Dangerous scene generation method and system based on simulator

Abstract

The invention discloses a simulator-based dangerous scene generation method and system, and relates to the technical field of automatic driving, and the method comprises the steps: collecting scene data through a data collection system; performing data processing on the scene data; carrying out weight distribution on the danger degree of the scene elements by utilizing an MOE expert network; decomposing and evaluating the risk contribution weight of each clustering scene element; mapping the time-space relationship of all traffic participants in the dangerous working condition data into a dynamic directed graph and carrying out system clustering; reconstructing the scene of the dangerous scene cluster; and building a scene in the simulator according to the fusion information and the static scene. According to the method, an MOE expert network is used for carrying out weight distribution on danger degrees of scene elements, scene danger level evaluation is realized, a clustering scene is reconstructed according to a scene danger level evaluation result, static scene elements outside a road are added at the same time, and dangerous scene information is reproduced in a simulator.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and more specifically, to a method and system for generating dangerous scenarios based on a simulator. Background Technology

[0002] With the acceleration of urbanization and the development of intelligent transportation technology, autonomous driving technology has become a core direction of the industry due to its goal of "zero accidents and low pollution". The Society of Automotive Engineers (SAE International) classifies autonomous driving technology into six levels from L0 to L5. Among them, L3 and above require that the system and the manufacturer be clearly defined as the driving entity and the responsible entity, respectively. Therefore, its safety verification has become a key prerequisite for commercialization.

[0003] Existing testing methods mainly rely on real vehicle testing and simulation testing: Real vehicle testing requires more than 2.4×10^8 km of mileage verification to ensure system robustness, which has inherent defects such as long testing cycle and high cost, making it difficult to meet the rapid iteration requirements of high-level autonomous driving; Simulation testing can improve testing efficiency through physical modeling and virtual scene generation technology, but its core problem lies in the insufficient ability to build scene library. Its traditional solution relies on expert experience to enumerate and generate scenes, resulting in limited scene coverage, lack of edge cases, and difficulty in truly reflecting complex road conditions (such as multi-target dynamic interaction, extreme weather and sudden abnormal events).

[0004] Existing research on generating simulation test scenarios mostly relies on manually deconstructing elements according to rules, lacking systematic modeling of the implicit logic and spatiotemporal coupling relationship of traffic participants' behavior, thus resulting in insufficient scenario structuring. When generating scenarios using natural driving data, problems such as difficulty in fusing multi-source heterogeneous data and strong noise interference are encountered, resulting in low scenario fidelity and thus limitations of data-driven approaches. Due to the lack of a quantitative assessment mechanism for the risks of generated scenarios, it is difficult to generate high-risk test cases in a targeted manner, resulting in a lack of an assessment system. Summary of the Invention

[0005] This invention provides a method and system for generating dangerous scenarios based on a simulator, in order to overcome at least one technical problem existing in the prior art.

[0006] In a first aspect, embodiments of the present invention provide a method for generating dangerous scenarios based on a simulator, comprising: Design a data acquisition system and use the data acquisition system to collect scene data; The scene data is processed to obtain a standardized scene description framework; the standardized scene description framework contains multiple clustered scene elements; the clustered scene elements include vehicle information, target information, environmental information, traffic information, and road information; Design an MOE expert network, which includes kinematics experts, dynamics experts, behavior prediction experts, system fault experts, environmental perception experts, and scene information experts; The MOE expert network is used to assign weights to the danger levels of scene elements to obtain the scene danger level. The risk contribution weight of each cluster scenario element output by the expert is decomposed and evaluated by calculating the SHAP value. Based on the hazard level of the scenario, obtain hazardous working condition data; By integrating the risk contribution weights of various clustering scenario elements and the Transformer-MOE node embedding encoder, the spatiotemporal relationship of all traffic participants in the hazardous condition data is mapped into a dynamic directed graph. The dynamic directed graph is systematically clustered using a multi-scale clustering method to obtain dangerous scene clusters. Based on the hazard level of the scene, the scenes of the dangerous scene clusters are reconstructed to obtain fused information; Based on the fusion information and the static scene, a scene is built in the simulator.

[0007] Optionally, the data acquisition system includes onboard sensors and V2X communication equipment; wherein, The vehicle-mounted sensors include at least a camera, lidar, GPS, and inertial navigation.

