Automatic driving scene generation method, system, equipment and medium

Abstract

The invention provides an automatic driving scene generation method, system and device and a medium, and relates to the field of virtual simulation scene generation. The method comprises the following steps: template extraction of semantic factors, and structured analysis of behavior, space and risk semantics from scene description; road partition map modeling is carried out, roads are divided into multiple types of areas such as main passing areas and intersection areas, and a topological connection map is constructed; coarse-grained dynamic trajectory generation: generating an initial feasible trajectory based on the simplified model; reconstructing a fine-grained trajectory, and performing multi-layer correction on the trajectory through road space constraint, dynamic continuity constraint and adaptive risk inference constraint; and multi-round iterative optimization is carried out, and a test scene which is physically executable, reasonable in space and adjustable in high risk is output through a generation-evaluation-reconstruction loop. According to the method, the problems of track physical distortion, insufficient road constraint expression and risk regulation deficiency in the prior art are effectively solved, and the authenticity and diversity of an automatic driving simulation test scene are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of virtual simulation scene generation technology, and in particular to a method, system, device and medium for generating autonomous driving scenes. Background Technology

[0002] As autonomous driving technology matures, its safety has become a core concern for the industry. The reliability of highly automated driving systems in complex road conditions, mixed traffic, and unexpected situations requires extensive testing through numerous high-risk, repeatable, and controllable simulation scenarios. However, existing scenario generation technologies remain limited in their capabilities, failing to meet the demands of automakers, research institutions, and industry regulators for "dangerous but reasonable" extreme scenarios.

[0003] Currently, there are three main technical approaches to simulation scene generation. The first approach is rule-based template-based scene construction, which typically uses predefined event templates and parameter combinations to generate scenes. However, this type of method relies too heavily on manual design and struggles to describe the complex, multi-factor, and highly interactive behaviors that occur in real traffic, resulting in a lack of diversity and innovation in the generated content.

[0004] The second category is adversarial scene generation methods based on reinforcement learning or optimization algorithms. These methods train adversarial agents to find trajectories or operational strategies that can disrupt autonomous driving systems. However, in most cases, the actions generated by these algorithms are prone to distortion, such as unreasonable extreme accelerations, frequent oscillating steering behaviors, or trajectories that directly cross impassable areas. This makes the generated results unexecutable in simulation systems and unusable as valid verification criteria.

[0005] The third type of method combines language models or generative models to drive scene generation through natural language descriptions. Although this type of method can improve semantic diversity, current generative models generally lack geometric and physical constraints, resulting in generated vehicle trajectories that fail to meet dynamic constraints, leading to the problem of "seemingly reasonable but difficult to execute." For example, vehicles may appear in impassable areas, trajectories may exhibit extreme curvature jumps, or the vehicle may not conform to road structures.

[0006] In summary, the existing technology has the following shortcomings: (1) the trajectory lacks physical consistency and does not conform to the laws of vehicle dynamics; (2) the road constraint expression ability is weak and it cannot accurately depict the passable spatial structure of high-precision maps; (3) it lacks an adaptive control mechanism for dangerous but reasonable scenarios, making it difficult for the scenario to approach the safety boundary. Summary of the Invention

[0007] This application provides a method, system, device, and medium for generating autonomous driving scenarios, aiming to propose a solution for generating complex autonomous driving scenarios that can simultaneously guarantee the physical executability of trajectories, the consistency of road structures, and the adjustability of risks. This application achieves the construction of high-value test scenarios for autonomous driving systems through a multi-layered feasibility constraint mechanism, structured road zoning map modeling, a two-stage trajectory generation strategy, and an adaptive risk inference model, overcoming the limitations of existing technologies in areas such as trajectory distortion, insufficient map representation, and lack of risk control.

[0008] This application is designed around the following core technical issues: First, existing traffic scene generation methods generally lack physical consistency, and vehicle trajectories cannot strictly follow the laws of dynamics. This application aims to address how to effectively constrain the acceleration, curvature, and response time of the trajectory to ensure that the generated trajectory is executable within the simulation system and the actual vehicle dynamics.

[0009] Secondly, traditional map constraint methods typically use centerlines or binary passable areas, resulting in weak structure and difficulty in accurately representing complex road topologies such as ramps, diversion zones, and complex intersections. This application aims to address how to express road traffic logic through highly structured area maps and use them for trajectory feasibility checks and guidance.

[0010] Furthermore, existing technologies lack adaptive risk control capabilities and cannot generate high-risk scenarios that approach the accident boundary but remain reasonable and legal. This application requires assessing the risk level during trajectory generation and dynamically adjusting the trajectory to optimize it towards higher risk.

[0011] Finally, the existing generation methods have weak iterative mechanisms, making it difficult to simultaneously meet the dual requirements of feasibility and high risk in a single generation step. This application proposes to construct a multi-round generation-evaluation-reconstruction loop process to iteratively approximate the optimal hazardous scenario.

