An automatic driving known unsafe scene library construction method based on self-learning derivation theory

Abstract

The application discloses a kind of based on self-learning derivation theory's automatic driving known unsafe scene library construction method, its method includes the steps as follows: first, expand known scene element;Second, establish explainable scene danger evaluation system;Third, derive known unsafe scene;Fourth, construct known unsafe scene library;Beneficial effect: compress the proportion of unknown scene in principle, effectively reduce the risk brought by unknown unsafe scene;Automatic division of safe scene and unsafe scene provides effective test scene selection basis for intelligent networked vehicle automatic driving system test;It can learn the characteristics of known unsafe scene by the method of deep reinforcement learning to learn derivation;Enough based on the known unsafe scene in scene library generalization derivation a series of known unsafe scene;Can greatly improve the effectiveness of simulation verification;Provide comprehensive reference for its optimization design.

Description

Technical Field

[0001] This invention relates to a method for constructing a library of known unsafe scenarios for autonomous driving, and particularly to a method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory. Background Technology

[0002] Currently, with the continuous development of automotive technology, the level of vehicle intelligence is increasing, and intelligent connected vehicles have become one of the main development directions of automotive technology. Intelligent technologies have been researched and applied in all aspects of vehicles. The development of intelligent connected vehicles is accompanied by the development of various intelligent technologies. According to the typical V-model development process for vehicles, effectively designing various testing stages, including HIL testing and MIL testing, during the development of these intelligent technologies can significantly improve the development efficiency of the entire vehicle and its various systems, thereby greatly shortening the development cycle. In the simulation testing of intelligent connected vehicles, setting appropriate operating scenarios is an extremely important step. Scientific and effective scenario settings can make the entire test more effective and convincing. The theory of autonomous driving test scenario library is a research theory with extremely high research and application value that has emerged against this background.

[0003] Based on existing research and practicality analysis of test scenarios, the international functional safety standard ISO PAS 21448 is expected to classify scenarios from two dimensions: safety and known factors. These two dimensions divide vehicle operation scenarios into four areas: known safe scenarios, known unsafe scenarios, unknown unsafe scenarios, and unknown safe scenarios. The safety of operating scenarios varies slightly depending on the driver's skills and driving style, making it a relatively subjective parameter. However, for ordinary drivers, the fluctuation range is not large, and for a given autonomous driving system, it can be quantitatively distinguished using certain evaluation indicators. Quantitatively distinguishing the safety of test operating scenarios is extremely important. In the autonomous driving testing process, especially simulation testing, a good autonomous driving system can usually handle safe scenarios well, while its ability to handle unsafe scenarios is often key to whether an autonomous driving system has high versatility. In other words, to verify whether the tested autonomous driving system has strong versatility, it needs to be tested through a series of unsafe scenarios. In the early stages of development, the proportion of both known and unknown unsafe scenarios was relatively high. Unknown unsafe scenarios are a major source of safety risks for autonomous driving systems, and this has become a significant pain point in the field of expected functional safety for autonomous driving. Unknown unsafe scenarios are characterized by the inability to define requirements and the difficulty in quantifying their evaluation. To reduce the proportion of unknown unsafe scenarios, transforming them into known unsafe scenarios through a series of methods has become an effective solution, indirectly reducing the risks posed by unknown unsafe scenarios.

[0004] One efficient method for compressing unknown unsafe scenarios into expanded known unsafe scenarios is to create a library of known unsafe scenarios. In the process of expanding this library, previously unknown unsafe scenarios are systematically discovered and transformed into known unsafe scenarios based on interpretable principles.

[0005] Chinese patent CN202011533463.6 discloses a method for modeling autonomous driving scenarios based on the domain-specific modeling language ADSML, providing a simple and easy-to-understand approach to autonomous driving scenario modeling and enabling efficient creation of instance models of autonomous driving scenarios. Chinese patent CN202210743020.2 discloses a VR scene generation method, device, equipment, and storage medium. This method uses scene images as input to train a target neural network model for VR scene generation. Chinese patent CN202210734738.5 discloses a scene 3D intelligent reconstruction system and method based on BIM and deep learning, overcoming the shortcomings of high cost and long reconstruction time in existing technologies using depth information for 3D reconstruction, and enriching the technical means of 3D reconstruction to a certain extent. However, all three patents can only construct a certain number of known scenes based on data, and cannot effectively improve the scene coverage of the constructed scene library, nor can they effectively reduce the potential unpredictable risks brought by unknown and unsafe scenes, thus failing to fundamentally solve the pain point of unknown and unsafe scenes. Summary of the Invention

[0006] The purpose of this invention is to propose a scheme for generating a library of known unsafe scenarios, which solves the problem that existing autonomous driving scenarios or scenario libraries can only build a certain number of known scenarios based on data, and cannot effectively improve the scenario coverage of the constructed scenario library, nor can they effectively reduce the potential unpredictable risks brought about by unknown unsafe scenarios. The invention provides a method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory.

[0007] The present invention provides a method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory, the method comprising the following steps:

[0008] The first step is to expand the known scene elements. The specific steps are as follows:

[0009] Step 1: Integration of Known Scene Element Library: Establish a library of scene elements appearing in known scenes, and integrate them according to various attributes of scene elements to form a library of known scene elements. This library provides data support for the expansion of scene elements in Steps 2 and 3. The sources of the expansion of known scene elements include, but are not limited to, various traffic scene-related standards, or extraction from various online and offline dictionaries based on rules or neural networks.

[0010] Step 2: Expanding the configuration of known static elements: Static elements are defined as follows: Static elements in this step refer to the various objectively existing scene elements in the scene when the scene is designed. They exist as static elements when separated from the specific scene, but they are not necessarily static in the specific dynamic scene.

[0011] Step 3: Expanding the Spatiotemporal Sequence of Known Dynamic Elements: Dynamic elements are defined as follows. In specific scenarios, the static elements defined in Step 2 often undergo a dynamic process to interact with other static elements in the scene, thereby forming an organic scene whole. This dynamic process is defined as the dynamic element of the scene.

[0012] Step 4: Expansion of Typical Physical Entity Elements: In Steps 2 and 3, a known static element library and a known dynamic element library were formed respectively. The known static element library includes a known traffic participant library, a known background environment library, and a special known static element sub-library. The known dynamic element library includes time series, spatial series, velocity series, and behavioral series. By arranging and combining some of the above-mentioned related sub-libraries, the known scene element library can be expanded. Subsequently, unreasonable scene element combinations are deleted based on rules to complete the expansion of typical physical entity elements and form a reasonable known scene element library.

[0013] The second step is to establish an interpretable scenario hazard assessment system, with the following specific steps:

[0014] Step 1: Scene Hazard Characterization Coefficient: Only by reasonably judging the safety of a scene can known unsafe scenes be correctly extracted from known scenes. For this purpose, an interpretable scene hazard evaluation system needs to be established. First, a characterization method, namely scene hazard characterization coefficient, needs to be proposed based on the factors that may cause scene unsafety. Based on the needs of the evaluation model, a series of scene hazard characterization coefficients are extracted from the scene for input into the evaluation model.

