Automatic derivation method, system and equipment of automatic driving test scene and medium

Abstract

The invention relates to an automatic derivation method, system and device for an automatic driving test scene and a medium, and the method comprises the steps: analyzing a test demand of an automatic driving system, and mapping the test demand to a multi-dimensional scene element hierarchical architecture, so as to construct a multi-layer road scene; reasoning a scene condition dependency relationship by using a dual conditional probability, and generating an initial scene fragment set in combination with probability distribution sampling; establishing a multi-objective optimization function based on the scene type distribution weight, the element redundancy penalty factor, the physical consistency constraint and the security rule, screening and recombining the multiple initial scene fragment sets, and outputting an optimal scene set; extracting a causal logical relationship of the optimal scene set according to the directed acyclic graph, and generating a scene evolution path in combination with topological sorting; and executing multiple safety verification on the scene evolution path, and determining the optimal scene passing the verification as the automatic driving test scene. The technical problems that an existing test scene is low in generation efficiency, poor in authenticity and unreasonable in event chain logic are solved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to an automated derivation method, system, device, and medium for autonomous driving test scenarios. Background Art

[0002] In recent years, autonomous driving technology has developed rapidly, and its safety verification has become a key challenge for industrial implementation. As the core vehicle for evaluating the performance of autonomous driving systems, the efficiency, coverage, and authenticity of test scenarios directly impact the effectiveness and credibility of the tests. However, current mainstream test scenario generation methods face many limitations:

[0003] 1. Traditional methods struggle to systematically analyze complex, multi-dimensional testing requirements (such as extreme weather interactions and traffic rule conflict scenarios) and lack the ability to model road environment elements in a layered manner. Furthermore, existing frameworks often organize scenario elements in a flat or hand-coded manner, resulting in inefficient and poorly adaptable scenario construction.

[0004] 2. Existing technologies do not deeply capture the constraint relationships between inter-layer elements (such as the boundary restrictions of road structures on vehicle trajectories) and the interactive dependencies between intra-layer elements (such as the probability of coordinated lane changes by multiple vehicles). They mainly rely on random sampling or scene fragments generated by predefined rules. They are difficult to simulate probabilistic interactions such as sudden behaviors (pedestrians peeking out) and complex decisions (unprotected left turns) in the real world, resulting in insufficient scene realism and poor excitability.

[0005] 3. A complete test scenario must include a continuous chain of event evolution. However, current technology lacks explicit modeling of causal relationships between events (such as the sudden braking of the front vehicle causing the rear vehicle to rear-end the rear vehicle) and timing logic (the traffic light changes before the vehicle starts). This results in logical jumps or unreasonable event chains in the generated scenario sequence, making it difficult to reproduce the gradual evolution process of a real accident.

[0006] Therefore, those skilled in the art are in urgent need of a scenario automation derivation method that can realize hierarchical modeling of multi-dimensional elements, dynamic coupling of constraint relationships, and explicit driving of temporal logic. Summary of the Invention

[0007] (1) Technical issues to be resolved

[0008] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method, system, device and medium for automated derivation of autonomous driving test scenarios, which solve the technical problems that the prior art has difficulty in implementing multi-dimensional factor hierarchical modeling, cannot dynamically couple constraint relationships and interactive dependencies, and lacks explicit driving of temporal logic and causal relationships, resulting in low scenario generation efficiency, poor authenticity, and unreasonable event chain logic.

[0009] (2) Technical solution

[0010] In order to achieve the above objectives, the main technical solutions adopted by the present invention include:

[0011] In a first aspect, an embodiment of the present invention provides a method for automatically deriving an autonomous driving test scenario, comprising:

[0012] Analyze the test requirements of the target autonomous driving system and map the analysis results to a preset multi-dimensional scenario element layered architecture to construct a multi-level road scenario that responds to external environment changes and test requirements;

[0013] Dual conditional probability is used to reason about the conditional dependencies between and within layers of multi-layer road scenes, and combined with probability distribution sampling, an initial set of scene fragments is generated that represents the dynamic interactions of multi-layer elements.

[0014] A multi-objective optimization function is established based on the set scene type distribution weights, element redundancy penalty factors, physical consistency constraints, and safety rules. Multiple initial scene fragment sets are screened and reorganized to output the optimal scene set that covers the target distribution and satisfies all constraints.

[0015] Extract the causal logic relationship of the optimal scenario set based on the directed acyclic graph, and perform topological sorting on the causal logic relationship to generate a scenario evolution path that conforms to the causal temporal logic;

[0016] Multiple safety verifications combining virtual data and real data are performed on the scenario evolution path, and the optimal scenario corresponding to the verified scenario evolution path is determined as a valid autonomous driving test scenario.

[0017] Optionally, the test requirements of the target autonomous driving system are analyzed, and the analysis results are mapped to a preset multi-dimensional scenario element layered architecture to construct a multi-level road scenario that responds to external environment changes and test requirements, including:

[0018] Analyze the test requirements of the target autonomous driving system and retrieve all scenario elements that meet the test requirements from the preset database based on the analysis results;

[0019] Mapping all scene elements to the corresponding hierarchical structure of the multi-dimensional scene element layered architecture to obtain an initial multi-level road scene;

[0020] The initial multi-level road scene is dynamically adjusted to environmental changes through the set environmental change timing parameters to obtain a multi-level road scene that responds to external environmental changes and test requirements in real time;

[0021] Among them, the multi-level road scene includes the static environment layer, dynamic participation layer, environmental condition layer and road facility layer.

[0022] Optionally, mapping all scene elements to the hierarchical structure of the corresponding multi-dimensional scene element layered architecture to obtain an initial multi-level road scene includes:

[0023] Road topology modeling is performed based on the test site map information in the scenario elements to generate a static environment layer, which is represented by an editable digital road network model.

[0024] Using probabilistic graphical models and reinforcement learning algorithms, combined with historical accident data from the test site, we generate risk dynamic scenarios and inject them into the digital road network model to build a dynamic participation layer.