[0012] Firstly, this application provides a method for generating autonomous driving scenarios, the method comprising: From the input scene description information, structured semantic factors are extracted based on predefined template and rule matching methods. The semantic factors include driving behavior factors, spatial structure factors, and risk triggering factors. Based on high-precision map information, the road environment is divided into multiple regional units with different traffic attributes, and the traffic connection relationship between regional units is defined through a graph structure to construct a road zoning map; wherein, the regional unit includes at least a main traffic area, an edge area, an intersection area, and a merging area; The initial trajectory value and behavioral trend are determined based on the extracted semantic factors, and a coarse-grained trajectory that satisfies the basic dynamic constraints is generated based on the simplified vehicle dynamics model. The coarse-grained trajectory is sequentially subjected to road zoning map spatial constraints, dynamic continuity constraints, and adaptive risk inference constraints, and fine-grained adjustments are made to obtain an intermediate trajectory. The adaptive risk inference constraints evaluate the risk level of the trajectory based on a multi-dimensional risk inference model, which integrates at least time collision, speed difference, occlusion ratio, and road pressure index. Based on the intermediate trajectory, the steps of trajectory generation, feasibility and risk assessment, and trajectory reconstruction based on assessment deviation are repeatedly executed. After multiple iterations, the target trajectory that meets the requirements of physical executability, spatial legality and preset risk is output.

[0013] In one possible design, the driving behavior factors include one or more of the following: going straight, changing lanes, overtaking, decelerating, emergency braking, merging lanes, and yielding; the spatial structure factor is used to characterize the road topology environment type; and the risk triggering factor is used to characterize the risk trend.

[0014] In one possible design, the road zoning modeling specifically includes: constructing a grid with a preset spatial resolution based on lane lines, curbs, and obstacle boundary information to divide the main traffic area, edge area, intersection area, and merging area; and establishing traversable edges between permitted traffic areas and prohibited traffic areas based on the road topology.

[0015] In one possible design, the simplified vehicle dynamics model includes longitudinal and lateral kinematic parameters; the basic dynamic constraints include threshold limits on acceleration, deceleration, angular velocity, and jerk.

[0016] In one possible design, the road zone map spatial constraint is used to ensure that each sampling point of the trajectory is within the passable area defined by the road zone map. If a trajectory point falls into a prohibited area, local adjustments are made through lateral offset correction, local interpolation, or resampling. The dynamic continuity constraint is used to ensure that the changes in the trajectory's velocity, acceleration, jerk, and angular velocity are continuous and within the vehicle's actuatory range. When the jerk exceeds a threshold or the angular velocity change is too large, spline interpolation or adjustment of the time step is used for smooth reconstruction.

[0017] In one possible design, the risk function of the adaptive risk inference constraint is expressed as: In the formula, For the comprehensive risk value, , , and These are the weighting coefficients. TTC For potential collision time, Due to speed difference, For the occlusion ratio, Road pressure; When the overall risk value is lower than the set threshold, the trajectory risk level is improved by fine-tuning the vehicle's longitudinal speed, lateral position, or the time it takes to enter the critical area.

[0018] In one possible design, the convergence condition generated by the multi-round iterative optimization includes at least one of the following: the number of iteration rounds reaches a preset upper limit, the trajectory satisfies all feasibility constraints, and the risk level of the trajectory reaches a preset target range.

[0019] Secondly, this application provides an autonomous driving scene generation system, the system comprising: The semantic factor extraction unit is configured to extract structured semantic factors from the input scene description information based on a predefined template and rule matching method. The semantic factors include driving behavior factors, spatial structure factors, and risk triggering factors. The road zoning map construction unit is configured to divide the road environment into multiple regional units with different traffic attributes based on high-precision map information, and to construct a road zoning map by defining the traffic connection relationship between regional units through a graph structure; wherein, the regional unit includes at least a main traffic area, an edge area, an intersection area, and a merging area; The coarse-grained trajectory generation unit is configured to determine the initial trajectory value and behavioral trend based on the extracted semantic factors, and generate a coarse-grained trajectory that satisfies the basic dynamic constraints according to the simplified vehicle dynamics model. The fine-grained trajectory reconstruction unit is configured to sequentially apply road zoning map spatial constraints, dynamic continuity constraints, and adaptive risk inference constraints to the coarse-grained trajectory to perform fine-grained adjustments and obtain an intermediate trajectory. The adaptive risk inference constraint evaluates the risk level of the trajectory based on a multi-dimensional risk inference model, which at least integrates time collision, speed difference, occlusion ratio, and road pressure index. The multi-round iterative optimization unit is configured to repeatedly execute trajectory generation, feasibility and risk assessment, and trajectory reconstruction based on assessment deviation based on the intermediate trajectory. After multiple rounds of iteration, it outputs a target trajectory that meets the requirements of physical executability, spatial legality, and preset risk.

[0020] Thirdly, embodiments of this application provide an electronic device, including: at least one processor and a memory; the memory stores computer execution instructions; the at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to perform the autonomous driving scene generation method as described in the first aspect and various possible designs of the first aspect.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the autonomous driving scenario generation method described in the first aspect and various possible designs of the first aspect.

[0022] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the autonomous driving scenario generation method described in the first aspect and various possible designs of the first aspect.

[0023] The autonomous driving scene generation method, system, device, and medium provided in this application have at least the following beneficial effects: First, by using a multi-layered feasibility constraint system, this application can strictly ensure that the generated trajectory conforms to vehicle dynamics constraints, and ensure that the trajectory can be executed on simulation platforms such as CARLA and LGSVL or on real vehicles, thus overcoming the common physical distortion problem of existing model-generated trajectories.

[0024] Secondly, this application uses road zoning map structure to express high-precision maps, which can fully express complex topological relationships and achieve map constraint capabilities with higher dimensions and more logical structure than traditional centerline representation, providing safe and complete environmental boundary information for the generation algorithm.

[0025] Furthermore, the adaptive risk inference mechanism of this application can assess the risk level in real time based on the interaction status between vehicles, and guide the scenario to approach the accident threshold by adjusting trajectory parameters, providing more valuable and more realistic extreme scenarios for the robustness testing of autonomous driving systems.