[0015] Step 2, Scene Hazard Assessment Model: Based on the scene hazard characterization coefficients from Step 1, an interpretable assessment model is needed to comprehensively evaluate the hazard factors in the scene in order to assess the hazard level of the scene. There are two main theories of scene hazard assessment: physical model-based methods and data-driven machine learning-based methods. Due to the lack of interpretability of machine learning-based methods, there is a possibility of incomprehensible false positives and false negatives when using them as scene hazard assessment models for known unsafe scene libraries. Therefore, physical model-based methods will be adopted as scene hazard assessment models for known unsafe scene libraries.

[0016] Step 3: Scene Hazard Regression Equation: In Step 2, a scene hazard assessment model based on a physical model was determined, mainly including the collision time method and the headway method. In order to comprehensively detect the hazard of the scene, a scene hazard regression equation was designed to integrate the two models, form a unified index, and highlight the hazard factors.

[0017] The third step involves deriving known unsafe scenarios. This step uses two different approaches to derive scenarios, resulting in known scenarios with broad coverage. The elements of these derived scenarios are based on the known scenario element library generated in the first step. This known scenario library serves as a candidate for the known unsafe scenario library. After deriving, it is sent to the fourth step for screening and verification before being integrated to form the known unsafe scenario library. The specific steps for deriving known unsafe scenarios are as follows:

[0018] Step 1: Scene Construction Based on Self-Learning Theory: The theoretical basis of this method is the machine learning method of reinforcement learning and Bayesian networks. By training the computer to learn a small number of existing but typical known unsafe scenes, a series of known scenes are derived based on the learning training.

[0019] Step 2: Scene construction based on self-extension theory: The theoretical basis of this method is mainly ontology. It generates all combinations of a small number of typical known unsafe scenes as the basis in the current parameter space by directly arranging and combining parameters.

[0020] The above steps one and two are cyclical. That is, the known scene sub-library derived in step one can serve as the self-extending data base required in step two, and the known scene sub-library derived in step two can also serve as the self-learning data base required in step one. This cycle iterates until scene construction based on self-learning and self-extending theories can no longer effectively generate scenes, at which point the iteration is completed, a known scene library is formed, and it is sent to step four.

[0021] Step 4: Construct a library of known insecure scenarios. The specific steps are as follows:

[0022] Step 1: Derivative Scenario Risk Assessment: Based on the interpretable scenario risk assessment system designed in Step 2, the risk assessment of the known scenario library generated in Step 3 can be performed.

[0023] Step 2, Derived Scenario Pattern Classification: In this step, the known scenario library generated in step 3 is identified and classified. First, safe and unsafe scenarios are distinguished from the known scenarios, thereby defining the scope of the known unsafe scenario library. Then, the defined known unsafe scenario library is classified by clustering to facilitate the integration and use of the scenario library.

[0024] Step 3: Integration of the Known Unsafe Scenario Library: The known unsafe scenario library obtained in Step 2 is sorted and organized according to the classification results, numbered, and recorded. The recording method includes, but is not limited to, Excel spreadsheets or XML extensible tag language files, so that the established known unsafe scenario library has good readability. The generated known unsafe scenario library is converted according to the OpenSCENARIO international autonomous driving test scenario standard to generate *.xodr or *.osgb files, so that the established known unsafe scenario library has good extensibility.

[0025] The specific steps of the first and second steps are as follows:

[0026] Step 1: Expanding the Known Static Element Types: When establishing the known unsafe scenario library, it is necessary to input a set number of typical known unsafe scenarios as a benchmark, such as ghost peek-out scenarios, severe weather scenarios, and sudden braking scenarios. These scenarios include different types of static elements such as vehicles, pedestrians, and other animals. Even if these static elements are in the same spatial location, changes in their type attributes can seriously affect the usability of the scenario. Therefore, the first step in expanding the known static element configuration is to expand it according to the type of static element, adding static elements with similar attributes to the same known static element sub-library, thus expanding it into multiple known static element sub-libraries.

[0027] Step Two: Expanding the Appearance of Known Static Elements: The appearance attributes of static elements, such as color and size, are extremely important parameters in a scene. There are two scenarios for this: First, the appearance of the static element is clearly defined. Typical examples of this include static elements like traffic cones, cars, and traffic signs. For these types of static elements, expansion should be based on relevant national standards or product information to comprehensively cover their various forms in the scene. Second, the appearance of the static element is not clearly defined or cannot be defined. Typical examples include static elements such as pedestrians, various animals, and road green belts. For these static elements, the static elements are expanded by discretizing important attributes. For example, the height of adult males is generally in the range of 160 to 185 centimeters. Therefore, the appearance size parameters of the static elements of adult males are expanded by taking a step size of 5 centimeters, that is, 160, 165, 170, 175, 180, and 185 centimeters. The size of the step size selected during the expansion directly affects the coverage of the scene library. In the first and second stages, a known traffic participant library of known static elements is formed.

[0028] Step 3: Expanding the Background Environment of Known Static Elements: Another crucial aspect of static elements is the background environment, which includes lighting conditions, weather conditions, and road conditions. These static elements have a significant impact on the safety of the scene, so they also need to be expanded. The expansion process mainly refers to autonomous driving-related standards, such as ISOPAS 21448. By expanding these static elements, a known background environment library is formed. This sub-library is independent of any other known static element sub-library.

[0029] Step 4: Further expansion of known static elements: Static elements are also expanded based on other criteria, such as lane line wear parameters. These attributes are parameters used for special processing of specific static elements, which can form a special known static element sub-library as the parameter basis when expanding a certain type of static element.

[0030] The specific steps of the first step and the third step are as follows:

[0031] Step 1: Time Series Expansion of Known Dynamic Elements: The concept of time series does not exist in real-world scenarios. However, for simulation test scenarios, since simulation calculations are performed based on time discreteness, it is necessary to finely divide the simulation time when designing a library of known unsafe scenarios in order to conduct detailed analysis of unsafe factors. Therefore, the first step in expanding the spatiotemporal sequence of known dynamic elements is to expand the scenarios based on time series.

[0032] Step 2: Expansion of Known Dynamic Element Spatial Sequence: In a scene, the relative positional relationship between different scene elements is one of the key factors in the interaction between various scene elements. The spatial position of the test vehicle and the surrounding scene elements also directly affects the safety of the known scene. Therefore, the most important step in expanding the spatiotemporal sequence of known dynamic elements is to expand the scene based on the spatial sequence. The range of values ​​for the spatial sequence is determined based on the maximum distance that will generate interaction in the known scene. Then, the step size of the spatial sequence is determined based on the accuracy requirements of building the scene library. With the range of values ​​and the step size, the dynamic spatial position of a scene element can be discretized and expanded.

[0033] Section 3: Expanding the Known Dynamic Element Velocity Sequence: In a dynamic scene, the scene library can be expanded by velocity sequences. By adding this dimension of velocity sequences, the overall dynamism of the scene is better demonstrated.

[0034] Section 4: Enlargement of Known Dynamic Element Behavior Sequences: In a dynamic scene, the behavior of scene elements, especially environmental vehicles, is often not a simple uniform motion, but a series of complex actions or even combinations of complex actions, such as overtaking, lane changing, and emergency braking. Based on the OpenDRIVE and OpenSCENARIO international autonomous driving scene modeling standards, the behavior of environmental vehicles in the scene is fully defined, and the behavior sequence of environmental vehicles in the scene is expanded according to the road environment.