[0008] Optionally, the scene data is processed, specifically as follows: The scene data is filtered using a Kalman filter to obtain filtered data. A spatiotemporal alignment algorithm is used to fuse the filtered data to obtain a standardized scene description framework.

[0009] Optionally, the kinematics expert is used to determine the motion trend based on the target vehicle speed and type; The dynamics expert is used to calculate the lateral force margin of this vehicle; The behavior prediction expert is used to predict the trajectory of other vehicles by combining the intersection shape and the number of lanes, and to calculate the probability of overlap with the current vehicle's path. The system failure expert is used to simulate the control delay time after the failure of a critical component. The environmental perception experts reconstruct perception confidence based on time period and visibility. The scene information expert is used to dynamically label road features.

[0010] Optionally, the dynamic directed graph is subjected to systematic clustering using a multi-scale clustering approach, including upper-level clustering, lower-level clustering, and union clustering; wherein, The upper layer is classified into scene categories using spectral clustering. The Dynamic Time Warping (DTW) algorithm is used to quantify trajectory similarity in the lower layer. Hierarchical clustering is used to merge classes that are closest to the DTW (Data Pointer) distance.

[0011] Optionally, the vehicle information includes speed, acceleration, braking pressure, and yaw rate; The target information includes type, distance, and relative speed; The environmental information includes weather, lighting, and visibility; The traffic information includes lane number and traffic light status; The road information includes road surface material, alignment, and intersection structure.

[0012] Secondly, the present invention also provides a simulator-based dangerous scene generation system, comprising: The first design module is used to design a data acquisition system and use the data acquisition system to collect scene data; The processing module is used to process the scene data to obtain a standardized scene description framework; the standardized scene description framework includes multiple clustered scene elements; the clustered scene elements include vehicle information, target information, environmental information, traffic information, and road information; The second design module is used to design the MOE expert network, which includes kinematics experts, dynamics experts, behavior prediction experts, system fault experts, environmental perception experts, and scene information experts. The allocation module is used to assign weights to the danger levels of scene elements using the MOE expert network to obtain the scene danger level. The first calculation module is used to decompose and evaluate the risk contribution weight of each cluster scenario element output by the experts by calculating the SHAP value. The first acquisition module is used to acquire hazardous working condition data according to the hazard level of the scenario; The first fusion module is used to fuse the risk contribution weights of various clustering scenario elements and the Transformer-MOE node embedding encoder, and map the spatiotemporal relationship of all traffic participants in the hazardous condition data into a dynamic directed graph. The clustering module is used to perform systematic clustering on the dynamic directed graph using a multi-scale clustering method to obtain dangerous scene clusters; The second fusion module is used to reconstruct the scenes of dangerous scene clusters to obtain fused information; The module is used to build a scene in the simulator based on the fusion information and the static scene.

[0013] Optionally, the processing module is specifically used for: The scene data is filtered using a Kalman filter to obtain filtered data. A spatiotemporal alignment algorithm is used to fuse the filtered data to obtain a standardized scene description framework.

[0014] Optionally, the kinematics expert is used to determine the motion trend based on the target vehicle speed and type; The dynamics expert is used to calculate the lateral force margin of this vehicle; The behavior prediction expert is used to predict the trajectory of other vehicles by combining the intersection shape and the number of lanes, and to calculate the probability of overlap with the current vehicle's path. The system failure expert is used to simulate the control delay time after the failure of a critical component. The environmental perception experts reconstruct perception confidence based on time period and visibility. The scene information expert is used to dynamically label road features.

[0015] Optionally, the dynamic directed graph is subjected to systematic clustering using a multi-scale clustering approach, including upper-level clustering, lower-level clustering, and union clustering; wherein, The upper layer is classified into scene categories using spectral clustering. The Dynamic Time Warping (DTW) algorithm is used to quantify trajectory similarity in the lower layer. Hierarchical clustering is used to merge classes that are closest to the DTW (Data Pointer) distance.

[0016] The innovative aspects of this invention include: 1. In this embodiment, scene data is collected through a data acquisition system, and the collected scene data is preprocessed to obtain a standardized scene description framework containing clustered scene elements. By designing an MOE expert network and using the MOE expert network to assign weights to the degree of danger of scene elements, the design of the scene danger domain is completed, and the assessment of the scene danger level is realized. This is one of the innovative points of this embodiment.