[0026] In addition, this application adopts an iterative reconstruction strategy, which refines the results of each generation based on the previous one, ensuring both the rationality of the trajectory and the continuous improvement of scenario risks, thus achieving a multi-objective generation effect that balances feasibility and risk.

[0027] In summary, this application can significantly improve the quality and importance of autonomous driving scenario generation, provide the industry with a new technological path for building virtual test scenarios, and is of great significance for improving the safety and reliability of autonomous driving systems. Attached Figure Description

[0028] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0029] Figure 1This is a schematic diagram of the overall process of an autonomous driving scene generation method provided in an embodiment of this application, showing the overall framework structure from semantic factor extraction, partition map modeling, trajectory generation to iterative optimization; Figure 2 A flowchart illustrating the implementation of an autonomous driving scene generation method provided in this application embodiment; Figure 3 This is a schematic diagram of the road zoning map structure provided in an embodiment of this application; Figure 4 This is a schematic diagram of coarse-grained dynamic trajectory generation provided in an embodiment of this application; Figure 5 This is a schematic diagram of the risk inference model structure provided in an embodiment of this application; Figure 6 This is a structural diagram of the autonomous driving scene generation system provided in an embodiment of this application.

[0030] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0031] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of systems and methods consistent with some aspects of this application as detailed in the appended claims.

[0032] The collection, storage, use, processing, transmission, provision, and disclosure of user data and other information involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0033] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0034] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0035] This application provides a method for generating autonomous driving scenarios. This method achieves multi-dimensional safety verification of autonomous driving systems through a multi-layered feasibility constraint model, structured road environment representation, vehicle dynamics consistency verification, and a trajectory reconstruction mechanism based on risk inference. This application can be used in virtual simulation platforms, autonomous driving function development processes, intelligent transportation system safety evaluation, and vehicle-road cooperative system testing, among other application scenarios.

[0036] like Figure 1 The diagram illustrates the overall process of an autonomous driving scene generation method provided in this application, showcasing the overall framework from semantic factor extraction, road zoning modeling, trajectory generation to iterative optimization. The overall framework for implementing this method consists of five parts: semantic factor template extraction, road zoning modeling, coarse-grained trajectory generation, fine-grained trajectory reconstruction with multi-layered feasibility constraints, and multi-round iterative generation optimization. First, based on predefined semantic factor templates, this application extracts behavioral semantics, spatial semantics, and risk-triggered semantics from the input scene description, providing structured input for subsequent scene prototype construction. Second, this application constructs a road zoning map with main regions, edge regions, intersection regions, and merging regions through region division and topological relationship modeling, used for spatial constraints and legality judgment of the trajectory. Subsequently, this application adopts a two-stage trajectory generation strategy: the first stage uses a simplified dynamics model to generate a basic executable trajectory; the second stage combines road feasibility, dynamic continuity, and risk inference results to fine-grainedly adjust the trajectory, gradually approximating a high-value test scenario. Furthermore, this application constructs a multi-dimensional risk inference model, dynamically assessing the risk level of the trajectory through indicators such as TTC, speed difference, occlusion ratio, and road congestion, and making guided corrections based on insufficient risk or deviation from the target direction. Finally, this application employs a multi-round generation-evaluation-reconstruction iterative mechanism, enabling the trajectory to be locally optimized based on deviations in each iteration, ultimately generating a target trajectory that simultaneously satisfies physical consistency, road legality, and high risk.

[0037] Specifically, the autonomous driving scenario generation method can be implemented through the following steps S10-S50.

[0038] S10: Extract structured semantic factors from the input scene description information based on predefined template and rule matching methods. Semantic factors include driving behavior factors, spatial structure factors, and risk triggering factors.

[0039] Step S10 is used to perform semantic factor template extraction. This step aims to extract structured semantic factors from the input scene description information, including driving behavior factors, spatial structure factors, and risk triggering factors. This embodiment uses a predefined template and rule matching method for semantic parsing, without relying on a deep language model, to ensure the method's controllability and interpretability.

[0040] First, based on traffic regulations and common driving behavior patterns, this embodiment sets a set of behavioral factors including going straight, changing lanes, overtaking, decelerating, emergency braking, merging, and yielding. Behavioral elements in the input description are identified through keyword matching and sentence structure analysis.

[0041] Secondly, spatial factors are used to characterize the topological environment of the road where the scene is located, including types such as "intersection", "ramp entrance", "roundabout", and "two-lane straight section". The system identifies region types through a pre-set dictionary and syntactic analysis, providing semantic input for zoning map modeling.

[0042] Furthermore, this embodiment extracts words that reflect risk trends, such as "suddenly," "obstruction," "approaching," and "close-range entry," for use as logical input in the subsequent risk inference module.

[0043] Through the above processing, step S10 outputs a list of structured semantic factors, providing a semantic constraint framework for subsequent trajectory generation.

[0044] S20: Based on high-precision map information, the road environment is divided into multiple regional units with different traffic attributes, and the traffic connection relationship between regional units is defined through a graph structure to construct a road zoning map; wherein, the regional unit includes at least the main traffic area, edge area, intersection area and merging area.

[0045] Step S20 is used to implement road zoning map modeling. Road zoning map modeling is a key innovation of this application in spatial structure representation. This model divides the road environment into regional units with different traffic attributes and defines the traffic relationships between regions through a graph structure, providing precise geometric and topological basis for trajectory generation and legality constraints. The zoning map consists of a set of nodes V and a set of edges E, representing road regions and their traversable connections, respectively.