[0035] In stages two and three, the known spatial and velocity sequences of dynamic elements only apply to the initial moment of the dynamic scene. After the initial moment, the scene elements move according to the known behavioral sequences of dynamic elements specified in stage four.

[0036] The specific steps of the third step, step one, are as follows:

[0037] Step 1: Designing Typical Known Unsafe Scenarios: Before learning and derivation, it is necessary to design a given number of typical known unsafe scenarios as the initial scenario data foundation to support subsequent self-learning and derivation work. Based on the test objects, the test scenarios are divided into four categories: autonomous driving perception layer test scenarios, autonomous driving planning layer test scenarios, autonomous driving control layer test scenarios, and autonomous driving system comprehensive test scenarios. According to the aforementioned classification, 100 unique scenarios are manually generated under each major category. For example, autonomous driving perception layer test scenarios generate a series of scenarios such as "ghost peek" scenarios, severe weather scenarios, and intersection scenarios with multiple traffic participants; autonomous driving decision layer test scenarios can generate a series of lane-changing scenarios, following scenarios, and scenarios of sudden braking by the vehicle in front; autonomous driving control layer test scenarios can generate a series of standard handling and stability test scenarios, icy and slippery road surface scenarios, and extreme working condition scenarios; and autonomous driving system comprehensive test scenarios can generate a series of regular urban road scenarios, regular rural road scenarios, and regular highway scenarios.

[0038] Step 2, Self-Learning Derivation: Based on self-learning theory, the scene derivation model is modeled to form a neural network with learning capabilities. First, a small number of typical known unsafe scenes generated in Step 1 are sampled for initialization. Random derivation is performed, and a deep reinforcement learning solver with adaptive stress testing is used to determine whether the derivation scenes are reasonable. Scene dissimilarity is estimated, and unreasonable scenes with large dissimilarity from other elements in the scene library are removed, while reasonable scenes with small dissimilarity are retained. This completes one scene library update. Then, the updated scene library is used as the initial scene library for sampling, derivation, testing, estimation, and updating in a loop.

[0039] The specific steps of the third step, step two, are as follows:

[0040] Step 1: Design typical known unsafe scenarios: Before deriving known scenarios, it is necessary to design a small number of typical known unsafe scenarios. The results of Step 1 in Step 3 are directly used as the initial scenario data basis to support subsequent self-extending derivation work.

[0041] Step Two: Self-Extending Rules. Self-extending rules include, but are not limited to, exhaustive self-extending rules and self-extending rules designed based on ontology theory. Among them, exhaustive self-extending rules are relatively simple. First, a typical known unsafe scenario designed in Step One is read. Then, the known scenario element library generated in Step One is read. By replacing the scene elements with similar attributes or functions in the typical known unsafe scenario with scene elements in the known scenario element library, a new scenario is derived. By exhaustively enumerating the scene elements in the scene element library by category, an exhaustive derivation with a large scene coverage is completed, realizing the self-extending derivation of known scenarios. Self-extending rules designed based on ontology theory first establish a knowledge representation model containing 5 levels: road level, traffic infrastructure level, temporary operation level, object level, and environment level. Specifically, using 284 classes, 762 logical axioms, and 75 semantic rules, an ontology-based automatic scenario derivation method can be realized to perform self-extending derivation of known scenarios.

[0042] The specific steps of step one in the fourth step are as follows:

[0043] Step 1: Discrete Calculation of Dynamic Scenarios: Based on the discrete time step of the scenario and the required test duration, calculate the relative positions of the main vehicle and traffic participants in the scenario at each moment during the dynamic process of the scenario. This position is used for the calculation of the hazard regression equation in Step 2. The calculation methods in this step include, but are not limited to, designing a program for iterative calculation and using mature autonomous driving simulation software for scenario calculation.

[0044] Step 2: Calculate the scene hazard regression equation: Calculate the scene hazard at each moment based on the regression equation for the dynamic scene discretely calculated in Step 1. To simplify the calculation and reduce the computational load, a safety distance is set. When there are no potentially dangerous traffic participants within the radius of the safety distance centered on the main vehicle, the moment is considered safe and the calculation is skipped. Only when there are potentially dangerous traffic participants within the safety distance is the scene hazard regression equation between the main vehicle and the traffic participant calculated. The calculation results are used for the search and judgment in Step 3.

[0045] Step 3: Searching for Dangerous Moments: After the scenario hazard regression equation for all discrete moments of each known scenario in the known scenario library is calculated, the discrete moments of each scenario are sorted in descending order according to the calculation results of the scenario hazard regression equation. The most dangerous moments in the scenario are searched, and these moments can help explain the unsafe factors of the scenario when used. In particular, if the top n dangerous moments obtained in this step are adjacent moments, they should be filtered out as duplicate data.

[0046] The specific steps of step two in the fourth step are as follows:

[0047] Step 1: Classifying Known Safe Scenarios: By setting scenario hazard thresholds, the most dangerous moment in a scenario is compared to see if the hazard threshold is met, thus distinguishing between safe and unsafe scenarios in the known scenarios. The known scenario library generated in the third step will be divided into two parts: a known safe scenario library and a known unsafe scenario library. The known safe scenario library is not the focus of the research and is therefore removed, leaving only the known unsafe scenario library.

[0048] Step 2: Classification of Known Unsafe Scenarios: Set up and train a classifier, and use clustering methods to cluster the known unsafe scenario library. Classify the known unsafe scenarios according to working conditions and hazard factors to facilitate the integration and use of the known unsafe scenario library.

[0049] The beneficial effects of this invention are:

[0050] 1) The method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory described in this invention can automatically expand the scenario library, thereby compressing the proportion of unknown scenarios in principle and effectively reducing the risks brought about by unknown unsafe scenarios.

[0051] 2) The method for constructing a library of known unsafe autonomous driving scenarios based on self-learning derivation theory provided by this invention can evaluate the safety of scenarios based on rules, automatically classify safe and unsafe scenarios, and provide an effective basis for selecting test scenarios for testing autonomous driving systems of intelligent connected vehicles.

[0052] 3) The autonomous driving known unsafe scenario library construction method based on self-learning derivation theory described in this invention has self-learning capability and can learn the features of known unsafe scenarios through deep reinforcement learning to perform self-learning derivation.

[0053] 4) The method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory provided by this invention has self-derived capabilities and can generalize and derive a series of known unsafe scenarios based on existing known unsafe scenarios in the scenario library.

[0054] 5) The method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory provided by this invention can provide a large number of known unsafe scenarios for simulation testing, which can greatly improve the effectiveness of simulation verification.

[0055] 6) The method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory provided by this invention can effectively identify security vulnerabilities in the tested autonomous driving system by increasing the coverage of test scenarios, and provide a comprehensive reference for its optimized design. Attached Figure Description

[0056] Figure 1 This is a schematic diagram illustrating the overall steps of the method for constructing a library of known unsafe scenarios for autonomous driving as described in this invention.

[0057] Figure 2 This is a schematic diagram of the overall architecture of the method for constructing a library of known unsafe scenarios for autonomous driving as described in this invention.

