Simulation test method and system for high-risk scene of automatic driving based on digital twinning

Abstract

This invention discloses a simulation testing method and system for high-risk scenarios of autonomous driving based on digital twins. The method includes: constructing a three-dimensional continuous traffic environment in a digital twin space based on a real road network topology and multi-source accident data; constructing an accident risk representation object based on the completeness of the multi-source accident data; generating high-risk test scenarios based on the types of accident risk representation objects; converting the high-risk test scenarios into test data corresponding to each functional module in the autonomous driving system, and injecting them into the corresponding functional modules; and outputting high-risk scenario test evaluation results based on the response results of the autonomous driving system. Because this invention constructs a continuous simulation environment by fusing real road network and accident data, and forms accident trigger chains or accident scenario fragments for complete and incomplete accident data respectively, it can comprehensively and accurately recreate various high-risk scenarios, effectively improving the coverage and evaluation accuracy of simulation tests of autonomous driving systems in complex high-risk scenarios.

Description

Technical Field

[0001] This invention relates to the field of automotive testing, and in particular to a simulation testing method and system for high-risk autonomous driving scenarios based on digital twins. Background Technology

[0002] With the development of intelligent connected vehicles and autonomous driving technologies, autonomous driving systems need to undergo simulation testing in numerous complex traffic scenarios to verify the safety response capabilities of functional modules such as perception, prediction, planning, and control under hazardous conditions. Among these, traffic accident scenarios, especially high-risk scenarios such as occluded crossings, sudden braking and cutting in, low visibility traffic, and conflicts at complex intersections, can reflect the safety performance of autonomous driving systems under long-tail risk conditions. Therefore, how to construct executable and evaluable high-risk simulation test scenarios based on real road traffic environments and accident data has become an important requirement in autonomous driving testing and verification.

[0003] Existing autonomous driving simulation testing methods typically generate test scenarios by manually editing scenario scripts, constructing static road environments based on high-precision maps, reproducing parts of the accident process by calling accident videos or accident reports, or by randomly disturbing vehicle speed, pedestrian positions, and obstacle positions. Some methods also input the generated scenario as a whole into the autonomous driving system to observe whether the autonomous vehicle experiences a collision, sudden braking, takeover, or abnormal trajectory, thereby judging the test results of the autonomous driving system in the corresponding scenario.

[0004] Existing practices suffer from the following significant drawbacks: First, real-world multi-source accident data often suffers from severe gaps or incompleteness. Existing methods lack adaptive assessment and differentiated processing mechanisms for data integrity, leading to the direct discarding of a large amount of incomplete, high-value accident data, or the distortion of scene logic due to forced piecing together, severely limiting the coverage of high-risk scenarios. Second, existing scene generation strategies often employ a single instantiation method, failing to execute matching generation strategies based on the inherent structural characteristics of the accident data, resulting in a lack of risk equivalence in the generated high-risk scenarios. Finally, existing test injection methods are mostly black-box, overall injections, unable to decouple the risk characteristics of high-risk scenarios and accurately transform them into specific test data required by different functional modules such as perception, prediction, planning, and control. This makes it impossible to conduct layered isolation testing, difficult to accurately locate specific failure nodes within the autonomous driving system, and severely lacking in the granularity and accuracy of test evaluation. Summary of the Invention

[0005] The main objective of this invention is to provide a simulation testing method and system for high-risk scenarios of autonomous driving based on digital twins. This aims to solve the technical problems of low utilization rate of accident data in existing technologies, difficulty in extracting key data from fragmented accident data, resulting in limited coverage of high-risk scenarios, and the inability of black-box test scenario injection to achieve accurate isolation testing and defect localization of multiple modules of autonomous driving systems.