[0046] In this implementation, the road is divided into four basic regions: Main Region, Edge Region, Conflict Region, and Merge Region. Each region is defined by its shape, size, location, and permitted traffic flow direction. The region division can be based on information such as lane lines, curbs, and obstacle boundaries from a high-precision map, using a spatial resolution of 0.1–0.5 meters to construct a grid to ensure sufficient geometric accuracy.

[0047] For connections between different areas, this embodiment constructs traversable edges based on the road topology. For example, there are bidirectional or unidirectional traversable links between the main traffic area and the merging area; there are potentially conflicting links between the main area and the intersection area; and prohibited traversable edges are established between the edge area and non-road space to constrain the trajectory from crossing road boundaries. Through the above steps, a road zoning map G = (V, E) with strong expressive power can be obtained, which is used for trajectory feasibility verification and spatial constraints.

[0048] like Figure 3 The diagram shown is a schematic representation of a road zoning map provided in this application embodiment, including a main traffic area, edge areas, intersection areas, merging areas, and their topological connections. Based on an exemplary road environment, this diagram clearly presents the various road area types and their spatial topological relationships defined by the method of this application. The main road in the diagram is divided into a main traffic area with clearly defined traffic attributes, specifically a structured main lane containing multiple lanes. Simultaneously, merging areas are clearly identified, such as ramp areas connected to the main lanes, and intersection areas used to handle traffic flow convergence conflicts. The diagram also exemplarily marks traffic participants located in different areas; for example, vehicle A is traveling in the main lane, and vehicle C is entering from the merging area, intuitively demonstrating the positional relationship between vehicles and road zones. This schematic diagram as a whole reveals the connection methods between the main traffic area, merging areas, and intersection areas, providing an intuitive basis for understanding the spatial feasibility constraints of trajectories in the method of this application.

[0049] S30: Determine the initial trajectory value and behavioral trend based on the extracted semantic factors, and generate a coarse-grained trajectory that satisfies the basic dynamic constraints based on the simplified vehicle dynamics model.

[0050] Step S30 is used to generate a coarse-grained dynamic trajectory. After parsing the scene semantics and map structure, this embodiment first generates a coarse-grained trajectory T0. This trajectory is determined by the vehicle dynamics model, ensuring that it meets basic executability requirements.

[0051] Coarse-grained trajectory generation is based on a simplified vehicle dynamics model, including longitudinal kinematics (velocity v, acceleration a) and lateral kinematics (heading angle θ, heading angular velocity ω). The generation algorithm simulates the vehicle's travel path over a given time range using a sampling frequency of 10–20 Hz. The dynamic parameters can be set within the following ranges: Maximum acceleration: 2–5 m / s² 2 ; Minimum deceleration: -4 to -8 m / s 2 ; Maximum angular velocity: 0.1–0.4 rad / s; Maximum speed: 2–6 m / s 3 ; The initial trajectory value is determined by semantic factors, such as the vehicle's initial position, speed, and direction in the scene. Then, the trajectory trend (such as going straight, changing lanes, cutting in, etc.) is determined by behavioral templates. The coarse trajectory only needs to meet basic dynamic constraints and does not need to meet complex road and risk conditions.

[0052] Figure 4 This schematic diagram illustrates the coarse-grained dynamic trajectory generation provided in this application, visually demonstrating the spatial morphology of the trajectory and the relationship between key dynamic parameters and time. The left side of the diagram shows the spatial trajectory of a vehicle in a road environment. This trajectory starts from a clearly marked starting point, crosses different lanes, and smoothly evolves towards the endpoint, reflecting the basic direction and continuity of the trajectory within the road structure. The right side of the diagram displays three types of dynamic parameter curves corresponding to this trajectory: the curve of velocity changing with time, the curve of acceleration changing with time, and the curve of jerk changing with time. These curves collectively reveal that in the coarse-grained trajectory generation stage, the vehicle's velocity, acceleration, and rate of change are all constrained within a reasonable range, and the changes are continuous and smooth, thus verifying that the trajectory meets the basic requirements for vehicle dynamics feasibility. Overall, this schematic diagram illustrates that the method of this application considers both the rationality of the spatial path and the feasibility of the dynamic behavior in the initial trajectory generation stage.

[0053] S40: Apply road zoning map spatial constraints, dynamic continuity constraints, and adaptive risk inference constraints sequentially to the coarse-grained trajectory, and make fine-grained adjustments to obtain the intermediate trajectory; the adaptive risk inference constraint evaluates the risk level of the trajectory based on a multi-dimensional risk inference model, which integrates at least time collision, speed difference, occlusion ratio, and road pressure index.

[0054] Step S40 is used to achieve fine-grained trajectory reconstruction under multi-layered feasibility constraints. After the coarse trajectory is generated, this embodiment performs fine-grained adjustments to the trajectory based on three types of constraints, so that it gradually approaches the final trajectory that is physically reasonable, spatially legal, and has appropriate risks: road zoning map spatial constraints, dynamic continuity constraints, and adaptive risk inference constraints.

[0055] Spatial feasibility constraints based on the partition map are used to ensure that each sampling point of the trajectory is within a passable area. If a trajectory point falls into a prohibited area or an inaccessible location between areas, the system locally adjusts the trajectory through lateral offset correction, local interpolation, or resampling.

[0056] For abrupt changes in heading that occur in the intersection area, this application uses local curvature minimization or cubic spline smoothing to adjust the heading angle to ensure the continuity of the trajectory space.

[0057] Dynamic continuity constraints are used to ensure that trajectory parameters such as velocity, acceleration, jerk, and angular velocity meet the vehicle's operational limits. Typical detection methods include: local curvature reconstruction when jerk exceeds a threshold; reallocation of the time step when angular velocity changes too much; and using smooth interpolation to control the magnitude of acceleration jumps.