[0058] Figure 3 This is a schematic diagram of the overall architecture of the first step described in this invention.

[0059] Figure 4 This is a schematic diagram of the overall architecture of the second step described in this invention.

[0060] Figure 5 This is a schematic diagram of the overall architecture of the third step described in this invention.

[0061] Figure 6 This is a schematic diagram of the overall architecture of the fourth step described in this invention.

[0062] Figure 7 This is a schematic diagram of the regression equation output in step three of the second step of the present invention.

[0063] Figure 8 This is a flowchart of the third step algorithm described in this invention.

[0064] Figure 9 This is a diagram of the enhanced deep learning neural network structure described in step one of the third step of this invention.

[0065] Figure 10 This is a partial example diagram of the known unsafe scenario library in step three of the fourth step of the present invention. Detailed Implementation

[0066] Please see Figures 1 to 10 As shown:

[0067] The present invention provides a method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory, the method comprising the following steps:

[0068] Step 1: Expand the known scene elements;

[0069] The second step is to establish an interpretable scenario hazard assessment system.

[0070] The third step is to derive known unsafe scenarios;

[0071] Step 4: Build a library of known insecure scenarios;

[0072] The specific steps for expanding the known scene elements in the first step are as follows:

[0073] Step 1: Integration of Known Scene Element Library: Establish a library of scene elements appearing in known scenes, and integrate them according to various attributes of the scene elements to form a known scene element library, providing data support for the expansion of scene elements in Steps 2 and 3. Sources for expanding known scene elements include, but are not limited to, various traffic scene-related standards, or extraction from various online and offline dictionaries based on rules or neural networks.

[0074] Step Two: Expanding the Configuration of Known Static Elements: Static elements are defined as follows: In this step, static elements refer to the various objectively existing scene elements that exist within the scene during the scene design process. They exist as static elements when separated from a specific scene, but in a dynamic scene, they are not necessarily static. The specific steps in Step Two are as follows:

[0075] Step 1: Expanding the Known Static Element Categories: When establishing the known unsafe scenario library, a small number of typical known unsafe scenarios need to be input as a benchmark, such as ghost-jumping-out scenarios, severe weather scenarios, and sudden braking scenarios. These scenarios typically include different types of vehicles, pedestrians, and other animals, among other static elements. Even if these static elements are in the same spatial location, changes in their category attributes can severely affect the usability of the scenario. Therefore, the first step in expanding the known static element configuration is to expand based on the types of static elements, adding static elements with similar attributes to the same known static element sub-library, thus expanding to form multiple known static element sub-libraries.

[0076] Step Two: Expanding the Appearance of Known Static Elements: The appearance attributes of static elements, such as color and size, are extremely important parameters in a scene. There are two scenarios for this. First, the appearance of static elements is clearly defined. Typical examples include traffic cones, cars, and traffic signs. For these, appearance expansion can be based on relevant national standards or product information to comprehensively cover the various forms of these static elements in the scene. Second, the appearance of static elements is not clearly defined or cannot be defined. Typical examples include pedestrians, various animals, and roadside green belts. For these, static element expansion can be done by discretizing important attributes. For example, the average height of adult males in my country is between 160 and 185 centimeters. Therefore, the appearance size parameter of adult male static elements can be expanded with a step size of 5 centimeters, i.e., 160, 165, 170, 175, 180, and 185 centimeters. The step size chosen in this case directly affects the coverage of the scene library. In stages one and two, a known traffic participant database was formed from a database of known static elements.

[0077] Step 3: Expanding the Background Environment of Known Static Elements: A crucial aspect of static elements is the background environment, which includes lighting conditions, weather conditions, and road conditions. These static elements have a significant impact on scene safety and therefore need to be expanded. This expansion can primarily refer to autonomous driving-related standards, such as ISO PAS 21448. By expanding these static elements, a known background environment library can be formed, independent of any other known static element library.

[0078] Step 4: Further expansion of known static elements: Static elements can be expanded based on other criteria, such as parameters like the wear condition of lane lines. These attributes are parameters used for special processing of specific static elements and can form a special known static element sub-library as the parameter basis when expanding a single type of static element.

[0079] Step 3: Expanding the Spatiotemporal Sequence of Known Dynamic Elements: Dynamic elements are defined as follows. In specific scenarios, the static elements defined in Step 2 often undergo a dynamic process to interact with other static elements in the scene, thus forming an organic scene whole. This dynamic process is defined as the dynamic element of the scene. The specific steps in Step 3 are as follows:

[0080] Step 1: Expanding the Time Series of Known Dynamic Elements: The concept of time series does not exist in real-world scenarios. However, for simulation test scenarios, since simulation calculations are performed based on time discreteness, it is necessary to finely divide the simulation time when designing a library of known unsafe scenarios in order to conduct detailed analysis of unsafe factors. Therefore, the first step in expanding the spatiotemporal series of known dynamic elements is to expand the scenarios based on time series.

[0081] Step Two: Expansion of Known Dynamic Element Spatial Sequence: In a scene, the relative positional relationships between different scene elements are one of the key factors in the interaction between various scene elements. The spatial position of the test vehicle and its surrounding scene elements also directly affects the safety of the known scene. Therefore, the most important step in expanding the spatiotemporal sequence of known dynamic elements is to expand the scene based on the spatial sequence. The range of values ​​for the spatial sequence is determined based on the maximum distance that will cause interaction in the known scene. Then, the step size of the spatial sequence is determined based on the accuracy requirements of building the scene library. With the range of values ​​and the step size, the dynamic spatial position of a scene element can be discretized and expanded.

[0082] Step 3: Expanding the known dynamic element velocity sequence: In a dynamic scene, the scene library can be expanded by velocity sequence. By adding the dimension of velocity sequence, the overall dynamics of the scene can be better displayed.

[0083] Step 4: Expanding Known Dynamic Element Behavior Sequences: In a dynamic scene, the behavior of scene elements, especially environmental vehicles, is often not a simple uniform motion, but a series of complex actions or even combinations of complex actions, such as overtaking, lane changing, and emergency braking. Based on the OpenDRIVE and OpenSCENARIO international autonomous driving scene modeling standards, the behavior of environmental vehicles in the scene can be fully defined, and the behavior sequence of environmental vehicles in the scene can be expanded according to the road environment.

[0084] Furthermore, the spatial and velocity sequences of the known dynamic elements in stages two and three should only apply to the initial moment of the dynamic scene, while the activities of scene elements after the initial moment should be based on the behavioral sequences of the known dynamic elements specified in stage four.

[0085] Step 4: Expansion of Typical Physical Entity Elements: Steps 2 and 3 respectively established a known static element library and a known dynamic element library. The known static element library includes a known traffic participant library, a known background environment library, and a special known static element sub-library. The known dynamic element library includes time series, spatial series, velocity series, and behavioral series. By arranging and combining these related sub-libraries, the known scene element library can be expanded. Subsequently, based on rules, unreasonable combinations of scene elements are removed to complete the expansion of typical physical entity elements and form a reasonable known scene element library.