[0006] To achieve the above objectives, this invention provides a simulation testing method for high-risk scenarios of autonomous driving based on digital twins, the method comprising the following steps: The topology and multi-source accident data of the real road network are obtained, and a three-dimensional continuous traffic environment is constructed in the digital twin space based on the topology and the multi-source accident data. An accident risk expression object is constructed based on the completeness of the multi-source accident data. When the multi-source accident data meets the preset completeness conditions, the accident risk expression object includes an accident trigger chain. When the multi-source accident data does not meet the preset completeness conditions, the accident risk expression object includes accident scene fragments. Based on the three-dimensional continuous traffic environment and the type of the accident risk expression object, a matching digital twin scenario instantiation strategy is executed on the accident risk expression object to generate a high-risk test scenario. The high-risk test scenarios are converted into test data corresponding to different functional modules in the autonomous driving system to be tested, and the test data is injected into the corresponding functional modules, including a perception module, a prediction module, a planning module, and a control module. Based on the response results of the autonomous driving system in the injected scenario, the test evaluation results of the high-risk scenario are output.

[0007] Furthermore, to achieve the above objectives, this invention also proposes a digital twin-based high-risk autonomous driving scenario simulation testing system that applies the digital twin-based high-risk autonomous driving scenario simulation testing method described above. The digital twin-based high-risk autonomous driving scenario simulation testing system includes: The traffic environment construction module is used to acquire the topology of the real road network and multi-source accident data, and to construct a three-dimensional continuous traffic environment in the digital twin space based on the topology and the multi-source accident data. The risk object construction module is used to construct an accident risk expression object based on the completeness of the multi-source accident data. When the multi-source accident data meets the preset completeness conditions, the accident risk expression object includes an accident trigger chain. When the multi-source accident data does not meet the preset completeness conditions, the accident risk expression object includes accident scene fragments. The test scenario construction module is used to generate high-risk test scenarios by executing a matching digital twin scenario instantiation strategy on the accident risk expression object based on the three-dimensional continuous traffic environment and the type of the accident risk expression object. The test data injection module is used to convert the high-risk test scenarios into test data corresponding to different functional modules in the autonomous driving system to be tested, and inject the test data into the corresponding functional modules. The functional modules include a perception module, a prediction module, a planning module, and a control module. The test evaluation module is used to output the test evaluation results of high-risk scenarios based on the response results of the autonomous driving system in the injected scenarios.

[0008] This invention constructs a three-dimensional continuous traffic environment using real road network topology and multi-source accident data, obtaining a simulation test foundation corresponding to real traffic conditions. It constructs accident trigger chains or accident scene fragments based on data completeness, compatible with both complete and incomplete accident data. High-risk test scenarios are generated through a matched digital twin scenario instantiation strategy, obtaining test inputs with accident risk mechanisms. This enables the formation of accident trigger chains or accident scene fragments for both complete and incomplete accident data, comprehensively and accurately recreating various high-risk scenarios. By converting high-risk test scenarios into test data corresponding to perception, prediction, planning, and control modules, a layered injection test process is formed. Through response result evaluation, high-risk scenario test evaluation results are output, obtaining safety response evaluation information for the autonomous driving system. This decouples high-risk scenarios into multi-dimensional test data for layered and precise injection and response evaluation of the perception, prediction, planning, and control modules of the autonomous driving system. This effectively overcomes the shortcomings of traditional testing, such as low accident data utilization, blind scenario generalization, and difficulty in locating internal defects in black-box overall testing. It significantly improves the scenario coverage, testing flexibility, and accuracy of system defect attribution in high-risk scenario simulation testing for autonomous driving. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the structure of a high-risk autonomous driving scenario simulation test device based on digital twin, which is part of the hardware operating environment of the embodiment of the present invention. Figure 2 This is a flowchart illustrating the first embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins according to the present invention. Figure 3 This is a flowchart illustrating the second embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins according to the present invention. Figure 4 This is a flowchart illustrating the third embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins according to the present invention. Figure 5 This is a flowchart illustrating the fourth embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins according to the present invention. Figure 6 This is a flowchart illustrating the fifth embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins according to the present invention. Figure 7 This is a flowchart illustrating the sixth embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins according to the present invention. Figure 8 This is a flowchart illustrating the seventh embodiment of the simulation and testing method for high-risk autonomous driving scenarios based on digital twins according to the present invention. Figure 9 This is a flowchart illustrating the eighth embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins according to the present invention. Figure 10 This is a flowchart illustrating the ninth embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins according to the present invention. Figure 11 This is a structural block diagram of the first embodiment of the high-risk scenario simulation and testing system for autonomous driving based on digital twins of the present invention.