[0025] Dynamically couple and adjust all environmental elements in the scene elements through the preset physical engine to generate environmental condition data, and then inject the environmental condition data into the digital road network model to construct the environmental condition layer;

[0026] Establish a digital twin of the roadside equipment in the test site, and simulate the roadside equipment signals through digital twin interaction to build the road facility layer.

[0027] Optionally, the conditional dependencies between and within layers of a multi-layer road scene are inferred using dual conditional probability, and combined with probability distribution sampling, to generate an initial scene segment set representing the dynamic interaction of multi-layer elements, including:

[0028] The dynamic Bayesian network is used to model the time-varying conditional dependencies and probabilistic constraint chains of upper-level elements on lower-level elements in multi-level road scenes, and a cross-level conditional probability distribution table is obtained.

[0029] By using Markov random fields to model the high-dimensional spatiotemporal correlation and conditional dependence between elements at the same level in multi-level road scenes based on physical constraints, traffic rules, and scene semantics, we obtain a joint probability energy function between elements that represents spatial proximity and temporal coherence.

[0030] Based on the cross-level conditional probability distribution table and the joint probability energy function between elements, the scene elements of the multi-level road scene are sampled with probability distribution to generate an initial scene fragment set with dynamic interaction of multi-level elements.

[0031] Optionally, a multi-objective optimization function is established based on the set scene type distribution weights, element redundancy penalty factors, physical consistency constraints, and safety rules to screen and reorganize multiple initial scene fragment sets. The output of the optimal scene set that covers the target distribution and satisfies all constraints includes:

[0032] The set scene type distribution weights and element redundancy penalty factors are integrated to construct the initial multi-objective optimization function;

[0033] Inject physical consistency constraints and safety rules as mandatory constraints of the initial multi-objective optimization function to form a comprehensive multi-objective optimization function with multi-dimensional verification capabilities;

[0034] Based on a comprehensive multi-objective optimization function, it drives the collaborative screening and evolutionary reorganization of multiple initial scene fragment sets until a Pareto optimal scene set is dynamically generated and output that covers the predetermined type of distribution targets, meets strong constraints of physical consistency and safety, and suppresses information redundancy.

[0035] Optionally, extracting the causal logic relationship of the optimal scenario set based on the directed acyclic graph, and topologically sorting the causal logic relationship to generate a scenario evolution path that conforms to the causal temporal logic includes:

[0036] The causal logic relationship of the optimal scenario set obtained based on directed acyclic graph analysis is modeled as state transition constraint rules, and a temporal logic model with causal constraint capabilities is constructed based on the state transition constraint rules;

[0037] Using the topological sorting results of the directed acyclic graph, a mandatory occurrence sequence of event elements in the optimal scenario set is generated. This sequence is then input into the temporal logic model as the core driving source to simulate the dynamic change process of the temporal logic model state in the temporal dimension.

[0038] Capture and analyze the sequence data of temporal state transitions generated during dynamic changes in real time, verify the sequence data based on state transfer constraint rules and preset temporal logic requirements, and output the scenario evolution path that meets all causal logic constraints and temporal constraints based on the verification results.

[0039] Optionally, performing multiple safety verifications on the scenario evolution path by combining virtual data and real data, and determining the optimal scenario corresponding to the verified scenario evolution path as a valid autonomous driving test scenario includes:

[0040] In the digital twin environment, the logical consistency and behavioral safety of the scenario evolution path are verified through the set temporal logic, and the logical safety score of the scenario evolution path is calculated based on the verification results;

[0041] In a simulation environment that is independent of the physical execution environment, the rule responsiveness of the scenario evolution path is tested using the set static traffic rules, and the rule compliance of the scenario evolution path is calculated based on the test results.

[0042] Perform a multimodal comparison between the actual drive test data of the scenario evolution path and the preset standard values, and calculate the physical feasibility of the scenario evolution path based on the comparison results;

[0043] When the logical safety score, rule compliance and physical feasibility of the scenario evolution path all meet the set thresholds, the optimal scenario corresponding to the scenario evolution path is determined as a valid autonomous driving test scenario.

[0044] In a second aspect, an embodiment of the present invention provides an automated derivation system for autonomous driving test scenarios, including:

[0045] The scenario requirement analysis and architecture mapping module is used to analyze the test requirements of the target autonomous driving system and map the analysis results to a preset multi-dimensional scenario element layered architecture to construct a multi-level road scenario that responds to external environment changes and test requirements;

[0046] The scene fragment generation module is used to use dual conditional probability reasoning to determine the conditional dependencies between and within layers of multi-layer road scenes, and combined with probability distribution sampling to generate an initial set of scene fragments that represent the dynamic interactions of multi-layer elements;

[0047] The multi-objective scenario optimization module is used to establish a multi-objective optimization function based on the set scenario type distribution weights, element redundancy penalty factors, physical consistency constraints, and safety rules. It screens and reorganizes multiple initial scenario fragment sets and outputs the optimal scenario set that covers the target distribution and satisfies all constraints.

[0048] The causal temporal path synthesis module is used to extract the causal logical relationship of the optimal scenario set based on the directed acyclic graph, and topologically sort the causal logical relationship to generate a scenario evolution path that conforms to the causal temporal logic;

[0049] The hybrid data security verification module is used to perform multiple security verifications on the scenario evolution path by combining virtual data and real data, and determine the optimal scenario corresponding to the verified scenario evolution path as a valid autonomous driving test scenario.

[0050] In a third aspect, an embodiment of the present invention provides an electronic device, including:

[0051] processor;

[0052] A memory storing steps of an automated derivative method for controlling the autonomous driving test scenario described above by a processor.

[0053] In a fourth aspect, an embodiment of the present invention provides a computer-readable medium having computer-executable instructions stored thereon. When the executable instructions are executed by a processor, the automated derivation method steps of the autonomous driving test scenario described above are implemented.