[0086] The specific steps for establishing an interpretable scenario hazard assessment system in the second step are as follows:

[0087] Step 1: Scene Hazard Characterization Coefficients. A reasonable assessment of scene safety is essential for correctly identifying known unsafe scenes from existing ones. Therefore, an interpretable scene hazard assessment system is needed. First, a characterization method, namely scene hazard characterization coefficients, needs to be proposed based on the factors that may cause scene unsafety. Then, a series of scene hazard characterization coefficients are extracted from the scenes based on the requirements of the assessment model and input into the model.

[0088] Step Two: Scene Hazard Assessment Model: Based on the scene hazard characterization coefficients from Step One, an interpretable assessment model is needed to comprehensively evaluate the hazard factors in a scene in order to assess its hazard level. Currently, the mainstream scene hazard assessment theories mainly include two types: physical model-based methods and data-driven machine learning-based methods. Due to the lack of interpretability of machine learning-based methods, when used as a scene hazard assessment model for a known unsafe scene library, there is a possibility of incomprehensible false positives and false negatives. Therefore, this patent adopts a physical model-based method as the scene hazard assessment model for a known unsafe scene library.

[0089] Step 3: Scene Hazard Regression Equation: In Step 2, a scene hazard assessment model based on a physical model was determined, mainly including the safe time model and the THW (time before head) model. To comprehensively detect the hazard level of a scene, a scene hazard regression equation needs to be designed to integrate the two models, forming a unified index and highlighting the hazard factors.

[0090] The scene hazard assessment model based on the physical model mentioned in step two includes the following two models:

[0091] Model 1: Safe Time Model. Time to Collision (TTC) refers to the remaining time before the vehicle collides with any other road obstacle. The formula for calculating TTC is as follows:

[0092] TTC=D / (V e -V f )

[0093] In the formula, D represents the distance between the main vehicle and the traffic participant ahead; V e V represents the speed of the main vehicle; f This indicates the speed of the traffic participant ahead. This algorithm can be used to calculate the collision time of traffic participants on the same lane.

[0094] For traffic participants not in the same lane, the collision time calculation formula changes to: TTC = D c / V e

[0095] In the formula, D c V represents the distance between the expected collision point of the two vehicles and the current position of this vehicle; e This is the speed of the vehicle.

[0096] Extensive statistical data analysis in human-machine interface research indicates that TTC = 2.5s is an effective threshold: providing a warning when TTC > 2.5s can effectively help drivers avoid impending forward collisions. Therefore, TTC = 2.5s can be considered the model boundary for distinguishing between safe and unsafe scenarios.

[0097] The safe time described in this model is an extrapolation of the collision time. In known unsafe scenarios, in addition to collision as the main hazard, there are other non-mainstream hazard factors such as falling. Therefore, the collision time is extrapolated as the safe time, which is defined as the remaining time after the following vehicle comes into contact with the hazard in front. The calculation method is exactly the same as the collision time.

[0098] Model 2: THW (Headway) Model. Headway represents the time difference between the front ends of two vehicles passing the same point. The calculation formula is as follows:

[0099] THW = D / V r

[0100] In the formula, D represents the distance between the front and rear of the vehicles; V r Indicates the speed of the vehicle behind.

[0101] THW is an important indicator for evaluating driving safety. It is closely related to traffic flow composition and driving behavior, and is an important basis for reflecting road capacity and service level. It can effectively evaluate the hazard level of a non-uniform speed dynamic scenario, such as a sudden braking scenario of a vehicle in front in urban conditions.

[0102] Based on the two scenario hazard assessment models in step two, the scenario hazard characterization coefficients that need to be extracted and calculated in step one include the vehicle speed sequence, the lateral and longitudinal speed and position sequences of surrounding traffic participants relative to the vehicle, and the lateral and longitudinal position sequences of surrounding obstacles and other hazard factors relative to the vehicle.

[0103] In order to better integrate the risk factors in the scenario and highlight the scenario's insecurity, the scenario risk regression equation defined in step three is expressed as follows:

[0104] SD = (1 - 0.8 * e) mTs-2.5 -0.8·e mTHW-4 ) / (1-0.8·e -2.5 -0.8·e -4 )

[0105] In the formula, SD represents the overall hazard level of the scenario; when SD is greater than 0, the scenario is unsafe. mTS represents the minimum safe time at that moment in the discrete-time dynamic scenario process. mTHW represents the minimum headway at that moment in the discrete-time dynamic scenario process. The formulas for calculating mTS and mTHW are as follows: mTS = min{TS1, TS2, TS3…}

[0106] mTHW=min{THW1, THW2, THW3...}

[0107] In the formula TS i This represents the safe time between the i-th scene element and the main vehicle; THW i This represents the time distance between the i-th scene element and the front of the main vehicle.

[0108] The above equation is only a special case of the scenario hazard regression equation. Depending on the usage requirements, scenario preferences, and the strictness of the judgment, a scenario hazard regression equation with combined requirements can be redesigned.

[0109] In the third step, known unsafe scenarios were derived through two different approaches. The derived scenarios were known scenarios with broad coverage. The element base for these derived scenarios was the known scenario element library generated in the first step. This known scenario library serves as a candidate for the known unsafe scenario library. After derivation, it is sent to the fourth step for screening and verification before being integrated to form the known unsafe scenario library. The specific steps for deriving known unsafe scenarios are as follows:

[0110] Step 1: Scene Construction Based on Self-Learning Theory: This method is primarily based on machine learning methods such as reinforcement learning and Bayesian networks. It trains a computer to learn from a small number of existing but typical known insecure scenarios, and based on this learning, derives a series of known scenarios. The specific steps of Step 1 are as follows:

[0111] Step 1: Design Typical Known Unsafe Scenarios: Before proceeding with learning and derivation, a small number of typical known unsafe scenarios need to be designed as initial scenario data to support subsequent self-learning and derivation work. Based on the test objects, test scenarios can be divided into four categories: autonomous driving perception layer test scenarios, autonomous driving planning layer test scenarios, autonomous driving control layer test scenarios, and autonomous driving system comprehensive test scenarios. According to the above classification, 100 unique scenarios are artificially generated under each major category. For example, autonomous driving perception layer test scenarios can generate a series of scenarios such as "ghost peek" scenarios, severe weather scenarios, and intersection scenarios with multiple traffic participants; autonomous driving decision-making layer test scenarios can generate a series of scenarios such as lane changing scenarios, following scenarios, and scenarios of sudden braking by the vehicle in front; autonomous driving control layer test scenarios can generate a series of standard handling and stability test scenarios, icy and slippery road surface scenarios, and extreme working condition scenarios; and autonomous driving system comprehensive test scenarios can generate a series of scenarios such as conventional urban road scenarios, conventional rural road scenarios, and conventional highway scenarios.

[0112] Step Two, Self-Learning Derivation: Based on self-learning theory, the scene construction model is modeled as a neural network with learning capabilities. First, a small number of typical known unsafe scenes generated in Step One are sampled for initialization. Random derivation is performed, and a deep reinforcement learning solver with adaptive stress testing is used to determine whether the derived scenes are reasonable. Scene dissimilarity is estimated, and unreasonable scenes with large dissimilarity from other elements in the scene library are removed, while reasonable scenes with small dissimilarity are retained. This completes one scene library update. Then, the updated scene library is used as the initial scene library for sampling, derivation, testing, estimation, and updating in a loop.