[0011] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0012] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0013] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a high-risk autonomous driving scenario simulation test device based on digital twins, which is part of the hardware operating environment of the embodiment of the present invention.

[0014] like Figure 1As shown, the digital twin-based high-risk autonomous driving scenario simulation test device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen and an input unit such as a keyboard; the user interface 1003 may also include standard wired and wireless interfaces. The network interface 1004 may optionally include standard wired and wireless interfaces (such as Wireless-Fidelity (Wi-Fi) interfaces). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0015] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the simulation test equipment for high-risk autonomous driving scenarios based on digital twins. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0016] like Figure 1 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a simulation test program for high-risk autonomous driving scenarios.

[0017] exist Figure 1 In the digital twin-based autonomous driving high-risk scenario simulation test device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and memory 1005 in the digital twin-based autonomous driving high-risk scenario simulation test device of the present invention can be set in the digital twin-based autonomous driving high-risk scenario simulation test device. The digital twin-based autonomous driving high-risk scenario simulation test device calls the autonomous driving high-risk scenario simulation test program stored in the memory 1005 through the processor 1001 and executes the autonomous driving high-risk scenario simulation test method provided in the embodiment of the present invention.

[0018] This invention provides a simulation testing method for high-risk autonomous driving scenarios based on digital twins, referring to... Figure 2 , Figure 2This is a flowchart illustrating the first embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins according to the present invention.

[0019] In this embodiment, the simulation testing method for high-risk autonomous driving scenarios based on digital twins includes the following steps: Step S10: Obtain the topology of the real road network and multi-source accident data, and construct a three-dimensional continuous traffic environment in the digital twin space based on the topology and the multi-source accident data.

[0020] It should be understood that the executing entity of this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a terminal electronic device capable of performing the above functions. The following description uses a digital twin-based high-risk autonomous driving scenario simulation test device (hereinafter referred to as the test device) as an example to illustrate this embodiment and the following embodiments.

[0021] It should be noted that the topology of a real road network refers to the data structure used to represent the connectivity between lanes, intersections, ramps, road boundaries, traffic signals, and directions of travel in a real road. Multi-source accident data can come from accident reports, accident videos, on-board sensors, roadside equipment, vehicle driving recorders, traffic management platforms, or accident-related data from manually labeled results.

[0022] A digital twin space refers to a virtual space used for virtual mapping, dynamic simulation, and testing of real road traffic scenarios. A three-dimensional continuous traffic environment refers to a simulation environment that continuously represents roads, vehicles, pedestrians, non-motorized vehicles, obstacles, and environmental factors in space, and supports dynamic changes of traffic objects in time.

[0023] In practical implementation, the testing equipment can acquire the topology of the real road network from high-precision maps, road design documents, traffic signal configuration files, or roadside perception systems, and obtain multi-source accident data from accident databases, accident videos, accident reports, vehicle driving records, roadside camera recordings, vehicle sensor recordings, or traffic management platforms. Then, the testing equipment maps road connectivity, number of lanes, lane boundaries, traffic signal locations, road speed limits, and intersection traffic directions into a digital twin space, and combines this with the accident location, the location of traffic participants, their movement trajectories, weather conditions, lighting conditions, and road visibility conditions to construct a three-dimensional continuous traffic environment in the digital twin space.