[0058] For example, when the jerkiness of a certain segment of the trajectory exceeds a preset threshold (e.g., 5 m / s), 3 When the angular velocity changes beyond the upper limit, this application uses cubic spline interpolation to reconstruct the trajectory segment to make its curvature continuous; when the angular velocity changes beyond the upper limit, it is automatically smoothed by adjusting the time step or reducing the steering amplitude.

[0059] Through such continuity constraints, this method ensures that the trajectory will not exhibit unnatural phenomena such as physical inconsistencies or extreme changes in motion.

[0060] The adaptive risk inference constraint is used to generate scenarios that are more meaningful for safety verification. To achieve this constraint, this embodiment constructs a multi-dimensional risk inference model to dynamically evaluate and adjust the trajectory risk level.

[0061] Figure 5 A schematic diagram of the risk inference model provided in this application embodiment is shown. The model takes generated trajectory and environmental information as input, including the vehicle's position, speed, surrounding obstacle status, and road structure. This information is processed in parallel by multiple dedicated analysis modules: a TTC calculation module assesses the potential collision time; a speed difference analysis module calculates the relative speed between the vehicle and surrounding road users; an occlusion assessment module analyzes the degree of visual restriction; and a road pressure module quantifies the sense of urgency posed by the surrounding environment to the vehicle. The analysis results from each module are then aggregated into a risk fusion and assessment unit, which comprehensively evaluates multiple indicators and ultimately outputs a risk guidance signal to guide trajectory adjustments.

[0062] The risk inference model involves the following four categories of risk indicators: TTC (Time-to-Collision): Reflects the potential collision time.

[0063] Speed ​​difference : Reflects the trend of interactive conflict.

[0064] Occlusion ratio This reflects the degree of limitation in one's field of vision.

[0065] Road pressure : Reflects the spatial compression state of the surrounding vehicles or road structure.

[0066] The risk function can be expressed as: In the formula, For the comprehensive risk value, , , and These are the weighting coefficients.

[0067] When the risk value is lower than the set threshold, the system fine-tunes the vehicle's longitudinal speed, lateral position, or the time of entry into the intersection zone to make the trajectory gradually approach the safe boundary conditions while still maintaining executability.

[0068] S50: Based on the intermediate trajectory, repeatedly execute the steps of trajectory generation, feasibility and risk assessment, and trajectory reconstruction based on assessment deviation. After multiple iterations, output the target trajectory that meets the requirements of physical executability, spatial legality and preset risk.

[0069] Step S50 is used to implement multi-round generation-evaluation-reconstruction iterative optimization. To achieve a balance between executability, legality, and risk in the trajectory, this embodiment adopts an iterative optimization strategy. The following steps are performed in each iteration: S501: Generate the initial trajectory or the trajectory adjusted in the previous round; S502: Perform spatial feasibility checks and dynamic continuity checks on the partition map; S503: Perform risk inference and assess the risk level of the trajectory; S504: If any constraint is not met, local reconstruction is performed according to the deviation direction, and the next iteration is initiated.

[0070] The iterative process terminates after the convergence condition is met, typically in no more than 10 iterations. This ultimately yields a high-quality target trajectory that simultaneously satisfies multiple constraints, which can be directly used in simulation systems such as CARLA, LGSVL, and Apollo.

[0071] The methodology of this application is based on mature theories such as vehicle dynamics, road topology modeling, and traffic behavior risk inference. Each module has its own independent scientific basis, and they are coupled together to form an overall scene generation capability.

[0072] First, regarding vehicle trajectory generation, the simplified dynamics model used in this application is based on common passenger vehicle dynamics characteristics, including physical parameters such as acceleration, deceleration, and jerk. These constraints are derived from the engineering specifications of real vehicles. Through a coarse trajectory generation stage and a fine-grained reconstruction stage, this method ensures that the trajectory conforms to the laws of vehicle dynamics, for example, avoiding unrealistic instantaneous speed changes or extreme acceleration jumps, thus making it physically feasible.

[0073] Secondly, the road zoning map modeling method introduced in this application has a strong logical structure. Compared with traditional rasterized mask or centerline models, zoning maps can express higher-dimensional road attributes, including regional properties, traffic priorities, conflict relationships, and merging boundaries. This type of graph structure can be combined with reachability analysis and path planning theory provided by mathematical graph theory to make feasibility verification more accurate. Simultaneously, the connectivity between nodes can be used to determine the legality of a trajectory, such as quickly detecting whether a trajectory crosses impassable edges or whether an infeasible turning maneuver was performed in an intersection area.

[0074] Furthermore, the risk inference mechanism is based on the theoretical foundations of traffic behavior and collision warning. TTC is one of the most commonly used risk indicators in ADAS and autonomous driving systems, with a larger reciprocal indicating greater danger. Occlusion factor can be obtained through line-of-sight analysis, reflecting the visibility of traffic participants limited by the environment. Speed ​​gradient and road pressure reflect the conflict trend of interactive behavior. By constructing multi-dimensional risk indicators, this application can continuously adjust the trajectory risk level, enabling the scenario to approach a dangerous but reasonable boundary area, thereby maximizing the testing value.

[0075] Ultimately, the multi-round iterative generation-evaluation-reconstruction mechanism is essentially a convergence process. Each iteration advances the trajectory along the direction of constraint satisfaction, gradually reaching a balance between road feasibility, dynamic feasibility, and risk. This multi-objective convergence mechanism ensures the stability, continuity, and high risk of the generated trajectory, and its process is similar to the Pareto optimization process in multi-objective optimization, making scene generation more scientific and reliable.