[0113] Step Two: Scene Construction Based on Self-Extension Theory: The theoretical basis of this method is mainly ontology. It generates all combinations of a small number of typical known unsafe scenes in the current parameter space through direct permutation and combination of parameters. The specific steps of Step Two are as follows:

[0114] Step 1: Design typical known unsafe scenarios. Similar to the scenario construction based on self-learning theory in Step 1, before deriving known scenarios, it is necessary to design a small number of typical known unsafe scenarios. Here, the results of Step 1 in Step 1 can be directly used as the initial scenario data basis to support subsequent self-extending and derivation work.

[0115] Step Two: Self-Extending Rules: The self-extending rules described in this patent include, but are not limited to, exhaustive self-extending rules and self-extending rules designed based on ontology theory. The exhaustive self-extending rules are relatively simple. First, a typical known unsafe scenario designed in Step One is read. Then, the known scenario element library generated in Step One is read. New scenarios are derived by replacing scene elements with similar attributes or functions in the typical known unsafe scenario with scene elements from the known scenario element library. Furthermore, through exhaustive derivation of scene elements in the scene element library by category, achieving a large scene coverage, the self-extending derivation of known scenarios is realized. The self-extending rules designed based on ontology theory first establish a knowledge representation model containing five levels: road level, traffic infrastructure level, temporary operation level, object level, and environment level. Specifically, using 284 classes, 762 logical axioms, and 75 semantic rules, an ontology-based automatic scenario derivation method can be implemented to perform self-extending derivation of known scenarios.

[0116] The self-learning derivative strategies described in step one are attached. Figure 8 As shown.

[0117] Adaptive stress testing is primarily used to simulate the complex computational problem of unsafe scenario simulation, treating it as a continuous decision-making problem by treating it as a black-box simulation. Setting up a black-box simulator S and a subset of interest E in the state space, the problem is transformed into finding the trajectory (s0, s1, s2, ... s...). n The trajectory ends at E, meaning that for a given S and E, the adaptive stress test can be expressed as a maximization problem as shown in the following formula:

[0118] maximizeP(a0,a1,a2,…a t ,s0,s1,s2,…s t )(s t ∈E)

[0119] In the formula, P(a0,a1,a2,…a…) t ,s0,s1,s2,…s t ) indicates that in the case of a t The parameters applied to the trajectory (s0,s1,s2,…s) generated by the S-black-box simulator t The probability of ).

[0120] The adaptive stress test setup mainly consists of three functions: an initialization function, a step function, and an endpoint determination function. The initialization function initializes the black-box simulator S according to certain rules and obtains its initial state s0; the step function determines the endpoint based on parameter a... i and current state s i The next step in calculating the trajectory is the next state s of S. i+1And calculate P(a0,a1,a2,…a…) i ,s0,s1,s2,…s i The endpoint determination function is used to determine whether the trajectory leads to the subset of interest E and whether E has been reached.

[0121] The aforementioned step function serves as the computational core of the self-learning derived strategy. It is calculated using a deep reinforcement learning neural network combining an LSTM (Long Short-Term Memory) model and an RNN (Recurrent Neural Network). The network structure is shown in the attached figure. Figure 9 As shown.

[0122] The above steps one and two are cyclical. The known scene sub-library derived in step one can serve as the self-expansion data foundation required in step two, and similarly, the known scene sub-library derived in step two can serve as the self-learning data foundation required in step one. This iterative process continues until scene construction based on self-learning and self-expansion theories can no longer effectively generate scenes, at which point the iteration is complete, forming a known scene library, which is then fed into step four.

[0123] The specific steps for building the known insecure scenario library in step four are as follows:

[0124] Step 1: Derivative Scenario Risk Assessment: Based on the interpretable scenario risk assessment system designed in Step 2, the risk of known scenarios generated in Step 3 can be assessed. Step 1 is detailed below:

[0125] Step 1: Discrete Calculation of Dynamic Scenarios. Based on the discrete time step of the scenario and the required test duration, calculate the relative positions of the main vehicle and traffic participants in the scenario at each moment during the dynamic process. This position is used for the calculation of the hazard regression equation in Step 2. The calculation methods in this step include, but are not limited to, iterative calculations using designed programs and scenario calculations using mature autonomous driving simulation software.

[0126] Step 2: Solving the Scene Hazard Regression Equation: The scene hazard is calculated based on the regression equation for each moment of the dynamic scene discretely calculated in Step 1. To simplify the calculation and reduce the computational load, a safety distance is set. When there are no potentially dangerous traffic participants within the radius of the safety distance centered on the main vehicle, the moment is considered safe and the calculation is skipped. Only when there are potentially dangerous traffic participants within the safety distance is the scene hazard regression equation between the main vehicle and the traffic participant calculated. The calculation results are used for the search and judgment in Step 3.

[0127] Step 3: Searching for Critical Moments: After calculating the scenario hazard regression equation for all discrete moments of each known scenario in the known scenario library, the discrete moments of each scenario are sorted in descending order according to the calculation results of the scenario hazard regression equation. This searches for the most dangerous moments in the scenario, which can help explain the unsafe factors of the scenario during use. In particular, if the top n dangerous moments obtained in this step are adjacent moments, they should be filtered out as duplicate data.

[0128] Step Two: Derived Scenario Pattern Classification: This step involves identifying and classifying the known scenario library generated in Step Three. First, it distinguishes between secure and insecure scenarios within the known scenarios, thus defining the scope of the known insecure scenario library. Then, it uses clustering to further classify the defined known insecure scenario library, facilitating its integration and use. Step Two is detailed below:

[0129] Step 1: Classifying Known Safe Scenarios. By setting scenario hazard thresholds and comparing whether the most dangerous moment in a scenario meets the threshold, safe and unsafe scenarios are distinguished from known scenarios. The known scenario library generated in Step 3 will be divided into two parts: a known safe scenario library and a known unsafe scenario library. The known safe scenario library is not the focus of the research and is therefore removed, leaving only the known unsafe scenario library.

[0130] Step 2: Classification of Known Unsafe Scenarios: Set up and train a classifier, and use clustering methods to cluster the known unsafe scenario library. Classify the known unsafe scenarios according to working conditions and hazard factors to facilitate the integration and use of the known unsafe scenario library.

[0131] Step 3: Integration of the Known Unsafe Scenario Library: The known unsafe scenario library obtained in Step 2 is sorted and organized according to the classification results, numbered, and recorded. Recording methods include, but are not limited to, Excel spreadsheets or XML extensible tag language files, ensuring the readability of the established known unsafe scenario library. The generated known unsafe scenario library can also be converted according to the OpenSCENARIO international autonomous driving test scenario standard to generate *.xodr or *.osgb files, ensuring good extensibility of the established known unsafe scenario library.