[0054] (3) Beneficial effects

[0055] The beneficial effects of the present invention are:

[0056] First, the present invention constructs a basic road scene with multi-level and multi-dimensional characteristics by analyzing and mapping the test requirements to a multi-dimensional scene element layered framework, which flexibly responds to changes in the external environment and diversified test targets.

[0057] Next, the present invention uses dual conditional probability to model inter-layer / intra-layer dependencies, which can deeply model complex dynamic interactive behaviors in the real world, and combines probabilistic sampling to generate high-fidelity initial scenes, further improving the realism and complexity of the generated scenes.

[0058] Then, the present invention integrates scenario distribution, redundancy penalty, physical constraints and safety rules to output a comprehensive and compliant optimal scenario set, ensuring that the generated scenarios have the characteristics of high quality, wide coverage and low redundancy.

[0059] Finally, the present invention performs a multiple security verification mechanism combining virtual data and real data on the generated scenario evolution path, which can evaluate the validity and boundaries of the scenario from different dimensions, significantly improving the security and excitability credibility of the test scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 A flowchart of an automated derivation method for an autonomous driving test scenario provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation methods in conjunction with the accompanying drawings.

[0062] refer to Figure 1 As shown, an embodiment of the present invention proposes an automated derivation method for autonomous driving test scenarios, which includes: parsing the test requirements of a target autonomous driving system and mapping the parsing results to a preset multi-dimensional scene element hierarchical architecture to construct a multi-level road scenario that responds to external environmental changes and test requirements; utilizing dual conditional probability reasoning to determine the conditional dependencies between and within the layers of the multi-level road scenario, and combining this with probability distribution sampling to generate an initial scene segment set that represents the dynamic interaction of the multi-level elements; establishing a multi-objective optimization function based on predefined scene type distribution weights, element redundancy penalty factors, physical consistency constraints, and safety rules to screen and reorganize multiple initial scene segment sets and output an optimal scene set that covers the target distribution and satisfies all constraints; extracting causal logical relationships within the optimal scene set based on a directed acyclic graph, and topologically sorting the causal logical relationships to generate a scene evolution path that conforms to causal temporal logic; performing multiple safety verifications on the scene evolution path using a combination of virtual data and real data, and determining the optimal scene corresponding to the verified scene evolution path as a valid autonomous driving test scenario.

[0063] This embodiment constructs a basic road scenario with multi-level and multi-dimensional characteristics by parsing and mapping the test requirements to a multi-dimensional scenario element layered framework, which flexibly responds to changes in the external environment and diversified test objectives. Next, this embodiment also uses dual conditional probability modeling to model inter-layer / intra-layer dependencies, which can deeply model complex dynamic interactive behaviors in the real world, and combines probabilistic sampling to generate high-fidelity initial scenarios, further improving the realism and complexity of the generated scenarios. Then, this embodiment integrates scenario distribution, redundancy penalties, physical constraints and safety rules to output a comprehensive and compliant optimal scenario set, ensuring that the generated scenarios are of high quality, wide coverage and low redundancy. Finally, this embodiment performs a multiple security verification mechanism combining virtual data and real data on the generated scenario evolution path, which can evaluate the validity and boundaries of the scenario from different dimensions, significantly improving the safety and excitability credibility of the test scenario.

[0064] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0065] Specifically, refer to Figure 1 As shown, an embodiment of the present invention provides an automated derivation method for autonomous driving test scenarios, which includes:

[0066] S100: Analyze the test requirements of the target autonomous driving system and map the analysis results to a preset multi-dimensional scenario element layered architecture to construct a multi-level road scenario that responds to external environment changes and test requirements.

[0067] This embodiment achieves precise understanding and structured expression of complex requirements by parsing and mapping the target autonomous driving system's test requirements into a multi-dimensional, hierarchical architecture of scenario elements. This mapping mechanism also flexibly responds to changes in the external environment and diverse test objectives, automatically constructing a multi-layered, multi-dimensional foundational road scenario, laying the structural foundation for the efficient and accurate generation of subsequent test scenarios.

[0068] In this embodiment, step S100 may further include the following sub-steps S110 to S130:

[0069] S110: Analyze the test requirements of the target autonomous driving system, and retrieve all scenario elements that meet the test requirements from a preset database based on the analysis results.

[0070] For example, when an autonomous driving system requires special road stress testing, the test requirement is to verify the vehicle's emergency avoidance capabilities and chassis control coordination on complex roads or with sudden obstacles. After analyzing the test requirements, the scenario elements that meet the test requirements are obtained, including vehicle test function requirements, road characteristics, and dynamic factors. Vehicle test function requirements include: emergency braking (AEB), trajectory replanning, and the timing of electronic stability control system (ESP) intervention; road characteristics: Class IV highway (road surface width ≤ 6.5m, design speed ≤ 40km / h); dynamic factors: oncoming agricultural vehicles (speed 15-30km / h, irregular lane changes).

[0071] S120 : Map all scene elements to the corresponding hierarchical structure of the multi-dimensional scene element layered architecture to obtain an initial multi-level road scene.

[0072] Furthermore, step S120 may further include the following sub-steps S121 to S124:

[0073] S121. Perform road topology modeling based on the test site map information in the scene elements to generate a static environment layer, where the static environment layer is represented by an editable digital road network model.

[0074] In this embodiment, after obtaining high-precision map data of the test site (such as high-definition raster maps, vector maps or OpenDRIVE / OSM format files), map parsing tools are used to read element information such as road structure (lane lines, intersection outlines, lane boundaries, etc.), buildings (office buildings, residential areas), and natural obstacles (mountains, rivers, green belts, traffic islands). Then, these elements are converted into precise geometric representations (points, lines, and surfaces). Secondly, the connection relationship of the roads is set, for example, the exit of lane A is connected to the entrance of lane B. Then, the road attributes are set, including lane width, curvature, slope, speed limit, traffic light location, etc. For example, the map shows an intersection with a main road consisting of four lanes and a branch road consisting of two lanes. Finally, an editable digital road network model is output. The static environment layer is represented by an editable digital road network model, allowing users to modify it directly on this digital model. The digital road network model built once can be used for countless different dynamic scene combinations.