[0017] 2. In this embodiment, the dangerous working condition data is mapped into a dynamic directed graph and systematically clustered. The clustered scene is reconstructed based on the scene hazard level assessment results. At the same time, static scene elements other than roads are added to reproduce the dangerous scene information in the simulator. This is one of the innovative points of this embodiment. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of a simulator-based hazardous scene generation method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a simulator-based hazardous scene generation system provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0021] It should be noted that the terms "comprising" and "having," and any variations thereof, in the embodiments and drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0022] This invention discloses a method and system for generating hazardous scenarios based on a simulator. These will be described in detail below.

[0023] Figure 1 A flowchart of a simulator-based hazardous scene generation method provided in this embodiment of the invention is shown below. Figure 1 The simulator-based hazardous scenario generation method provided in this embodiment includes: Step 1: Design a data acquisition system and use the data acquisition system to collect scene data; Step 2: Process the scene data to obtain a standardized scene description framework; the standardized scene description framework contains multiple clustered scene elements; the clustered scene elements include vehicle information, target information, environmental information, traffic information, and road information; Step 3: Design the MOE expert network, which includes kinematics experts, dynamics experts, behavior prediction experts, system fault experts, environmental perception experts, and scene information experts; Step 4: Use the MOE expert network to assign weights to the danger levels of scene elements to obtain the scene danger level; Step 5: Decompose and evaluate the risk contribution weight of each cluster scenario element output by the experts by calculating the SHAP value; Step 6: Obtain hazardous operating condition data based on the hazard level of the scenario; Step 7: Integrate the risk contribution weights of each cluster scenario element and the Transformer-MOE node embedding encoder to map the spatiotemporal relationship of all traffic participants in the hazardous conditions data into a dynamic directed graph; Step 8: Perform systematic clustering on the dynamic directed graph using multi-scale clustering to obtain clusters of hazardous scenarios; Step 9: Based on the scene hazard level, reconstruct the scenes of the hazardous scene clusters to obtain fused information; Step 10: Based on the fusion information and the static scene, build the scene in the simulator.

[0024] For details, please refer to Figure 1 The simulator-based hazardous scenario generation method provided in this embodiment requires the collection of scenario data. Therefore, a data acquisition system is first designed in step 1, and then the scenario data is collected using this data acquisition system. In this embodiment, the data acquisition system integrates vehicle-mounted sensors and V2X communication equipment. The vehicle-mounted sensors include cameras, radar, GPS, and inertial navigation. This data acquisition system can collect at least 6 channels of LVDS (Low Voltage Differential Signaling) high-definition video, 2 channels of high-frame-rate CAN signals from vehicles, 1 channel of GPS information, 1 channel of forward-looking video, and detection results in real time. It also provides vehicle dynamic parameters, traffic participant behavior trajectories, and environmental characteristics (weather, road surface), perfectly solving the scenario data collection problem. At the same time, it can also provide vehicle status information, GPS signals, etc.

[0025] After acquiring the scene data, step 2 involves data processing. For example, Kalman filtering is first used to filter the scene data and eliminate sensor noise, resulting in filtered data. To ensure that data from the same time period is used during scene construction, a spatiotemporal alignment algorithm is employed to fuse the filtered data, thereby aligning the data from multiple sensors to obtain a standardized scene description framework. This standardized scene description framework contains multiple clustered scene elements, including vehicle information, target information, environmental information, traffic information, and road information.

[0026] Vehicle information includes, for example, image frame number, vehicle speed, acceleration, brake pedal signal, brake pressure, accelerator signal, gear information, GPS information, vehicle yaw rate, and lateral acceleration. Object information includes, for example, image frame number, object ID, object type, forward distance from the object to the vehicle (with the foremost point of the vehicle's centerline as the origin), lateral distance from the object to the vehicle, object speed, and relative speed. Environmental information includes weather, lighting conditions, and visibility. Traffic information includes lane number and traffic light status. Road information includes road surface material, alignment, and intersection structure.

[0027] In this invention, the MOE (Mixture of Experts) network is used to assess the danger level of scene elements. Therefore, after obtaining the standardized scene description framework, the MOE network is designed through step 3.

