static Risk Field Analysis

[0142] 1) Road network generation: along Isopotential lines are used to fit the boundary of the temporary bus stop area, and the point cloud coordinates of the stop area are output. n depends on the point cloud resolution, and the bus lane is marked along the gradient direction (3.2m).

[0143] 2) Equipotential surfaces Designate pedestrian walkway areas;

[0144] 3. Dynamic Risk Field Analysis

[0145] Dirac pulse: Pedestrian model "sprint behavior" and spatiotemporal coordinates of blind spots Coupling ensures blind spot window > ;

[0146] Social force model based on pedestrian behavior fuzzy coefficient Add pedestrian lateral velocity mutation ( (Item Enhancement)

[0147] 4. Environmental variable control

[0148] Environment parameter settings

[0149] Visibility: Morning fog ;

[0150] Road surface adhesion: dry road surface ;

[0151] Sensor Correction: Laser Point Cloud Retention Rate ;

[0152] 5. Scenario Validation Metrics: Based on trajectory prediction error Adjusting the fuzzy coefficient of pedestrian behavior

[0153] ;

[0154] Adjust blind spot events according to the proportion of camera visual obstruction area. ;

[0155] Scenario 3: Oncoming vehicle loses control on an icy curve in a mountainous area.

[0156] 1. Construction of a mixed risk field

[0157] 1) Static field input source: physical feature extraction (high-incidence accident areas in mountainous areas);

[0158] Extracted parameters: Risk of mountain curves ;

[0159] Extreme point of potential energy at the apex of the curve ;

[0160] Potential gradient correction on the shaded side of the mountain ;

[0161] 2) Dynamic field input source: semantic command input (loss of control due to icing on a mountain bend);

[0162] Generation parameters: lateral filtering coefficient Out-of-control events ;

[0163] 2. Static Risk Field Analysis

[0164] 1) Road network generation: Generate S-shaped mountain bends along the potential energy gradient normal (curvature transition R1=150m, R2=180m);

[0165] Along the high-risk equipotential line ( Fit the lane centerline;

[0166] Output curvature change rate ;

[0167] 2) Obstacle generation

[0168] Delineating the distribution range: equipotential surfaces Designate icing zones

[0169] 3. Dynamic Risk Field Analysis

[0170] Correcting the horizontal filtering coefficient (Curve + loss of control), lateral acceleration of the vehicle ;

[0171] Out-of-control vehicle type mapping: Take an equipotential platform at the apex of the opposite curve ( )radius Classified as medium-sized trucks / high-speed small passenger vehicles;

[0172] 4. Environmental variable control

[0173] Environment parameter settings

[0174] Road surface adhesion: ice ;

[0175] Light intensity: Shaded side during the day ;

[0176] Lateral control has almost failed: ;

[0177] Sensor correction: Equivalent sensing delay of 0.5s for camera overexposure / underexposure;

[0178] 5. Scenario Validation Metrics

[0179] Risk of adjusting curvature based on trajectory offset ;

[0180] Adjust lateral risk based on the trigger ratio of avoidance strategies (emergency braking / lane departure / shoulder avoidance) (change). );

[0181] Adjust the penalty mechanism around the icy zone based on the spatiotemporal coordinates of the out-of-control event to force the event to be triggered in the icy zone;

[0182] Parameter optimization implementation process:

[0183] 1. Data Acquisition: Record obstacle collision coordinates and time, sensor misjudgment event types and frequencies, and waypoint deviations between the planned trajectory and the reference path in each round of testing;

[0184] 2. Sensitivity Analysis: Calculate the factors influencing the safety indicators of the parameters.

[0185] ;

[0186] For parameters to be optimized, For the hazards of collision incidents, The function represents the inhibitory effect of parameter adjustment and collision occurrence on training convergence, prioritizing the adjustment of parameters. The parameters are used to construct tests with diverse risks.

[0187] 3. Stability verification: Check the variance of parameter changes every 5 iterations.

[0188] ;

[0189] Comparative experiment on the generation of scenes in highway construction areas during nighttime heavy rain:

[0190] 1. Experimental setup

[0191] Experimental environment: NVIDIA GeForce RTX 5080+ high-speed SSD storage array;

[0192] Comparison methods: Method A: Manual scene editing based on Prescan and Carla platforms;

[0193] Method B: Random search generation (used by NHTSA);

[0194] Method C: Risk-free field-guided GAN generation with preset parameters;

[0195] This invention: Risk-field driven dual-channel GAN ​​dynamic-static hybrid generation

[0196] Evaluation Metrics: Scenario Coverage: The percentage of effective scenarios that trigger security features such as AEB / LDW;

[0197] Generation efficiency: Average generation time per scene (including rendering);

[0198] Scenario diversity: Differences in the distribution of JSD metric parameters;

[0199] 2. Parameter settings

[0200] To ensure the applicability of the comparison method, the scene parameters are adjusted as follows:

[0201] The frequency of dynamic risk events has been adjusted to once per minute, and the parameters for sudden events have been adjusted to... ;

[0202] Static field: Reconstructing the cone-shaped gradient layout of the construction area based on historical accident data;

[0203] Random search generates a validity evaluation function:

[0204] ;

[0205] Where TTC is the collision time; JSD is the Jessen-Shannon distance; weights , , ;

[0206] 3. Experimental Results

[0207] A 5 100% 0.08 0.62 B 34 34.7% 0.57 0.33 C 105 61.5% 1.75 0.19 This invention 322 89.3% 5.37 0.09

[0208] In the above embodiments, the vehicles are all small passenger cars. This is only for the convenience of explanation and does not mean that the present invention cannot be used for medium, large, heavy vehicles, special vehicles, etc. In specific implementation, the parameters need to be corrected according to the vehicle parameters. In particular, when designing risk modeling, the vehicle model and vehicle condition will significantly affect the risk field distribution.