[0075] S122. Utilize probabilistic graphical models and reinforcement learning algorithms, combined with historical accident data from the test site, to generate risk dynamic scenarios, which are then injected into the digital road network model to construct a dynamic participation layer.

[0076] In this embodiment, first, a probabilistic graphical model is used to analyze historical accident data: factors related to the accident are identified; the conditional probability relationship between these factors is calculated, for example, at specific intersections on rainy days and during evening rush hours, the probability of a vehicle suddenly braking and resulting in a rear-end collision is higher; and the key pattern sequence for the occurrence of the accident is identified. Next, reinforcement learning is used to simulate traffic participants (cars, motorcycles, bicycles, pedestrians) exploring and learning in a road network model with various variable combinations (dropped cargo, construction cones, vehicle trajectory changes, pedestrian holiday paths, etc.). Then, reinforcement learning uses the prior knowledge provided by the probabilistic graphical model as a guide for its exploration and learning process, and generates risk dynamic scenarios through learning (traffic interactions during morning and evening rush hours, sudden pedestrian crossings in school areas). Finally, the generated risk dynamic scenarios are injected into the digital road network model to construct a dynamic participation layer, so that the subsequently generated road scenarios contain all interactively movable traffic participants and their interaction rules.

[0077] S123. Dynamically couple and adjust all environmental elements in the scene elements through a preset physical engine to generate environmental condition data, and inject the environmental condition data into the digital road network model to construct an environmental condition layer.

[0078] In this embodiment, environmental elements include lighting elements (strong sunlight at noon, low light at dusk), meteorological elements (heavy rain, heavy fog, snow accumulation), and time elements (day-night patterns, seasonal changes). The physics engine (CARLA / SUMO) uses the input meteorological elements, lighting elements, and time elements as initial conditions or driving variables, and simulates the dynamic effects of these environmental elements and their mutual influence in real time according to the laws of physics. Ultimately, it outputs environmental physical state data that changes with time or location, namely environmental condition data. Example scenarios include road surface reflection interference caused by heavy rain and reduced visibility in the winter evening. Finally, the environmental condition data is injected into the digital road network model to construct the environmental condition layer, which defines the physical environmental conditions of each point on the digital road network model at different times.

[0079] S124. Establish a digital twin of the roadside equipment in the test site, and interactively simulate the roadside equipment signals through the digital twin to construct a road facility layer.

[0080] In this embodiment, based on the information of roadside equipment deployed on the real site (such as communication protocol: DSRC / C-V2X, perception equipment: camera / radar, model range, signal light control logic, etc.), a digital twin of the equipment is established at the corresponding position of the virtual scene. The twin not only accurately replicates the geometric shape and installation posture of the equipment, but more importantly, it performs simulation modeling at the functional level: for example, implementing the V2X communication model, creating a virtual perception model to reproduce the control model of the traffic light, etc. These twins run interactively in the simulation: the perception model scans the virtual world to obtain target information, the control model executes the preset logic (such as the signal light changes), and the communication model generates and broadcasts simulated signal messages (such as the remaining green light time SPaT message). All twins together constitute the road facility layer

[0081] S130 , dynamically adjusting the initial multi-level road scene to the environmental changes by using the set environmental change timing parameters, to obtain a multi-level road scene that responds to the external environmental changes and test requirements in real time.

[0082] S200: Utilize dual conditional probability reasoning to determine the conditional dependencies between and within layers of a multi-layer road scene, and combine this with probability distribution sampling to generate an initial scene segment set representing the dynamic interactions of the multi-layer elements.

[0083] This embodiment utilizes dual conditional probabilistic reasoning technology (i.e., simultaneously considering the conditional dependencies between both inter-layer and intra-layer elements), enabling in-depth modeling of complex dynamic interactions in the real world. Combined with probability distribution sampling, it automatically generates a set of initial scene segments that highly characterize the dynamic interactions of multiple layers of elements, encompassing the behavioral decisions of traffic participants and responses to environmental changes. This significantly enhances the realism and complexity of the generated scenes.

[0084] In this embodiment, step S200 may further include the following sub-steps S210 to S230:

[0085] S210. Modeling the time-varying conditional dependency and probabilistic constraint chain of upper-layer elements on lower-layer elements in a multi-layer road scene through a dynamic Bayesian network to obtain a cross-layer conditional probability distribution table.

[0086] Furthermore, this embodiment uses a dynamic Bayesian network to model the time-varying conditional dependencies and probabilistic constraint chains of upper-level elements on lower-level elements in multi-level road scenes, where the state evolution of upper-level elements is regarded as the causal driving factor for the generation and evolution of lower-level elements, and an accurate cross-level conditional probability distribution table is obtained through Bayesian reasoning. In this embodiment, the dependency relationship of the state evolution of inter-layer elements over time is modeled, rather than just static dependency, so that the output results are closer to the essence of dynamic scene generation. The expression of the conditional dependency relationship between inter-layer elements is:

[0087] P(S,D,E,I)=P(S)·P(D|S)·P(E|S,D)·P(I|S,D,E) (1)

[0088] In formula (1), S represents the static environment layer elements, D represents the dynamic participation layer elements, E represents the environmental condition layer elements, and I represents the road facility layer elements.

[0089] S240. By using Markov random fields to model the high-dimensional spatiotemporal correlation and conditional dependence between elements at the same level in multi-level road scenes based on physical constraints, traffic rules and scene semantics, a joint probability energy function between elements is obtained that represents spatial proximity and temporal coherence.

[0090] Furthermore, this embodiment employs a learnable Markov random field to model the high-dimensional spatiotemporal correlations and conditional dependencies between elements within the same hierarchy, based on physical constraints, traffic rules, and scene semantics. The Markov random field automatically captures local dependency patterns between elements through graph structure learning and generates a joint probability energy function between elements that characterizes spatial proximity and temporal coherence. By adaptively capturing complex dependency patterns between elements at the same level, rather than relying on a fixed, pre-set structure, this embodiment significantly enhances the model's flexibility and real-world applicability.