[0028] The MOE expert network of this invention includes kinematics experts, dynamics experts, behavior prediction experts, system fault experts, environmental perception experts, and scene information experts. Among them, kinematics experts are responsible for identifying target vehicle types (such as vehicles, pedestrians, etc.) and judging motion trends (such as acceleration, deceleration) based on the target vehicle's speed and type; dynamics experts can calculate the vehicle's lateral force margin, such as the difference between the vehicle's limit handling capability and the current steering force, thereby assessing the risk of loss of vehicle control; behavior prediction experts combine traffic flow, lane numbers, traffic lights, intersection shapes, and the number of lanes to predict the trajectories of other vehicles and calculate the probability of overlap with the vehicle's path; system fault experts can simulate the control delay time after the failure of key components; environmental perception experts reconstruct perception confidence (such as the target recognition error rate in foggy weather) based on time periods (such as day and night) and visibility; and scene information experts can combine raw data to analyze road surface material and road alignment, thereby dynamically labeling road features (such as the risk of narrowing emergency lanes on highways).

[0029] After designing the MOE expert network, in step 4, the MOE expert network can be used to fuse scene elements. For example, by analyzing scene elements, it can be used to determine which experts are important in the current scene, that is, to assign weights based on their risk level and output a set of weight values. At the same time, each expert independently processes scene elements and outputs a risk assessment score. The weight values ​​and risk assessment scores output by each expert are weighted and summed to obtain the scene risk level. In step 5, the SHAP value is calculated for the MOE expert network. The SHAP value assigns a value to each scene element, which is the contribution of each feature to risk prediction, thus decomposing and evaluating the risk contribution weights output by each expert.

[0030] In step 6, high-frequency and high-risk scenarios can be identified based on the scenario hazard level, thus obtaining hazardous working condition data. In step 7, dynamic directed graph modeling is performed. During modeling, each traffic participant (this vehicle, the target vehicle, pedestrians, etc.) is used as a node, and the interaction strength output by the expert network is used as the edge weight (e.g., if the target vehicle cuts in at a distance of less than 3 meters, a high-weight conflict edge is obtained). The node embedding is achieved by fusing the risk feature vector of the expert network through the Transformer-MOE encoder. That is, by fusing the risk contribution weights of each clustered scenario element and the Transformer-MOE node embedding encoder, the spatiotemporal relationship of all traffic participants in the hazardous working condition data is mapped into a dynamic directed graph.

[0031] After obtaining the dynamic directed graph, in step 8, hierarchical clustering of the dynamic directed graph yields clusters of hazardous scenarios. In this invention, a multi-scale clustering approach is used for hierarchical clustering, which involves upper-level clustering, lower-level clustering, and merging. The upper level can employ spectral clustering to classify scenarios based on macroscopic road network structures (such as intersections and highways). The lower level can use the Dynamic Time Warping (DTW) algorithm to quantify trajectory similarity; the smaller the dynamic time warping distance, the more similar the trajectory patterns. Merging can use hierarchical clustering to merge the classes with the closest DTW distance. For example, during merging, right-turn conflicts and straight-ahead conflicts are grouped into the same subclass due to similar trajectory shapes.

[0032] After obtaining the dangerous scene clusters, in step 9, the scenes of the dangerous scene clusters are reconstructed according to the scene danger level to obtain fusion information. Furthermore, since other elements besides roads, such as houses, shops, and green vegetation, were not included in the clustering analysis, this invention also needs to refer to the static scene elements in the original data sample to make the scene content closer to the actual scene and reproduce the scene where the vehicle is operating. In step 10, when building the scene in the simulator, static scene elements are also considered. The scene is built based on the fusion information and the static scene, thereby enriching the simulator screen.

[0033] The present invention provides a simulator-based method for generating hazardous scenarios. This method collects scenario data through a data acquisition system and preprocesses the collected data to obtain a standardized scenario description framework containing clustered scenario elements. By designing a MOE (Modular Object Explanation) expert network and using it to assign weights to the hazard levels of scenario elements, the method completes the design of the scenario hazard domain, enabling the assessment of the scenario hazard level. Hazardous condition data is mapped to a dynamic directed graph and systematically clustered. Based on the scenario hazard level assessment results, the clustered scenario is reconstructed, and static scenario elements other than roads are added to reproduce the hazardous scenario information in the simulator.