[0024] For example, in a pedestrian crossing accident at an urban intersection, the testing equipment can construct a three-dimensional model of the intersection based on the lane connections, zebra crossing location, traffic light location, and vehicle travel direction of the actual intersection. Based on the accident video and vehicle travel records, the equipment can configure the accident vehicle, pedestrian, obstructing vehicle, and ambient lighting status in the three-dimensional model.

[0025] Step S20: Construct an accident risk representation object based on the completeness of the multi-source accident data.

[0026] It should be noted that the completeness of multi-source accident data refers to the extent to which multi-source accident data covers the pre-accident process, accident location, accident object status, traffic environment status, and system response status.

[0027] Preset integrity conditions refer to pre-defined data integrity judgment conditions, which may include time continuity conditions, spatial coverage conditions, object status coverage conditions, and pre-accident process coverage conditions.

[0028] An accident risk representation object refers to a structured object used to represent the accident risk formation process or local factors of accident risk, which may include accident triggering chains and accident scene fragments.

[0029] Among them, the accident triggering chain refers to the chain of accident evolution formed by multiple risk events according to time sequence, spatial conflict relationship, and risk transmission relationship. Accident scene fragments refer to reusable local risk elements extracted from incomplete accident data.

[0030] It should be noted that when the multi-source accident data meets the preset integrity conditions, the accident risk expression object includes the accident trigger chain; when the multi-source accident data does not meet the preset integrity conditions, the accident risk expression object includes accident scene fragments.

[0031] It is understood that this embodiment analyzes the completeness of multi-source accident data to separate complete and incomplete accident data, thereby enabling compatibility with accident data of different quality and forming an accident risk expression object that is adapted to subsequent scenario instantiation.

[0032] In practical implementation, the testing equipment can detect whether multi-source accident data includes key time periods before the accident, accident object trajectories, accident spatial locations, road environment conditions, traffic participant states, and autonomous driving response states. If the multi-source accident data contains data that can reconstruct the accident evolution process, the testing equipment can extract the risk change process before the accident from the multi-source accident data and organize events such as target appearance, occlusion removal, deceleration of the vehicle in front, cutting in from the side, pedestrian crossing, traffic light changes, and autonomous driving system braking response into an accident trigger chain according to time sequence and risk transmission relationships. If the multi-source accident data lacks complete trajectories, complete video, or key time periods before the accident, the testing equipment can extract local risk elements such as road structure, traffic objects, object behavior, environmental conditions, occlusion relationships, conflict relationships, and autonomous driving responses from the existing data, and organize these local risk elements into accident scene fragments.

[0033] For example, if the accident data includes vehicle trajectories, pedestrian trajectories, intersection videos, and vehicle braking records before the accident, an accident trigger chain can be constructed: "obstructed vehicle parking - pedestrian emerging from the obstructed area - delayed recognition of the main vehicle - insufficient braking of the main vehicle." If the accident data only includes accident reports and accident scene photos, fragments of accident scenarios such as "bus obstruction," "nighttime intersection," "pedestrian crossing," and "insufficient vehicle braking" can be extracted.

[0034] Step S30: Based on the three-dimensional continuous traffic environment and the type of the accident risk expression object, execute a matching digital twin scenario instantiation strategy on the accident risk expression object to generate a high-risk test scenario.

[0035] It should be noted that the digital twin scenario instantiation strategy refers to the processing method of converting accident risk representation objects into scenario objects, scenario parameters, and scenario processes that can be executed in the digital twin space.

[0036] High-risk test scenarios refer to autonomous driving test scenarios that can cause collision risks, emergency braking risks, takeover risks, planning anomaly risks, or control execution risks.

[0037] It is understood that this embodiment can generate high-risk test scenarios with accident risk mechanisms by selecting a matching instantiation strategy according to the type of accident risk expression object.