[0076] In summary, this application is not only theoretically feasible, but also has significant technical advantages because it can effectively improve the quality, risk and execution capabilities of autonomous driving test scenarios through multi-module collaboration.

[0077] To verify the comprehensive performance of the proposed method in terms of trajectory physical executability, spatial legality, risk control capabilities, and scenario diversity, this section constructs multiple typical traffic scenarios based on a unified simulation platform and systematically compares them with two representative methods: traditional rule-based template methods and adversarial reinforcement learning methods. All experiments are conducted under the same map, simulation configuration, and consistent initial conditions to ensure the objectivity and reliability of the comparison results. The experimental data below are only used to demonstrate the technical effectiveness of the proposed method and do not constitute a limitation on the scope of protection.

[0078] I. Experimental Scenario Setup.

[0079] The experiments covered typical traffic environments, including urban roads, highways, and mixed-use structures, encompassing multiple representative traffic scenarios. These included: urban intersections with multi-vehicle interaction; high-speed lane-changing scenarios where ramps merge into main roads; low-speed following and lateral cutting scenarios on straight urban roads; and intersection scenarios with limited visibility due to building obstructions. Different initial speeds, surrounding traffic densities, and road topology changes were set for each scenario to test the adaptability of the proposed method under varying levels of complexity and risk.

[0080] II. Evaluation Index System.

[0081] To comprehensively evaluate the quality of scene generation, this embodiment employs a four-dimensional evaluation index system: physical executability, spatial legitimacy, risk controllability, and scene diversity. Regarding physical executability, the focus is on monitoring the dynamic consistency of the trajectory, including whether acceleration, curvature, and jerkiness are within the executable range. For spatial legitimacy, attention is paid to whether the trajectory enters impassable areas and crosses prohibited topological connecting edges. In terms of risk indicators, the level of danger of the scene is measured by the average TTC value and the proportion of TTC falling into the high-risk zone. Regarding diversity, it is measured by the number of scene variants that can be generated under the same semantic input and the complexity of interactive behaviors. These indicators allow for a comprehensive evaluation of the performance differences between different methods from multiple perspectives.

[0082] III. Comparison of experimental results.

[0083] Regarding physical executability, the method in this application relies on dynamic continuity constraints to ensure that the generated trajectory remains highly executable even in complex interactive scenarios. In experiments with a crossroads interference scenario, adversarial reinforcement learning methods, due to the lack of continuity in action generation, resulted in approximately 28% of trajectory segments becoming unexecutable, such as excessive acceleration or drastic curvature changes; while the method in this application, through multi-layered dynamic constraints, significantly reduced such unreasonable segments to below 2%. Rule-based template methods, while possessing high executability due to behavioral constraints, lack complexity in actions, making it difficult to reveal potential risks in the system.

[0084] Regarding spatial legitimacy, experimental results show that the proposed method, relying on structured modeling of "road zoning maps," effectively prevents trajectories from falling outside the road or crossing impassable connecting edges. In tests on complex road environments such as roundabouts and lane merges, adversarial reinforcement learning methods generated approximately 12% of out-of-bounds trajectories, while the proposed method maintained trajectories entirely within the passable area in all experiments. In contrast, while the rule-based template method has a low out-of-bounds rate, its performance is limited in scenarios with weak road structure representation capabilities (such as multi-layered lane merges or complex intersection areas).

[0085] In terms of risk control, the method presented in this application demonstrates stable risk adjustability. Trajectories generated by rule-based template methods are generally safer, with average TTC values ​​concentrated between 4 and 5 seconds, making it difficult to create high-risk scenarios approaching accident boundaries. Adversarial reinforcement learning methods often exhibit extreme behavior with TTC values ​​below 1 second, rendering the trajectory unexecutable. However, the method presented in this application, through risk inference and trajectory fine-tuning mechanisms, ensures that the TTC of the generated scenarios stably falls within the 1.5–2.0 second range, more closely resembling the pre-accident state of real traffic accidents, achieving both high risk and maintainability.

[0086] Regarding scene diversity, the method in this application, under the same semantic input conditions, can generate more than 10 variants with different risk patterns and spatial interaction structures, including different conflict point entry timings, different speed difference configurations, and different spatial pressure modes. While adversarial reinforcement learning methods can generate a large variety of behaviors, they lack stability and controllability; rule template methods have a limited number of variants, making it difficult to meet the needs of systematic testing. This application achieves a good balance between diversity and controllability, making scene library construction more practically valuable.

[0087] IV. Comparative Analysis and Summary.

[0088] In summary, the method presented in this application outperforms the comparative methods in several key performance indicators. Regarding physical executability, the method effectively constrains trajectory dynamics, making it robust and executable. In terms of spatial legality, it accurately represents complex road topologies through road zoning maps. Regarding risk control, the method can generate high-risk scenarios that approach safety boundaries while maintaining executability. Regarding diversity, the method possesses the ability to generate multiple interaction structures and risk patterns, significantly improving the richness of simulation test data. Therefore, the method presented in this application can serve as an important approach for constructing autonomous driving test scenarios, providing reliable support for system performance verification and robustness evaluation.

[0089] The following examples are intended to further illustrate the specific application of the method of this application, and to demonstrate the generation effect of the method under different road environments and traffic interaction conditions. Those skilled in the art can make appropriate adjustments to the parameters and forms based on the following embodiments, and all such adjustments should be considered to fall within the protection scope of this application.