[0091] S230 , performing probability distribution sampling on scene elements of the multi-level road scene based on the cross-level conditional probability distribution table and the joint probability energy function between elements, and generating an initial scene segment set of dynamic interaction of multi-level elements.

[0092] Furthermore, this embodiment employs a hierarchical sequential Gibbs sampling algorithm for probability distribution sampling. First, the upper-level element states are iteratively sampled in hierarchical topological order. Next, the conditional sampling space for the lower-level elements is dynamically constructed based on the upper-level sampling results and the cross-level conditional probability distribution table. Second, within the conditional sampling space, the Metropolis-Hastings criterion based on the joint probability energy function is applied to accept and reject element states, generating lower-level element states that meet spatiotemporal coupling constraints. Finally, the algorithm iterates until all levels of element states are sampled, ultimately generating an initial set of scene fragments with dynamic interaction between multiple levels of elements.

[0093] S300: Establish a multi-objective optimization function based on the set scene type distribution weights, element redundancy penalty factors, physical consistency constraints, and safety rules, screen and reorganize multiple initial scene fragment sets, and output the optimal scene set that covers the target distribution and meets all constraints.

[0094] This embodiment establishes a multi-objective optimization function that integrates scene type distribution weights, element redundancy penalty factors, physical consistency constraints, and safety rules. This allows for automated screening, reorganization, and output of an optimal scene set from a large initial set of scene fragments. This scene set maximizes coverage of the target distribution while strictly satisfying all pre-set constraints, ensuring high-quality, wide-ranging coverage, minimal redundancy, and realistic feasibility.

[0095] In this embodiment, step S300 may further include the following sub-steps S310 to S330:

[0096] S310: Integrate the set scene type distribution weights and element redundancy penalty factors to construct an initial multi-objective optimization function.

[0097] The expression of the initial multi-objective optimization function is:

[0098]

[0099] In formula (2), N represents the total number of key scenario types to be covered. If 10 types of edge cases are defined (such as sudden braking, lane change conflict, sensor failure, etc.), then N = 10; α k Represents the weight coefficient of the kth category scenario (reflecting the importance or risk level of the scenario, high-risk scenarios (such as children suddenly crossing) are given high weights (α k =0.9); Coverage(C k ) represents C k The coverage function of the class scenario is used to quantify the coverage ability of the generated scenario. The function output range is [0,1]. If the generated scenario library contains 80% of the heavy rain AEB test cases, then Coverage = 0.8; C k Represents the abstract category of the kth scenario, for example, C1 = "AEB triggering in heavy rain", C2 = "V2X signal loss and degradation"; β represents the redundancy penalty coefficient, β∈[0.1,0.5]; Redundancy represents the redundancy measure of the scenario library, Redundancy=1 / M∑i ≠ jsim(S i ,S j ), M is the total number of scenes in the scene library, S i is the feature vector of the i-th scene, S j is the feature vector of the jth scene, sim(S i ,S j ) is the scene similarity function, which is used to quantify the degree of repetition between two scenes, ∑i ≠ j is the sum of all different scene pairs.

[0100] S320. Inject physical consistency constraints and safety rules as mandatory constraints of the initial multi-objective optimization function to form a comprehensive multi-objective optimization function with multi-dimensional verification capabilities.

[0101] The expression of the comprehensive multi-objective optimization function is:

[0102]

[0103] In formula (3), PhysicalConsistency represents the physical consistency constraint, such as the vehicle cannot pass through the building, and SafetyViolation represents the safety rule constraint, such as the speeding risk threshold in rainy days.

[0104] S330, based on a comprehensive multi-objective optimization function, drives the collaborative screening and evolutionary reorganization of multiple initial scene fragment sets until a Pareto optimal scene set is dynamically generated and output that covers the predetermined type of distribution targets, meets the strong constraints of physical consistency and safety, and suppresses information redundancy.

[0105] S400. Extract the causal logical relationship of the optimal scenario set based on the directed acyclic graph, and topologically sort the causal logical relationship to generate a scenario evolution path that conforms to the causal temporal logic.

[0106] In this embodiment, the causal logical relationship between each event / state in the optimal scenario set is extracted through a directed acyclic graph (DAG), and topologically sorted, thereby generating a scenario evolution path that conforms to the causal temporal logic. This ensures that the occurrence of events in the test scenario has a reasonable logical relationship and sequence, greatly improving the fidelity and logical coherence of the test scenario, making the evaluation more scientific and effective. Among them, the expression of the scenario evolution path G is:

[0107]

[0108] In formula (4), G represents a directed acyclic graph; V represents a vertex set, i.e., all possible combinations of scene element states; F represents a directed edge set, i.e., connecting compatible vertices and representing legal transition paths between states); S i Represents the i-th instantiation state of the static environment layer element, such as S1: straight section of highway, S2: intersection with left-turn lane; D j Represents the jth instantiation state of the dynamic participation layer element, such as D1: a truck goes straight at 80 km / h, D2: a pedestrian crosses the road at 1.2 m / s; F k Represents the kth instantiation state of the environmental condition layer element, such as F1: sunny day with 100,000 lux, F2: heavy rain (visibility <50m); I lRepresents the first instantiation state of the road facility layer element, such as I1: traffic light is normal, I2: RSU (roadside unit) sends "accident ahead" message); p , v q Represents any vertex, that is, in S i ,D j ,F k ,I l A specific state in, such as v p =S1 (highway), v q =D1(truck goes straight); Compatible(v p , v q ) represents the vertex compatibility judgment function, verifying v p to v q Whether the transfer complies with physical / traffic regulations.