[0038] In its implementation, the testing equipment can identify the type of accident risk representation object. If the accident risk representation object is an accident trigger chain, the testing equipment maps each trigger node in the accident trigger chain to the road location, traffic object, object behavior, and environmental state in a three-dimensional continuous traffic environment, and generates a basic accident scenario based on the temporal relationship between the nodes. Subsequently, risk-equivalent perturbations can be applied to the target appearance time, traffic object speed, object spacing, occlusion release time, braking delay, and road surface adhesion state to generate multiple high-risk test scenarios. If the accident risk representation object is an accident scene fragment, the testing equipment can first complete the missing time, missing location, missing trajectory, or missing occlusion relationship based on the road geometric constraints, vehicle motion constraints, traffic rule constraints, and sensor visibility constraints in the three-dimensional continuous traffic environment, and then splice together multiple accident scene fragments according to conflict geometric relationships, temporal compatibility relationships, and risk action relationships to generate a high-risk test scenario.

[0039] For example, for a complete accident trigger chain, while maintaining the risk mechanism of "pedestrian crossing after obstruction is removed causing insufficient braking of the main vehicle," the timing of the pedestrian's appearance and the initial speed of the main vehicle can be adjusted to create multiple test scenarios with different levels of danger. For fragmented accident scenarios, fragments such as "low light at night in rainy weather," "obstruction by large vehicles," "lateral intrusion by non-motorized vehicles," and "limited space for the main vehicle to avoid obstacles" can be combined into rare high-risk test scenarios.

[0040] Step S40: Convert the high-risk test scenarios into test data corresponding to different functional modules in the autonomous driving system to be tested, and inject the test data into the corresponding functional modules.

[0041] It should be noted that the functional modules of the autonomous driving system under test include a perception module, a prediction module, a planning module, and a control module.

[0042] The perception module refers to the functional module used to identify the road, traffic objects, obstacles, lane lines and traffic signals around the vehicle.

[0043] The prediction module refers to the functional module used to predict the future trajectory or behavioral intentions of traffic participants.

[0044] The planning module refers to the functional module used to generate driving paths, speed strategies, and obstacle avoidance strategies for autonomous vehicles.

[0045] A control module is a functional module used to control a vehicle to perform steering, acceleration, deceleration, or braking operations.

[0046] It is understood that this embodiment achieves precise stimulation and isolation testing of each independent module within the autonomous driving system by inputting test data in layers to each functional module of the autonomous driving system, thereby improving the flexibility and targeting of the test.

[0047] In practical implementation, the testing equipment can parse the road environment, traffic object trajectories, environmental states, risk triggering nodes, and vehicle operating states in high-risk test scenarios. Based on the input interface formats of each functional module in the autonomous driving system under test, it converts the high-risk test scenarios into different types of test data. For the perception module, it can generate camera images, point cloud data, radar target data, or multi-sensor fusion data; for the prediction module, it can generate traffic participant trajectories, traffic participant behavioral intentions, and motion probability information; for the planning module, it can generate road boundaries, dynamic obstacles, drivable areas, and traffic rule constraints; for the control module, it can generate target trajectories, target speeds, braking response parameters, and steering response parameters. Then, different test data are injected into the corresponding functional modules according to their respective interfaces.

[0048] For example, in a scenario where pedestrians are obstructed while crossing the road, the testing equipment can inject images and point cloud data at the moment the obstruction is removed into the perception module, inject the pedestrian crossing trajectory into the prediction module, inject the pedestrian movement area and road boundary into the planning module, and inject the emergency braking target speed and target trajectory into the control module.

[0049] Step S50: Based on the response results of the autonomous driving system in the injected scenario, output the high-risk scenario test evaluation results.

[0050] It should be noted that the response result refers to the perception, prediction, planning, control, or overall vehicle operation results generated by the autonomous driving system after input in a high-risk test scenario.