[0090] Example 1: Generation of a traffic congestion scenario when merging lanes on a highway.

[0091] This example is used to verify the risk control capability and trajectory executability of the proposed method in a highway merging scenario. The scenario consists of the lead vehicle (Ego), following vehicles, and vehicles merging from the right. Based on the semantic description, the proposed method first identifies semantic factors such as "lead vehicle traveling straight at high speed," "following vehicle rapidly approaching," and "vehicle on the right attempting to merge," and then generates a road zoning map containing the main traffic area and merging area based on the road structure.

[0092] In the coarse-grained trajectory generation stage, the system generates an initial trajectory that conforms to high-speed operating conditions based on the vehicle dynamics model. Because the initial trajectory has a low risk level (TTC value of approximately 4.2 seconds), the risk inference module identifies insufficient risk. In the fine-grained reconstruction stage, the system adjusts the longitudinal speed of the lead vehicle and appropriately delays its entry into the lane-merging area, reducing the TTC between the lead vehicle and the merging vehicle to approximately 1.8 seconds, while avoiding unenforceable extreme maneuvers. The final trajectory is dynamically continuous and spatially legal, effectively constructing a high-speed lane-merging scenario that approaches the safety boundary.

[0093] Example 2: Scenarios where the view is obstructed at urban intersections.

[0094] This example verifies the risk mitigation capability of the proposed method under conditions of limited visibility and potential conflict. The scenario is a typical urban intersection where some buildings at the corners cause obstruction, easily creating the risk of "invisible oncoming vehicles." The proposed method first extracts semantic factors such as "obstruction," "intersection," and "left-turn / straight-through interaction" based on semantics, and then marks the intersection and obstruction areas as key risk areas in the partition map.

[0095] The coarse trajectory is generated along a straight path, while the risk inference module detects a high occlusion ratio (O>0.6) and a potential collision point with oncoming vehicles. In the fine-grained trajectory reconstruction stage, the method in this application adjusts the time when the main vehicle enters the intersection zone, reducing the time difference between it and the oncoming vehicle at the collision point to within the range of 0.3–0.6 seconds. This adjustment maintains trajectory feasibility while bringing the scene closer to the potential collision boundary, thereby generating urban occlusion hazard scenarios with high validation value.

[0096] Example 3: A scenario where a car suddenly cuts into a low-speed following vehicle.

[0097] This example verifies the ability of the proposed method to construct entry risk scenarios under low-speed urban conditions. The scenario consists of the lead vehicle, following vehicles, and vehicles attempting to cut in from the left. The semantic factor extraction stage identifies semantic information such as "low-speed following," "short distance," and "sudden entry," and constructs a partition map structure containing the main traffic area and the lateral entry area based on this information.

[0098] In the coarse trajectory stage, the lead vehicle maintains a normal following relationship with the vehicle in front, and the risk inference module detects an average TTC value of approximately 4 seconds, indicating insufficient risk. In the fine-grained trajectory reconstruction stage, the method in this application coordinates the lateral trajectory of the cutting vehicle and the longitudinal speed of the lead vehicle, causing the cutting vehicle to enter the position in front of the lead vehicle with a time difference of approximately 1.5 seconds, thus reconstructing a realistic and executable "sudden cut-in" scenario. This scenario is widely used to test the responsiveness and decision-making robustness of autonomous driving systems.

[0099] This application also provides an autonomous driving scene generation system, such as... Figure 6 As shown, the autonomous driving scenario generation system includes: The semantic factor extraction unit 601 is configured to extract structured semantic factors from the input scene description information based on a predefined template and rule matching method. The semantic factors include driving behavior factors, spatial structure factors, and risk triggering factors. The road zoning map construction unit 602 is configured to divide the road environment into multiple regional units with different traffic attributes based on high-precision map information, and to construct a road zoning map by defining the traffic connection relationship between regional units through a graph structure; wherein, the regional unit includes at least a main traffic area, an edge area, an intersection area, and a merging area; The coarse-grained trajectory generation unit 603 is configured to determine the initial trajectory value and behavioral trend based on the extracted semantic factors, and generate a coarse-grained trajectory that satisfies the basic dynamic constraints according to the simplified vehicle dynamics model. The fine-grained trajectory reconstruction unit 604 is configured to sequentially apply road zoning map spatial constraints, dynamic continuity constraints, and adaptive risk inference constraints to the coarse-grained trajectory to perform fine-grained adjustments and obtain an intermediate trajectory; the adaptive risk inference constraints evaluate the risk level of the trajectory based on a multi-dimensional risk inference model, and the risk inference model integrates at least time collision, speed difference, occlusion ratio, and road pressure index. The multi-round iterative optimization unit 605 is configured to repeatedly execute trajectory generation, feasibility and risk assessment, and trajectory reconstruction based on assessment deviation based on the intermediate trajectory. After multiple rounds of iteration, it outputs a target trajectory that meets the requirements of physical executability, spatial legality, and preset risk.

[0100] This application provides an electronic device. The electronic device may include a processor and a memory, wherein the processor and the memory can communicate; exemplarily, the processor and the memory communicate via a communication bus.

[0101] The processor executes computer execution instructions stored in memory, causing the processor to perform the scheme in the above embodiments. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0102] The communication bus can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. Transceivers are used to enable communication between the database access system and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include random access memory (RAM) and may also include non-volatile memory.

[0103] The electronic device provided in this application embodiment can be the terminal device described in the above embodiments.

[0104] This application also provides a computer-readable storage medium storing computer instructions. When the computer instructions are executed on a computer, the computer performs the technical solution of the autonomous driving scene generation method described above.