[0209] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for automatically generating ADAS test scenarios based on parameterized preset risk fields, characterized in that: The physical parameters of the risk field are extracted from real accident data through the physical feature extraction channel to construct a static risk field; natural language instructions are mapped to traffic parameters of the risk field through the semantic big model parsing channel. A hybrid GAN generative adversarial network is generated to generate edge cases using a dual-channel hybrid approach of physical feature extraction channel and semantic big model parsing channel to construct a dynamic risk field. The static risk field modeling for the road network generation is as follows: ； in, Represents the spatially discretized risk matrix; The total number of obstacles; For the first Extreme coordinates of the potential energy distribution of each obstacle; Risk intensity indicates the inherent danger level of an obstacle; The risk radius reflects the range of impact from obstacles and is related to the vehicle's status and speed. When multiple static risk fields are superimposed, the risk of multi-obstacle coupling is added, and the modified static risk field is used to avoid misjudging the risk of gaps between multiple obstacles, generating local optima, and leading to aggressive driving behavior. The model is as follows: ； in, Represents a static risk field; For coupling enhancement coefficient, ; This represents the maximum risk intensity within the directional sensing range; This represents the gradient of the potential energy field. ; The normalized baseline value is the average static risk intensity of the scenario. The dynamic risk field is modeled as follows: In multi-vehicle interaction scenarios, a spatiotemporally coupled risk propagation equation is introduced, and a social force model is used to construct a hybrid risk field that conforms to real-world traffic flow. Based on the social force model, a time-varying dynamic expectation of vehicles is constructed. ； in, This indicates the expected rate of correction for the social force model; This represents the expected speed of the k-th participant; Let this be the social interaction term between the k-th participant and the j-th participant; The potential energy field influence coefficient; Driven by the physical and traffic parameters of the risk field, a standard road network is generated based on the gradient distribution of the potential energy field. The Poisson point process is implemented to generate obstacles constrained by the static risk field. Traffic flow and pedestrian behavior data are generated by combining the social force model with the dynamic risk field and the spatiotemporally coupled risk propagation equation. Environmental variables are injected, and the risk field parameters are adjusted based on the test results of the ADAS system in the generated scenario and the training performance of the generator.

2. The method for automatically generating ADAS test scenarios based on a parameterized preset risk field according to claim 1, characterized in that: Static risk field modeling for static obstacle generation includes: The gradient risk arising from the superimposed non-point structure of the road, which includes road markings and edges, is modeled as follows: ； in, This represents the total number of road markings. This represents the risk intensity coefficient for different types of grading lines. This is a normalized baseline value, representing the average risk of the road network, which depends on the road type; This indicates the normal distance between the vehicle and the road markings.

3. The method for automatically generating ADAS test scenarios based on a parameterized preset risk field according to claim 1, characterized in that: Spatiotemporal Coupled Risk Propagation Equation: Defines the spatiotemporal propagation law of the risk field to improve the accuracy of the risk field description mapping to scene elements: ； in, Let α be the spatiotemporal coupling tensor, β be the static risk diffusion coefficient, and β be the dynamic risk propagation coefficient. The term represents a sudden risk event, and the Dirac function is used to record significant risks that occur within a very short period of time to prevent unexpectedly large data points from arising from sudden changes in intensity. ； Where m represents the number of emergencies; γ is the risk weight of the emergencies; The spatiotemporal coordinates of the risk field measurement points; The coordinates of the sudden event are denoted by δ; the Dirac function represents the step risk caused by the spatiotemporal location jump of the sudden event. Indicates the duration of an emergency.

4. The method for automatically generating ADAS test scenarios based on a parameterized preset risk field according to claim 1, characterized in that: When there is a risk of interaction between traffic participants, the risk kinetic energy field of traffic participants is used for description, and the model is as follows: ； in, This is a longitudinal risk term, representing the risk arising from the expected longitudinal movement of traffic participants; This is a lateral risk term, representing the risk arising from the expected lateral movement of traffic participants; This is the lateral filtering coefficient, reflecting the likelihood of lateral movement of the vehicle.

5. An automatic ADAS test scenario generation system based on parameterized preset risk fields, characterized in that: For implementing the method of any one of claims 1-4, the system comprises: The dual-channel input interface includes a physical feature extraction channel connected to the accident database and a semantic large model parsing channel, as well as a hybrid GAN generative adversarial network for generating edge cases; A hybrid risk field model integrates a static risk field and a dynamic risk field that incorporates spatiotemporal coupling equations; The scene element reverse generation module includes: Road topology generation unit generates a standard road network based on the potential energy field gradient distribution; The obstacle layout unit generates obstacles constrained by a static risk field through a Poisson point sampling process. The traffic flow simulation unit uses a social force model combined with a dynamic risk field and a spatiotemporally coupled risk propagation equation to generate traffic flow and pedestrian behavior data. The 3D rendering adapter injects environment variables, and the closed-loop verification and optimization module adjusts the risk field parameters based on the test results of the ADAS system in the generated scene and the generator training performance, converting the generated traffic flow and pedestrian behavior data into a scene description file. The execution logic of the closed-loop verification optimization module includes: When the ADAS system fails to respond within a preset risk threshold, the potential energy field strength in the corresponding area is increased. When the error generation rate of the ADAS system is too high, optimize and adjust the key parameters of the risk propagation equation; when the traffic elements are incorrectly mapped to the risk field parameters, reduce the confidence of the erroneous elements corresponding to the training parameter intervals of the offspring.

Images