[0051] The evaluation results of high-risk scenario tests refer to the evaluation information used to characterize the safety response capability of autonomous driving systems in high-risk test scenarios.

[0052] Scene evaluation metrics can include collision risk, minimum safe distance, minimum collision time, detection misses, prediction deviations, planned trajectory deviations, and control tracking errors.

[0053] In practical implementation, the testing equipment can collect the response results of the autonomous driving system in the injected scenario. The response results can include target recognition results, target tracking results, behavior prediction results, planned trajectory, control output, actual vehicle trajectory, collision results, takeover results, emergency braking results, and traffic rule violation results. Then, the testing equipment can calculate scenario evaluation indicators based on the response results, such as collision risk, minimum safe distance, minimum collision time, perception misses, prediction deviations, planned trajectory deviations, and control tracking errors, and output high-risk scenario test evaluation results based on the scenario evaluation indicators.

[0054] For example, if an autonomous driving system has issues such as missed detections, braking delays, and insufficient minimum safe distance when pedestrians are crossing the road, the testing equipment can output the risk level, abnormal functional modules, and abnormal response types corresponding to that scenario.

[0055] This embodiment maps the real road network topology and multi-source accident data to a digital twin space, enabling a continuous three-dimensional traffic environment that simultaneously includes road connectivity, traffic participant status, accident environmental conditions, and accident occurrence process information. This avoids the lack of real accident constraints when generating scenarios solely based on static maps or manual scripts, thus providing subsequent high-risk test scenarios with a foundation in real road structures and accident risk sources. Accident trigger chains and accident scenario fragments are constructed based on the completeness of the multi-source accident data, allowing complete accident data to express the gradual accumulation of risk in the accident evolution process, while incomplete accident data can also participate in scenario generation through local risk elements, preventing the direct discarding of incomplete accident data. Simultaneously, a matching digital twin scenario instantiation strategy is executed based on the type of accident risk expression object, enabling accident trigger chains to be transformed into high-risk scenarios with continuous risk transmission relationships, and accident scenario fragments to be transformed into combined scenarios with consistent risks. Finally, high-risk test scenarios are converted into test data corresponding to the perception, prediction, planning, and control modules, allowing the same accident risk mechanism to be injected and verified at different functional levels, thereby achieving synergy between high-risk scenario generation, module-level testing, and anomaly source localization.

[0056] refer to Figure 3 , Figure 3 This is a flowchart illustrating the second embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins according to the present invention.

[0057] Based on the first embodiment described above, in this embodiment, step S10 further includes: Step S101: Obtain the topology of the real road network and multi-source accident data, and generate a road network topology map based on the topology of the real road network.

[0058] It should be noted that a road network topology map refers to graph structure data used to represent the relationships between lanes, intersections, traffic signals, and road traffic in a real road, including lane nodes, road connecting edges, and traffic signal nodes. The nodes and edges in the road network topology map are associated with road boundary attributes and road traffic rule attributes.

[0059] Lane nodes are graph nodes that represent a single lane or lane segment; road connection edges are graph edges that represent the traffic connection relationship between different lane nodes; road boundary attributes can be boundary information such as lane lines, guardrails, curbs, and medians; road traffic rule attributes can be traffic constraints such as speed limits, yielding, turning restrictions, and traffic light control.

[0060] In practical implementation, the testing equipment obtains the real road network topology of the target test area from high-precision maps, road design documents, traffic signal configuration files, or road management platforms, and obtains multi-source accident data from accident databases, accident videos, vehicle driving records, roadside perception records, or traffic management platforms. Then, the testing equipment converts lane centerlines, lane boundaries, intersection connections, ramp connections, traffic light positions, and road traffic directions into graph structure data, representing lanes, intersections, and traffic lights as nodes, and lane connections, turning relationships, and traffic relationships as edges. It also configures attributes such as road width, lane boundaries, speed limit rules, turning rules, yielding rules, and no-entry rules for nodes and edges, thereby generating a road network topology map.