[0109] In this embodiment, step S400 may further include the following sub-steps S410 to S430:

[0110] S410. Model the causal logic relationship of the optimal scenario set obtained based on the directed acyclic graph analysis as a state transition constraint rule, and construct a temporal logic model with causal constraint capabilities based on the state transition constraint rule.

[0111] For example, the causal logic relationship extracted from a directed acyclic graph is: pedestrian intruding into the lane → vehicle identifying the target → triggering a braking decision → completing deceleration → rerouting. This causal logic relationship is converted into state transition constraints and a temporal logic model is constructed: Rule 1: If the system is in the normal driving state and a pedestrian intruding into the lane occurs, it must enter the emergency target recognition state; Rule 2: If the system is in the emergency target recognition state and a target recognition success event occurs, it must enter the braking decision activation state; Rule 3: If the system is in the braking decision activation state and the vehicle speed drops to a safety threshold, it is allowed to enter the rerouting state.

[0112] S420. Utilize the topological sorting results of the directed acyclic graph to generate a mandatory occurrence sequence of event elements in the optimal scenario set, and input the sequence as the core driving source into the temporal logic model to simulate the dynamic change process of the temporal logic model state in the temporal dimension.

[0113] For example, the mandatory sequence of event elements in the optimal scenario is: pedestrian entering the lane → vehicle identifying the target → triggering a braking decision → completing a deceleration → rerouting. This sequence drives the temporal logic model: first, the vehicle's normal driving speed is obtained; then, multiple events are injected in sequence: a pedestrian entering the lane, a successful target identification event, and a vehicle speed reduction to the minimum idle speed; finally, based on the model's activated states after each injection, a state transition chain is generated: normal driving → emergency target identification → activation of a braking decision → rerouting.

[0114] S430: Capture and analyze the sequence data of the temporal state transitions generated during the dynamic transition process in real time, and verify the sequence data based on the state transfer constraint rules and the preset temporal logic requirements. Based on the verification results, output the scenario evolution path that meets all causal logic constraints and temporal constraints.

[0115] For example, while generating the state transition chain, sequential state data is captured. The timing for normal driving is 0.0 seconds, the timing for emergency target recognition is 1.2 seconds, the timing for braking decision activation is 1.5 seconds, and the timing for path replanning is 3.0 seconds. The sequence data is then verified based on the state transition constraint rules and temporal logic requirements. For example, the system checks whether the braking decision activation strictly occurs after successful target recognition and verifies whether the time it takes for the vehicle speed to drop from normal speed to minimum idle speed meets the 2-second safety threshold. Based on the verification results, the final scenario evolution path is output: normal driving (0.0 seconds) → pedestrian entering the lane (1.2 seconds) → successful target recognition (1.5 seconds) → braking activation (1.5 seconds) → deceleration completed (3.0 seconds) → path replanning (3.2 seconds).

[0116] S500. Perform multiple safety verifications on the scenario evolution path by combining virtual data and real data, and determine the optimal scenario corresponding to the verified scenario evolution path as a valid autonomous driving test scenario.

[0117] In this embodiment, the generated scenario evolution path undergoes a multi-step safety verification mechanism that combines virtual data (simulation) with real data (such as road traffic data and accident data). This allows the scenario's validity and boundaries to be assessed from different perspectives (model accuracy, physical realism, and actual risk alignment). Only scenarios corresponding to evolutionary paths that pass all verification steps are ultimately determined to be valid test scenarios, significantly improving the safety and susceptibility of the test scenarios.

[0118] In this embodiment, step S500 may further include the following sub-steps S510 to S540:

[0119] S510. In the digital twin environment, the logical consistency and behavioral safety of the scenario evolution path are verified through the set timing logic, and the logical safety score of the scenario evolution path is calculated based on the verification results.

[0120] In this embodiment, virtual roads, sensor models (LiDAR / camera noise simulation), and other twins are constructed through real simulation platforms such as CARLA, Prescan, and VTD. In this way, the logical consistency and behavioral safety of the scene are verified through spatiotemporal logic in the digital twin environment to eliminate semantic errors that are not related to physics. The expression of spatiotemporal logic is:

[0121]

[0122] In formula (5), Φ represents the logic verification type; [t1,t2] Represents a global constraint operator, which always satisfies the sub-conditions in the time period [t1, t2], such as the "full speed limit" rule; [t3,t4] represents a finality operator, whose sub-conditions must be satisfied at least once within the time period [t3, t4], such as the requirement of "finally passing the green light." Speed(v) represents the speed function of vehicle v. f(Weather) represents an environment-dependent speed threshold function, with the speed limit calculated dynamically based on the weather. For example, on rainy days, f("rain") = 0.7 × the legal speed limit. t1 and t2 represent the start and end points of the time window for the section where the event occurred, i.e., absolute timestamps or relative event trigger times. For example, t1 represents the time when the vehicle enters the construction zone, and t2 represents the time when the vehicle leaves the construction zone. t3 and t4 represent the start and end points of the vehicle's time window at the intersection. For example, t3 represents the time when the vehicle is 50 meters from the intersection, and t4 represents the time when the vehicle reaches the stop line. TrafficLight represents the traffic light state variable, reflecting the current light color (Red / Yellow / Green). Green represents the green state of the light.

[0123] S520 . In a simulation environment separated from the physical execution environment, the rule responsiveness of the scenario evolution path is tested using set static traffic rules, and the rule compliance of the scenario evolution path is calculated based on the test results.

[0124] In this embodiment, in a simulation environment independent of the physical execution environment, the scenario generated by the virtual simulation is converted into a structured data stream (such as the OpenScenario format) and then injected into the perception / decision-making module to verify the scenario's static rule responsiveness and code robustness. For example, in verifying whether a scenario complies with traffic regulations based on the knowledge graph, the turn signal triggering logic in the lane change scenario is checked to ensure that it meets the "turn on 3 seconds in advance" rule.

[0125] S530 : Perform a multimodal comparison between the actual drive test data of the scenario evolution path and the preset standard value, and calculate the physical feasibility of the scenario evolution path based on the comparison result.