[0034] Based on the same inventive concept, this invention also provides a simulator-based dangerous scene generation system. Figure 2 A schematic diagram of a simulator-based hazardous scene generation system provided in an embodiment of the present invention is shown below. Figure 2 The simulator-based hazardous scene generation system 100 provided in this embodiment of the invention includes: The first design module is used to design a data acquisition system and use the data acquisition system to collect scene data; The processing module is used to process scene data to obtain a standardized scene description framework. The standardized scene description framework contains multiple clustered scene elements. The clustered scene elements include vehicle information, target information, environmental information, traffic information, and road information. The second design module is used to design the MOE expert network, which includes kinematics experts, dynamics experts, behavior prediction experts, system fault experts, environmental perception experts, and scene information experts. The allocation module is used to assign weights to the danger levels of scene elements using the MOE expert network to obtain the scene danger level. The first calculation module is used to decompose and evaluate the risk contribution weight of each cluster scenario element output by the experts by calculating the SHAP value. The first acquisition module is used to acquire hazardous working condition data based on the hazard level of the scenario; The first fusion module is used to fuse the risk contribution weights of elements in each clustering scenario and the Transformer-MOE node embedded encoder, mapping the spatiotemporal relationship of all traffic participants in the hazardous conditions data into a dynamic directed graph. The clustering module is used to perform systematic clustering of dynamic directed graphs using multi-scale clustering methods to obtain clusters of hazardous scenarios. The second fusion module is used to reconstruct the scenes of dangerous scene clusters to obtain fused information; The setup module is used to build scenes in the simulator based on fusion information and static scenes.

[0035] For details, please refer to Figure 2 The simulator-based hazardous scene generation system 100 provided in this embodiment of the invention requires the collection of scene data. Therefore, a data acquisition system is first designed through the first design module, and then the scene data is collected using this data acquisition system. In this embodiment, the data acquisition system integrates vehicle-mounted sensors and V2X communication equipment. The vehicle-mounted sensors include cameras, radar, GPS, and inertial navigation. Through this data acquisition system, at least 6 channels of LVDS (Low Voltage Differential Signaling) high-definition video, 2 channels of vehicle high-frame-rate CAN signals, 1 channel of GPS information, 1 channel of forward-looking video, and detection results can be collected in real time. It also provides vehicle dynamic parameters, traffic participant behavior trajectories, and environmental characteristics (weather, road surface), perfectly solving the scene data collection problem. At the same time, it can also provide vehicle status information, GPS signals, etc.

[0036] After acquiring scene data, the processing module performs data processing. For example, Kalman filtering is first used to filter the scene data and eliminate sensor noise, resulting in filtered data. To ensure that data from the same time period is used during scene construction, a spatiotemporal alignment algorithm is employed to fuse the filtered data, thereby aligning data from multiple sensors to obtain a standardized scene description framework. This standardized scene description framework contains multiple clustered scene elements, including vehicle information, target information, environmental information, traffic information, and road information.

[0037] Vehicle information includes, for example, image frame number, vehicle speed, acceleration, brake pedal signal, brake pressure, accelerator signal, gear information, GPS information, vehicle yaw rate, and lateral acceleration. Object information includes, for example, image frame number, object ID, object type, forward distance from the object to the vehicle (with the foremost point of the vehicle's centerline as the origin), lateral distance from the object to the vehicle, object speed, and relative speed. Environmental information includes weather, lighting conditions, and visibility. Traffic information includes lane number and traffic light status. Road information includes road surface material, alignment, and intersection structure.

[0038] In this invention, the danger level of scene elements is evaluated using the MOE (Mixture of Experts) expert network. Therefore, after obtaining the standardized scene description framework, the MOE expert network is designed through the second design module.

[0039] The MOE expert network of this invention includes kinematics experts, dynamics experts, behavior prediction experts, system fault experts, environmental perception experts, and scene information experts. Among them, kinematics experts are responsible for identifying target vehicle types (such as vehicles, pedestrians, etc.) and judging motion trends (such as acceleration, deceleration) based on the target vehicle's speed and type; dynamics experts can calculate the vehicle's lateral force margin, such as the difference between the vehicle's limit handling capability and the current steering force, thereby assessing the risk of loss of vehicle control; behavior prediction experts combine traffic flow, lane numbers, traffic lights, intersection shapes, and the number of lanes to predict the trajectories of other vehicles and calculate the probability of overlap with the vehicle's path; system fault experts can simulate the control delay time after the failure of key components; environmental perception experts reconstruct perception confidence (such as the target recognition error rate in foggy weather) based on time periods (such as day and night) and visibility; and scene information experts can combine raw data to analyze road surface material and road alignment, thereby dynamically labeling road features (such as the risk of narrowing emergency lanes on highways).