[0105] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the autonomous driving scene generation method in the above embodiments.

[0106] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or modules may be electrical, mechanical, or other forms.

[0107] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0108] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0109] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0110] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0111] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0112] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Architecture (EISA) buses, etc. Buses can be categorized into address buses, data buses, control buses, etc.

[0113] The aforementioned storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0114] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic control unit or main control device.

[0115] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for generating autonomous driving scenarios, characterized in that, The method includes: From the input scene description information, structured semantic factors are extracted based on predefined template and rule matching methods. The semantic factors include driving behavior factors, spatial structure factors, and risk triggering factors. Based on high-precision map information, the road environment is divided into multiple regional units with different traffic attributes, and the traffic connection relationship between regional units is defined through a graph structure to construct a road zoning map; wherein, the regional unit includes at least a main traffic area, an edge area, an intersection area, and a merging area; The initial trajectory value and behavioral trend are determined based on the extracted semantic factors, and a coarse-grained trajectory that satisfies the basic dynamic constraints is generated based on the simplified vehicle dynamics model. The coarse-grained trajectory is sequentially subjected to road zoning map spatial constraints, dynamic continuity constraints, and adaptive risk inference constraints, and fine-grained adjustments are made to obtain an intermediate trajectory. The adaptive risk inference constraints evaluate the risk level of the trajectory based on a multi-dimensional risk inference model, which integrates at least time collision, speed difference, occlusion ratio, and road pressure index. Based on the intermediate trajectory, the steps of trajectory generation, feasibility and risk assessment, and trajectory reconstruction based on assessment deviation are repeatedly executed. After multiple iterations, the target trajectory that meets the requirements of physical executability, spatial legality and preset risk is output.

2. The autonomous driving scene generation method according to claim 1, characterized in that, The driving behavior factors include one or more of the following: going straight, changing lanes, overtaking, decelerating, emergency braking, merging lanes, and yielding; the spatial structure factors are used to characterize the road topology environment type; and the risk triggering factors are used to characterize risk trends.

3. The autonomous driving scene generation method according to claim 1, characterized in that, The road zoning modeling specifically includes: constructing a grid with a preset spatial resolution based on lane lines, road edges, and obstacle boundary information to divide the main traffic area, edge area, intersection area, and merging area; and establishing passable edges between permitted traffic areas and prohibited traffic edges between prohibited traffic areas according to the road topology.

4. The autonomous driving scene generation method according to claim 1, characterized in that, The simplified vehicle dynamics model includes longitudinal and lateral kinematic parameters; the basic dynamic constraints include threshold limits on acceleration, deceleration, angular velocity, and jerk.

5. The autonomous driving scene generation method according to claim 1, characterized in that, The road zone map spatial constraints are used to ensure that each sampling point of the trajectory is within the passable area defined by the road zone map. If a trajectory point falls into a prohibited area, local adjustments are made through lateral offset correction, local interpolation, or resampling. The dynamic continuity constraints are used to ensure that the changes in the trajectory's velocity, acceleration, jerk, and angular velocity are continuous and within the vehicle's executable range. When the jerk exceeds the threshold or the angular velocity changes too much, spline interpolation or adjustment of the time step is used for smooth reconstruction.

6. The autonomous driving scene generation method according to claim 1, characterized in that, The risk function of the adaptive risk inference constraint is expressed as: In the formula, For the comprehensive risk value, , , and These are the weighting coefficients. TTC For potential collision time, Due to speed difference, For the occlusion ratio, Road pressure; When the overall risk value is lower than the set threshold, the trajectory risk level is improved by fine-tuning the vehicle's longitudinal speed, lateral position, or the time it takes to enter the critical area.

7. The autonomous driving scene generation method according to claim 1, characterized in that, The convergence conditions generated by the multi-round iterative optimization include at least one of the following: the number of iteration rounds reaches a preset upper limit, the trajectory satisfies all feasibility constraints, and the risk level of the trajectory reaches a preset target range.

8. An autonomous driving scene generation system, characterized in that, The system includes: The semantic factor extraction unit is configured to extract structured semantic factors from the input scene description information based on a predefined template and rule matching method. The semantic factors include driving behavior factors, spatial structure factors, and risk triggering factors. The road zoning map construction unit is configured to divide the road environment into multiple regional units with different traffic attributes based on high-precision map information, and to construct a road zoning map by defining the traffic connection relationship between regional units through a graph structure; wherein, the regional unit includes at least a main traffic area, an edge area, an intersection area, and a merging area; The coarse-grained trajectory generation unit is configured to determine the initial trajectory value and behavioral trend based on the extracted semantic factors, and generate a coarse-grained trajectory that satisfies the basic dynamic constraints according to the simplified vehicle dynamics model. The fine-grained trajectory reconstruction unit is configured to sequentially apply road zoning map spatial constraints, dynamic continuity constraints, and adaptive risk inference constraints to the coarse-grained trajectory to perform fine-grained adjustments and obtain an intermediate trajectory. The adaptive risk inference constraint evaluates the risk level of the trajectory based on a multi-dimensional risk inference model, which at least integrates time collision, speed difference, occlusion ratio, and road pressure index. The multi-round iterative optimization unit is configured to repeatedly execute trajectory generation, feasibility and risk assessment, and trajectory reconstruction based on assessment deviation based on the intermediate trajectory. After multiple rounds of iteration, it outputs a target trajectory that meets the requirements of physical executability, spatial legality, and preset risk.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the autonomous driving scene generation method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the autonomous driving scene generation method as described in any one of claims 1-7.

Images