[0126] In this embodiment, the scenario's physical feasibility and closed-loop performance are verified in a real physical environment. Hardware-in-the-loop (HIL) testing involves connecting a real engine management system (ECU) to a simulator to test vehicle dynamics responses. Vehicle-in-the-loop (VIL) testing involves equipping a real vehicle with test equipment to replicate high-risk scenarios in a closed area. A mobile platform (such as a SoftCar) is used to simulate an unexpected vehicle, triggering real traffic signal control via V2X communication. Finally, uncovered boundary data is extracted based on real road test data, and the scenario library is incrementally updated.

[0127] S540: When the logical safety score, rule compliance, and physical feasibility of the scenario evolution path all meet the set thresholds, the optimal scenario corresponding to the scenario evolution path is determined as a valid autonomous driving test scenario.

[0128] In addition, this embodiment also proposes an automated derivation system for autonomous driving test scenarios, which includes:

[0129] The scenario requirement analysis and architecture mapping module is used to analyze the test requirements of the target autonomous driving system and map the analysis results to a preset multi-dimensional scenario element layered architecture to construct a multi-level road scenario that responds to external environment changes and test requirements.

[0130] The scene fragment generation module is used to use dual conditional probability reasoning to infer the conditional dependencies between inter-layer and intra-layer elements of multi-layer road scenes, and combine it with probability distribution sampling to generate an initial set of scene fragments that represent the dynamic interaction of multi-level elements.

[0131] The multi-objective scenario optimization module is used to establish a multi-objective optimization function based on the set scene type distribution weights, element redundancy penalty factors, physical consistency constraints and safety rules, screen and reorganize multiple initial scene fragment sets, and output the optimal scene set that covers the target distribution and meets all constraints.

[0132] The causal temporal path synthesis module is used to extract the causal logical relationship of the optimal scenario set based on the directed acyclic graph, and topologically sort the causal logical relationship to generate a scenario evolution path that conforms to the causal temporal logic.

[0133] The hybrid data security verification module is used to perform multiple security verifications on the scenario evolution path by combining virtual data and real data, and determine the optimal scenario corresponding to the verified scenario evolution path as a valid autonomous driving test scenario.

[0134] In addition, this embodiment also proposes an electronic device, including: a processor; and a memory storing automated derivative method steps for the processor to control the above-mentioned autonomous driving test scenario.

[0135] Finally, this embodiment also proposes a computer-readable medium on which computer-executable instructions are stored. When the executable instructions are executed by a processor, the automated derivation method steps of the autonomous driving test scenario described above are implemented.

[0136] In summary, the present embodiment proposes an automated derivation method, system, device, and medium for autonomous driving test scenarios. Through systematic hierarchical modeling, dynamic interaction of probabilistic reasoning modeling, multi-objective optimization to screen high-quality scenarios, construction of causal temporal logic paths, and implementation of multiple safety verifications, it ultimately achieves high-efficiency, automated, high-fidelity, high-coverage, and high-safety generation of autonomous driving test scenarios, significantly improving the depth, breadth, and reliability of autonomous driving system testing.

[0137] Since the systems / devices described in the above embodiments of the present invention are systems / devices used to implement the methods of the above embodiments of the present invention, those skilled in the art will be able to understand the specific structures and variations of these systems / devices based on the methods described in the above embodiments of the present invention, and thus will not be described in detail here. All systems / devices used in the methods of the above embodiments of the present invention are within the scope of protection of the present invention.

[0138] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions.

[0140] It should be noted that, in the description of the present invention, the word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The present invention can be implemented by means of hardware comprising several distinct components and by means of a suitably programmed computer. The use of the words first, second, third, etc., is merely for convenience and does not imply any order. These words should be understood as part of the component name.

[0141] In addition, it should be noted that, in the description of this specification, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0142] Although preferred embodiments of the present invention have been described, those skilled in the art will be able to make additional changes and modifications to these embodiments after obtaining the basic inventive concepts.

[0143] Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the invention.

Claims

1. An automated derivation method for autonomous driving test scenarios, characterized in that: include: Analyze the test requirements of the target autonomous driving system and map the analysis results to a preset multi-dimensional scenario element layered architecture to construct a multi-level road scenario that responds to external environment changes and test requirements; Dual conditional probability is used to reason about the conditional dependencies between and within layers of multi-layer road scenes, and combined with probability distribution sampling, an initial set of scene fragments is generated that represents the dynamic interactions of multi-layer elements. A multi-objective optimization function is established based on the set scene type distribution weights, element redundancy penalty factors, physical consistency constraints, and safety rules. Multiple initial scene fragment sets are screened and reorganized to output the optimal scene set that covers the target distribution and satisfies all constraints. Extract the causal logic relationship of the optimal scenario set based on the directed acyclic graph, and perform topological sorting on the causal logic relationship to generate a scenario evolution path that conforms to the causal temporal logic; Multiple safety verifications combining virtual data and real data are performed on the scenario evolution path, and the optimal scenario corresponding to the verified scenario evolution path is determined as a valid autonomous driving test scenario.

2. The method according to claim 1, wherein Analyze the test requirements of the target autonomous driving system and map the analysis results to a preset multi-dimensional scenario element layered architecture to construct a multi-level road scenario that responds to external environment changes and test requirements, including: Analyze the test requirements of the target autonomous driving system and retrieve all scenario elements that meet the test requirements from the preset database based on the analysis results; Mapping all scene elements to the corresponding hierarchical structure of the multi-dimensional scene element layered architecture to obtain an initial multi-level road scene; The initial multi-level road scene is dynamically adjusted to environmental changes through the set environmental change timing parameters to obtain a multi-level road scene that responds to external environmental changes and test requirements in real time; Among them, the multi-level road scene includes the static environment layer, dynamic participation layer, environmental condition layer and road facility layer.