Claims

1. A method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory, characterized in that: The method includes the following steps: The first step is to expand the known scene elements. The specific steps are as follows: Step 1: Integration of Known Scene Element Library: Establish a library of scene elements appearing in known scenes, and integrate them according to various attributes of scene elements to form a known scene element library. This library provides data support for the expansion of scene elements in Steps 2 and 3. The sources of the expansion of known scene elements include various traffic scene-related standards, or extraction from various online and offline dictionaries based on rules or neural networks. Step 2: Expanding the configuration of known static elements: Static elements are defined as follows: Static elements in this step refer to the various objectively existing scene elements in the scene when the scene is designed. They exist as static elements when separated from the specific scene, but they are not necessarily static in the specific dynamic scene. Step 3: Expanding the Spatiotemporal Sequence of Known Dynamic Elements: Dynamic elements are defined as follows. In specific scenarios, the static elements defined in Step 2 often undergo a dynamic process to interact with other static elements in the scene, thereby forming an organic scene whole. This dynamic process is defined as the dynamic element of the scene. Step 4: Expansion of Typical Physical Entity Elements: In Steps 2 and 3, a known static element library and a known dynamic element library were formed respectively. The known static element library includes a known traffic participant library, a known background environment library, and a special known static element sub-library. The known dynamic element library includes time series, spatial series, velocity series, and behavior series. By arranging and combining the above-mentioned related sub-libraries, the known scene element library can be expanded. Subsequently, unreasonable scene element combinations are deleted based on rules to complete the expansion of typical physical entity elements and form a reasonable known scene element library. The second step is to establish an interpretable scenario hazard assessment system, with the following specific steps: Step 1: Scene Hazard Characterization Coefficient: Only by reasonably judging the safety of a scene can known unsafe scenes be correctly extracted from known scenes. For this purpose, an interpretable scene hazard evaluation system needs to be established. First, a characterization method, namely scene hazard characterization coefficient, needs to be proposed based on the factors that may cause scene unsafety. Based on the needs of the evaluation model, a series of scene hazard characterization coefficients are extracted from the scene for input into the evaluation model. Step 2, Scene Hazard Assessment Model: Based on the scene hazard characterization coefficients from Step 1, an interpretable assessment model is needed to comprehensively evaluate the hazard factors in the scene in order to assess the hazard level of the scene. There are two main theories of scene hazard assessment: physical model-based methods and data-driven machine learning-based methods. Due to the lack of interpretability of machine learning-based methods, there is a possibility of incomprehensible false positives and false negatives when using them as scene hazard assessment models for known unsafe scene libraries. Therefore, physical model-based methods will be adopted as scene hazard assessment models for known unsafe scene libraries. Step 3: Scene Hazard Regression Equation: In Step 2, a scene hazard assessment model based on a physical model was determined, mainly including the collision time method and the headway method. In order to comprehensively detect the hazard of the scene, a scene hazard regression equation was designed to integrate the two models, form a unified index, and highlight the hazard factors. The third step involves deriving known unsafe scenarios. This step uses two different approaches to derive scenarios, resulting in known scenarios with broad coverage. The elements of these derived scenarios are based on the known scenario element library generated in the first step. This known scenario library serves as a candidate for the known unsafe scenario library. After deriving, it is sent to the fourth step for screening and verification before being integrated to form the known unsafe scenario library. The specific steps for deriving known unsafe scenarios are as follows: Step 1: Scene Construction Based on Self-Learning Theory: The theoretical basis of this method is the machine learning method of reinforcement learning and Bayesian networks. By training the computer to learn a small number of existing but typical known unsafe scenes, a series of known scenes are derived based on the learning training. Step 2: Scene construction based on self-extension theory: The theoretical basis of this method is mainly ontology. It generates all combinations of a small number of typical known unsafe scenes as the basis in the current parameter space by directly arranging and combining parameters. The above steps one and two are cyclical. That is, the known scene sub-library derived in step one can serve as the self-extending data base required in step two, and the known scene sub-library derived in step two can also serve as the self-learning data base required in step one. This cycle iterates until scene construction based on self-learning and self-extending theories can no longer effectively generate scenes, at which point the iteration is completed, a known scene library is formed, and it is sent to step four. Step 4: Construct a library of known insecure scenarios. The specific steps are as follows: Step 1: Derivative Scenario Risk Assessment: Based on the interpretable scenario risk assessment system designed in Step 2, the risk assessment of the known scenario library generated in Step 3 can be performed. Step 2, Derived Scenario Pattern Classification: In this step, the known scenario library generated in step 3 is identified and classified. First, safe and unsafe scenarios are distinguished from the known scenarios, thereby defining the scope of the known unsafe scenario library. Then, the defined known unsafe scenario library is classified by clustering to facilitate the integration and use of the scenario library. Step 3: Integration of the Known Unsafe Scenario Library: The known unsafe scenario library obtained in Step 2 is sorted and organized according to the classification results, numbered, and output as records. The recording method includes Excel spreadsheets or XML extensible tag language files, ensuring the readability of the established known unsafe scenario library. The generated known unsafe scenario library is then converted according to the OpenSCENARIO international autonomous driving test scenario standard, generating *.xodr or *.osgb files, ensuring the extensibility of the established known unsafe scenario library.

2. The method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory according to claim 1, characterized in that: The specific steps of the first and second steps are as follows: Step 1: Expanding the Known Static Element Types: When establishing the known unsafe scenario library, it is necessary to input a set number of typical known unsafe scenarios as a benchmark, including ghost peek scenarios, severe weather scenarios, and sudden braking scenarios. These scenarios include different types of static elements such as vehicles, pedestrians, and other animals. Even if these static elements are in the same spatial location, changes in their type attributes can seriously affect the usability of the scenario. Therefore, the first step in expanding the known static element configuration is to expand it according to the type of static element, adding static elements with similar attributes to the same known static element sub-library, thus expanding it into multiple known static element sub-libraries. Step Two: Expanding the Appearance of Known Static Elements: The appearance attributes of static elements, including color and size, are extremely important parameters in a scene. There are two scenarios for this: First, the appearance of static elements is clearly defined, including static elements such as traffic cones, cars, and traffic signs. For these types of static elements, the appearance expansion should be based on relevant national standards or product information to comprehensively cover the forms in which these static elements exist in the scene. Second, the appearance of static elements is not clearly defined or cannot be defined. The scenario includes static elements such as pedestrians, various animals, and roadside green belts. For these static elements, the static elements are expanded by discretizing important attributes. The height of adult males is generally in the range of 160 to 185 centimeters. Therefore, the appearance size parameters of the static elements of adult males are expanded by taking a step size of 5 centimeters, that is, 160, 165, 170, 175, 180, and 185 centimeters. The step size selected during the expansion of this case directly affects the coverage of the scenario library. In the first and second stages, a known traffic participant library of known static elements is formed. Step 3: Expanding the Background Environment of Known Static Elements: Another crucial aspect of static elements is the background environment, which includes lighting conditions, weather conditions, and road conditions. These static elements have a significant impact on the safety of the scene, so they also need to be expanded. The expansion process mainly refers to autonomous driving-related standards, including the ISOPAS 21448 standard. By expanding these static elements, a known background environment library is formed. This sub-library is independent of any other known static element sub-library. Step 4: Further expansion of known static elements: Static elements are also expanded based on other criteria, including lane line wear parameters. These attributes are parameters used for special processing of specific static elements, which can form a special known static element sub-library as the parameter basis when expanding a certain type of static element.