[0040] After designing the MOE expert network, the allocation module uses the MOE expert network to fuse scene elements, such as assigning weights to the hazard levels of scene elements to obtain the scene hazard level. Simultaneously, the first calculation module calculates the SHAP value and uses the SHAP value to calculate the contribution of each feature to risk prediction, thereby decomposing and evaluating the risk contribution weights output by each expert.

[0041] The first acquisition module identifies high-frequency, high-risk scenarios based on the scene's hazard level, thus obtaining hazardous condition data. The first fusion module performs dynamic directed graph modeling. During modeling, each traffic participant (this vehicle, the target vehicle, pedestrians, etc.) is treated as a node, and the interaction strength output by the expert network is used as the edge weight (e.g., if the target vehicle cuts in at a distance of less than 3 meters, a high-weight conflict edge is obtained). The risk feature vector of the expert network is fused through the Transformer-MOE encoder to achieve node embedding. That is, by fusing the risk contribution weights of each clustered scene element and the Transformer-MOE node embedding encoder, the spatiotemporal relationships of all traffic participants in the hazardous condition data are mapped into a dynamic directed graph.

[0042] After obtaining the dynamic directed graph, the clustering module performs systematic clustering on the graph to obtain clusters of hazardous scenarios. In this invention, a multi-scale clustering approach is used for systematic clustering, which involves upper-level clustering, lower-level clustering, and merging. The upper level can employ spectral clustering to classify scenarios based on macroscopic road network structures (such as intersections and highways). The lower level can use the Dynamic Time Warping (DTW) algorithm to quantify trajectory similarity; the smaller the dynamic time warping distance, the more similar the trajectory patterns. Merging can use hierarchical clustering to merge the clusters with the closest DTW distance. For example, during merging, right-turn conflicts and straight-ahead conflicts are grouped into the same subclass due to similar trajectory shapes.

[0043] After obtaining the dangerous scene clusters, the second fusion module reconstructs the scenes within these clusters based on their danger levels, obtaining fusion information. Furthermore, since the clustering analysis did not include other elements besides roads, such as static scene elements like buildings, shops, and green vegetation, this invention, in order to make the scene content more closely resemble the actual scene and reproduce the scene where the vehicle is operating, also needs to refer to the static scene elements in the original data sample. When building the scene in the simulator, the construction module also considers the static scene elements, building the scene based on the fusion information and the static scene elements, thereby enriching the simulator's visuals.

[0044] The present invention provides a simulator-based hazardous scene generation system. This system collects scene data through a data acquisition system and preprocesses the collected data to obtain a standardized scene description framework containing clustered scene elements. By designing a MOE (Modular Object Explanation) expert network and using it to assign weights to the hazard levels of scene elements, the system completes the design of the scene hazard domain, enabling the assessment of scene hazard levels. Hazardous condition data is mapped to a dynamic directed graph and systematically clustered. Based on the scene hazard level assessment results, the clustered scenes are reconstructed, and static scene elements other than roads are added to reproduce the hazardous scene information in the simulator.

[0045] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0046] Those skilled in the art will understand that the modules in the system of the embodiments can be distributed in the system of the embodiments as described in the embodiments, or they can be located in one or more systems different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating dangerous scenarios based on a simulator, characterized in that, include: Design a data acquisition system and use the data acquisition system to collect scene data; The scene data is processed to obtain a standardized scene description framework; The standardized scene description framework includes multiple clustering scene elements; the clustering scene elements include vehicle information, target information, environmental information, traffic information, and road information; Design an MOE expert network, which includes kinematics experts, dynamics experts, behavior prediction experts, system fault experts, environmental perception experts, and scene information experts; The MOE expert network is used to assign weights to the danger levels of scene elements to obtain the scene danger level. The risk contribution weight of each cluster scenario element output by the expert is decomposed and evaluated by calculating the SHAP value. Based on the hazard level of the scenario, obtain hazardous working condition data; By integrating the risk contribution weights of various clustering scenario elements and the Transformer-MOE node embedding encoder, the spatiotemporal relationship of all traffic participants in the hazardous condition data is mapped into a dynamic directed graph. The dynamic directed graph is systematically clustered using a multi-scale clustering method to obtain dangerous scene clusters. Based on the hazard level of the scene, the scenes of the dangerous scene clusters are reconstructed to obtain fused information; Based on the fusion information and the static scene, a scene is built in the simulator.

2. The method for generating hazardous scenarios based on a simulator according to claim 1, characterized in that, The data acquisition system includes vehicle-mounted sensors and V2X communication equipment; wherein... The vehicle-mounted sensors include at least a camera, lidar, GPS, and inertial navigation.