3. The method according to claim 2, wherein Map all scene elements to the hierarchical structure of the corresponding multi-dimensional scene element layered architecture to obtain the initial multi-level road scene including: Road topology modeling is performed based on the test site map information in the scenario elements to generate a static environment layer, which is represented by an editable digital road network model. Using probabilistic graphical models and reinforcement learning algorithms, combined with historical accident data from the test site, we generate risk dynamic scenarios and inject them into the digital road network model to build a dynamic participation layer. Dynamically couple and adjust all environmental elements in the scene elements through the preset physical engine to generate environmental condition data, and then inject the environmental condition data into the digital road network model to construct the environmental condition layer; Establish a digital twin of the roadside equipment in the test site, and simulate the roadside equipment signals through digital twin interaction to build the road facility layer.

4. The method according to claim 1, wherein The conditional dependencies between and within layers of multi-layer road scenes are inferred using dual conditional probability, and combined with probability distribution sampling, an initial set of scene fragments representing the dynamic interactions of multi-layer elements is generated, including: The dynamic Bayesian network is used to model the time-varying conditional dependencies and probabilistic constraint chains of upper-level elements on lower-level elements in multi-level road scenes, and a cross-level conditional probability distribution table is obtained. By using Markov random fields to model the high-dimensional spatiotemporal correlation and conditional dependence between elements at the same level in multi-level road scenes based on physical constraints, traffic rules, and scene semantics, we obtain a joint probability energy function between elements that represents spatial proximity and temporal coherence. Based on the cross-level conditional probability distribution table and the joint probability energy function between elements, the scene elements of the multi-level road scene are sampled with probability distribution to generate an initial scene fragment set with dynamic interaction of multi-level elements.

5. The method according to claim 1, wherein A multi-objective optimization function is established based on the set scene type distribution weights, element redundancy penalty factors, physical consistency constraints, and safety rules. Multiple initial scene fragment sets are screened and reorganized to output the optimal scene set that covers the target distribution and satisfies all constraints. The following are included: The set scene type distribution weights and element redundancy penalty factors are integrated to construct the initial multi-objective optimization function; Inject physical consistency constraints and safety rules as mandatory constraints of the initial multi-objective optimization function to form a comprehensive multi-objective optimization function with multi-dimensional verification capabilities; Based on a comprehensive multi-objective optimization function, it drives the collaborative screening and evolutionary reorganization of multiple initial scene fragment sets until a Pareto optimal scene set is dynamically generated and output that covers the predetermined type of distribution targets, meets strong constraints of physical consistency and safety, and suppresses information redundancy.

6. The method according to claim 1, wherein The causal logic relationships of the optimal scenario set are extracted based on the directed acyclic graph, and the causal logic relationships are topologically sorted to generate scenario evolution paths that conform to the causal temporal logic, including: The causal logic relationship of the optimal scenario set obtained based on directed acyclic graph analysis is modeled as state transition constraint rules, and a temporal logic model with causal constraint capabilities is constructed based on the state transition constraint rules; Using the topological sorting results of the directed acyclic graph, a mandatory occurrence sequence of event elements in the optimal scenario set is generated. This sequence is then input into the temporal logic model as the core driving source to simulate the dynamic change process of the temporal logic model state in the temporal dimension. Capture and analyze the sequence data of temporal state transitions generated during dynamic changes in real time, verify the sequence data based on state transfer constraint rules and preset temporal logic requirements, and output the scenario evolution path that meets all causal logic constraints and temporal constraints based on the verification results.

7. The method according to claim 1, wherein Multiple safety verifications are performed on the scenario evolution path using both virtual and real data. The optimal scenario corresponding to the verified scenario evolution path is identified as a valid autonomous driving test scenario, including: In the digital twin environment, the logical consistency and behavioral safety of the scenario evolution path are verified through the set temporal logic, and the logical safety score of the scenario evolution path is calculated based on the verification results; In a simulation environment that is independent of the physical execution environment, the rule responsiveness of the scenario evolution path is tested using the set static traffic rules, and the rule compliance of the scenario evolution path is calculated based on the test results. Perform a multimodal comparison between the actual drive test data of the scenario evolution path and the preset standard values, and calculate the physical feasibility of the scenario evolution path based on the comparison results; When the logical safety score, rule compliance and physical feasibility of the scenario evolution path all meet the set thresholds, the optimal scenario corresponding to the scenario evolution path is determined as a valid autonomous driving test scenario.

8. An automated derivation system for autonomous driving test scenarios, characterized in that: include: The scenario requirement analysis and architecture mapping module is used to analyze the test requirements of the target autonomous driving system and map the analysis results to a preset multi-dimensional scenario element layered architecture to construct a multi-level road scenario that responds to external environment changes and test requirements; The scene fragment generation module is used to use dual conditional probability reasoning to determine the conditional dependencies between and within layers of multi-layer road scenes, and combined with probability distribution sampling to generate an initial set of scene fragments that represent the dynamic interactions of multi-layer elements; The multi-objective scenario optimization module is used to establish a multi-objective optimization function based on the set scenario type distribution weights, element redundancy penalty factors, physical consistency constraints, and safety rules. It screens and reorganizes multiple initial scenario fragment sets and outputs the optimal scenario set that covers the target distribution and satisfies all constraints. The causal temporal path synthesis module is used to extract the causal logical relationship of the optimal scenario set based on the directed acyclic graph, and topologically sort the causal logical relationship to generate a scenario evolution path that conforms to the causal temporal logic; The hybrid data security verification module is used to perform multiple security verifications on the scenario evolution path by combining virtual data and real data, and determine the optimal scenario corresponding to the verified scenario evolution path as a valid autonomous driving test scenario.

9. An electronic device, characterized in that: include: processor; A memory storing steps for a processor to control an automated derivation method of an autonomous driving test scenario according to any one of claims 1 to 7.

10. A computer-readable medium having computer-executable instructions stored thereon, characterized in that: When the executable instructions are executed by the processor, the steps of the automated derivation method of the autonomous driving test scenario as described in any one of claims 1 to 7 are implemented.

Images