3. The method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory according to claim 1, characterized in that: The specific steps of the first step, step three are as follows: Step 1: Time Series Expansion of Known Dynamic Elements: The concept of time series does not exist in real-world scenarios. However, for simulation test scenarios, since simulation calculations are performed based on time discreteness, it is necessary to finely divide the simulation time when designing a library of known unsafe scenarios in order to conduct detailed analysis of unsafe factors. Therefore, the first step in expanding the spatiotemporal sequence of known dynamic elements is to expand the scenarios based on time series. Step 2: Expansion of Known Dynamic Element Spatial Sequence: In a scene, the relative positional relationship between different scene elements is one of the key factors in the interaction between various scene elements. The spatial position of the test vehicle and the surrounding scene elements also directly affects the safety of the known scene. Therefore, the most important step in expanding the spatiotemporal sequence of known dynamic elements is to expand the scene based on the spatial sequence. The range of values ​​for the spatial sequence is determined based on the maximum distance that will generate interaction in the known scene. Then, the step size of the spatial sequence is determined based on the accuracy requirements of building the scene library. With the range of values ​​and the step size, the dynamic spatial position of a scene element can be discretized and expanded. Section 3: Expanding the Known Dynamic Element Velocity Sequence: In a dynamic scene, the scene library can be expanded by velocity sequences. By adding this dimension of velocity sequences, the overall dynamism of the scene is better demonstrated. Section 4: Enlargement of Known Dynamic Element Behavior Sequences: In a dynamic scene, the behavior of environmental vehicles in the scene elements is often not a simple uniform motion, but a series of complex actions or even combinations of complex actions, including overtaking, lane changing and emergency braking. Based on the OpenDRIVE and OpenSCENARIO international autonomous driving scene modeling standards, the behavior of environmental vehicles in the scene is fully defined, and the behavior sequence of environmental vehicles in the scene is expanded according to the road environment. In stages two and three, the known spatial and velocity sequences of dynamic elements only apply to the initial moment of the dynamic scene. After the initial moment, the scene elements move according to the known behavioral sequences of dynamic elements specified in stage four.

4. The method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory according to claim 1, characterized in that: The specific steps of the third step, step one, are as follows: Step 1: Designing Typical Known Unsafe Scenarios: Before learning and derivation, it is necessary to design a given number of typical known unsafe scenarios as the initial scenario data foundation to support subsequent self-learning and derivation work. Based on the test objects, the test scenarios are divided into four categories: autonomous driving perception layer test scenarios, autonomous driving planning layer test scenarios, autonomous driving control layer test scenarios, and autonomous driving system comprehensive test scenarios. According to the aforementioned classification, 100 unique scenarios are manually generated under each major category. These include: autonomous driving perception layer test scenarios generating a series of scenarios such as "ghost peek" scenarios, severe weather scenarios, and intersection scenarios with multiple traffic participants; autonomous driving decision layer test scenarios generating a series of lane-changing scenarios, following scenarios, and scenarios of sudden braking by the vehicle in front; autonomous driving control layer test scenarios generating a series of standard handling and stability test scenarios, icy and slippery road surface scenarios, and extreme working condition scenarios; and autonomous driving system comprehensive test scenarios generating a series of regular urban road scenarios, regular rural road scenarios, and regular highway scenarios. Step 2, Self-Learning Derivation: Based on self-learning theory, the scene derivation model is modeled to form a neural network with learning capabilities. First, a small number of typical known unsafe scenes generated in Step 1 are sampled for initialization. Random derivation is performed, and a deep reinforcement learning solver with adaptive stress testing is used to determine whether the derivation scenes are reasonable. Scene dissimilarity is estimated, and unreasonable scenes with large dissimilarity from other elements in the scene library are removed, while reasonable scenes with small dissimilarity are retained. This completes one scene library update. Then, the updated scene library is used as the initial scene library for sampling, derivation, testing, estimation, and updating in a loop.

5. The method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory according to claim 1, characterized in that: The specific steps of the third step, step two, are as follows: Step 1: Design typical known unsafe scenarios: Before deriving known scenarios, it is necessary to design a small number of typical known unsafe scenarios. The results of Step 1 in Step 3 are directly used as the initial scenario data basis to support subsequent self-extending derivation work. Step Two: Self-Extending Rules: Self-extending rules include exhaustive self-extending rules and self-extending rules designed based on ontology theory. Among them, exhaustive self-extending rules are relatively simple. First, a typical known unsafe scenario designed in Step One is read. Then, the known scenario element library generated in Step One is read. By replacing the scene elements with similar attributes or functions in the typical known unsafe scenario with scene elements in the known scenario element library, a new scenario is derived. By exhaustively enumerating the scene elements in the scene element library by category, an exhaustive derivation with a large scene coverage is completed, realizing the self-extending derivation of known scenarios. Self-extending rules designed based on ontology theory first establish a knowledge representation model containing 5 levels: road level, traffic infrastructure level, temporary operation level, object level, and environment level. Using 284 classes, 762 logical axioms, and 75 semantic rules, an ontology-based automatic scenario derivation method can be realized to perform self-extending derivation of known scenarios.

6. The method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory according to claim 1, characterized in that: The specific steps of the fourth step, step one, are as follows: Step 1: Discrete Calculation of Dynamic Scenarios: Based on the discrete time step of the scenario and the required test duration, calculate the relative positions of the main vehicle and traffic participants in the scenario at each moment during the dynamic process of the scenario. This position is used for the calculation of the hazard regression equation in Step 2. The calculation method in this step includes designing a program for iterative calculation. Step 2: Calculate the scene hazard regression equation: Calculate the scene hazard at each moment based on the regression equation for the dynamic scene discretely calculated in Step 1. To simplify the calculation and reduce the computational load, a safety distance is set. When there are no potentially dangerous traffic participants within the radius of the safety distance centered on the main vehicle, the moment is considered safe and the calculation is skipped. Only when there are potentially dangerous traffic participants within the safety distance is the scene hazard regression equation between the main vehicle and the traffic participant calculated. The calculation results are used for the search and judgment in Step 3. Step 3: Searching for Dangerous Moments: After the regression equations for the scene hazard levels of all discrete moments in the known scene library are calculated, the discrete moments of each scene are sorted in descending order according to the results of the scene hazard regression equations. The most dangerous moments in the scene are then searched. These moments can help explain the unsafe factors of the scene when used. In this step, if the top n dangerous moments obtained by sorting are adjacent moments, they should be filtered out as duplicate data.

7. The method for constructing a library of known unsafe scenarios for autonomous driving based on self-learning derivation theory according to claim 1, characterized in that: The specific steps of the fourth step, step two, are as follows: Step 1: Classifying Known Safe Scenarios: By setting scenario hazard thresholds, the most dangerous moment in a scenario is compared to see if the hazard threshold is met, thus distinguishing between safe and unsafe scenarios in the known scenarios. The known scenario library generated in the third step will be divided into two parts: a known safe scenario library and a known unsafe scenario library. The known safe scenario library is not the focus of the research and is therefore removed, leaving only the known unsafe scenario library. Step 2: Classification of Known Unsafe Scenarios: Set up and train a classifier, and use clustering methods to cluster the known unsafe scenario library. Classify the known unsafe scenarios according to working conditions and hazard factors to facilitate the integration and use of the known unsafe scenario library.

Images