3. The method for generating dangerous scenarios based on a simulator according to claim 1, characterized in that, The scene data is processed as follows: The scene data is filtered using a Kalman filter to obtain filtered data. A spatiotemporal alignment algorithm is used to fuse the filtered data to obtain a standardized scene description framework.

4. The method for generating dangerous scenarios based on a simulator according to claim 1, characterized in that, The kinematics experts are used to determine the motion trend based on the target vehicle speed and type; The dynamics expert is used to calculate the lateral force margin of this vehicle; The behavior prediction expert is used to predict the trajectory of other vehicles by combining the intersection shape and the number of lanes, and to calculate the probability of overlap with the current vehicle's path. The system failure expert is used to simulate the control delay time after the failure of a critical component. The environmental perception experts reconstruct perception confidence based on time period and visibility. The scene information expert is used to dynamically label road features.

5. The method for generating hazardous scenarios based on a simulator according to claim 1, characterized in that, The dynamic directed graph is systematically clustered using a multi-scale clustering approach, including upper-level clustering, lower-level clustering, and union clustering; wherein, The upper layer is classified into scene categories using spectral clustering. The Dynamic Time Warping (DTW) algorithm is used to quantify trajectory similarity in the lower layer. Hierarchical clustering is used to merge classes that are closest to the DTW (Data Pointer) distance.

6. The method for generating hazardous scenarios based on a simulator according to claim 1, characterized in that, The vehicle information includes speed, acceleration, braking pressure, and yaw rate; The target information includes type, distance, and relative speed; The environmental information includes weather, lighting, and visibility; The traffic information includes lane number and traffic light status; The road information includes road surface material, alignment, and intersection structure.

7. A simulator-based hazardous scene generation system, characterized in that, include: The first design module is used to design a data acquisition system and use the data acquisition system to collect scene data; The processing module is used to process the scene data to obtain a standardized scene description framework; the standardized scene description framework includes multiple clustered scene elements; the clustered scene elements include vehicle information, target information, environmental information, traffic information, and road information; The second design module is used to design the MOE expert network, which includes kinematics experts, dynamics experts, behavior prediction experts, system fault experts, environmental perception experts, and scene information experts. The allocation module is used to assign weights to the danger levels of scene elements using the MOE expert network to obtain the scene danger level. The first calculation module is used to decompose and evaluate the risk contribution weight of each cluster scenario element output by the experts by calculating the SHAP value. The first acquisition module is used to acquire hazardous working condition data according to the hazard level of the scenario; The first fusion module is used to fuse the risk contribution weights of various clustering scenario elements and the Transformer-MOE node embedding encoder, and map the spatiotemporal relationship of all traffic participants in the hazardous condition data into a dynamic directed graph. The clustering module is used to perform systematic clustering on the dynamic directed graph using a multi-scale clustering method to obtain dangerous scene clusters; The second fusion module is used to reconstruct the scenes of dangerous scene clusters to obtain fused information; The module is used to build a scene in the simulator based on the fusion information and the static scene.

8. The simulator-based hazardous scene generation system according to claim 7, characterized in that, The processing module is specifically used for: The scene data is filtered using a Kalman filter to obtain filtered data. A spatiotemporal alignment algorithm is used to fuse the filtered data to obtain a standardized scene description framework.

9. The simulator-based hazardous scene generation system according to claim 7, characterized in that, The kinematics experts are used to determine the motion trend based on the target vehicle speed and type; The dynamics expert is used to calculate the lateral force margin of this vehicle; The behavior prediction expert is used to predict the trajectory of other vehicles by combining the intersection shape and the number of lanes, and to calculate the probability of overlap with the current vehicle's path. The system failure expert is used to simulate the control delay time after the failure of a critical component. The environmental perception experts reconstruct perception confidence based on time period and visibility. The scene information expert is used to dynamically label road features.

10. The simulator-based hazardous scene generation system according to claim 7, characterized in that, The dynamic directed graph is systematically clustered using a multi-scale clustering approach, including upper-level clustering, lower-level clustering, and union clustering; wherein, The upper layer is classified into scene categories using spectral clustering. The Dynamic Time Warping (DTW) algorithm is used to quantify trajectory similarity in the lower layer. Hierarchical clustering is used to merge classes that are closest to the DTW (Data Pointer) distance.

Images