[0061] Step S102: Extract the temporal features of the accident scene from the multi-source accident data.

[0062] It should be noted that the temporal characteristics of an accident scene refer to a set of features that express the movement of the accident object, changes in the environment, changes in road visibility, and changes in sensor observations in chronological order, including the trajectory of the accident object, the state of the accident environment, the state of road visibility, and the state of sensor observations.

[0063] Accident object trajectory refers to the sequence of positional changes of accident vehicles, pedestrians, non-motorized vehicles, or obstacles during the accident process; road visibility status refers to the visible, partially visible, or invisible status of road areas, traffic objects, or conflict areas in the sensor's field of view; sensor observation status refers to the observation results of traffic objects and road environment by sensors such as cameras, lidar, and millimeter-wave radar.

[0064] It is understandable that this embodiment unifies the object movement, environmental conditions, road visibility, and sensor observation status in the accident data into a time series, so that the subsequent scene construction can express the process of accident risk accumulation over time, avoiding the problem of missing risk evolution information caused by constructing a scene only based on the instant the accident occurs.

[0065] In its implementation, the testing equipment timestamps, marks the data source, and aligns the objects in the multi-source accident data. It also extracts the movement trajectories of accident vehicles, pedestrians, non-motorized vehicles, obstructions, and other traffic objects before and after the accident from accident videos, vehicle trajectory records, roadside perception records, and accident reports. Simultaneously, it extracts information such as weather, lighting, road surface conditions, obstruction range, visibility distance, sensor detection results, and sensor missed detection records, and forms the temporal characteristics of the accident scene according to the time sequence before and after the accident.

[0066] Step S103: Construct a traffic semantic graph based on the road network topology map and the temporal features of the accident scenario.

[0067] It should be noted that a traffic semantic graph refers to graph-structured data that simultaneously expresses the relationships between road structure, traffic participants, environmental conditions, and observation information, including road nodes, traffic participant nodes, environmental nodes, and observation nodes.

[0068] Road nodes refer to nodes used to represent road components such as lanes, intersections, zebra crossings, and ramps; traffic participant nodes refer to nodes used to represent dynamic objects such as vehicles, pedestrians, and non-motorized vehicles; environmental nodes refer to nodes used to represent environmental factors such as weather, lighting, road surface wetness, and obstruction; and observation nodes refer to nodes used to represent the observation results of sensors on roads or traffic objects.

[0069] It is understood that this embodiment expresses road topology, accident objects, environmental factors and sensor observations in a unified heterogeneous graph structure, so that spatial relationships, behavioral relationships and observational relationships in the accident scene can participate in subsequent coding together, overcoming the problem of fragmented expression of road modeling and accident data.

[0070] In its implementation, the testing equipment uses the road network topology map as the basis for the road structure. It converts the vehicles, pedestrians, non-motorized vehicles, static obstacles, weather conditions, lighting conditions, occlusion conditions, and sensor observation results in the temporal features of the accident scene into different types of semantic nodes. Semantic edges are established based on the lane where the object is located, the positional relationship between the object and the road boundary, the relative motion relationship between objects, the influence of environmental conditions on visibility, and the correspondence of sensor observations, thereby constructing a traffic semantic map.

[0071] Step S104: Input the traffic semantic graph into the spatiotemporal graph neural network to encode the node relationships and obtain the spatiotemporal correlation features of the road network.

[0072] It should be noted that spatiotemporal graph neural networks refer to neural networks used to jointly encode spatial connectivity and temporal variation relationships in graph structure data; node relationship encoding refers to converting node attributes, node connectivity, and node temporal changes into feature vectors; and road network spatiotemporal correlation features refer to features that can express the correlation between the real road network structure and the dynamic changes of accident scenarios.

[0073] It is understandable that this embodiment uses a spatiotemporal graph neural network to jointly encode the traffic semantic graph, so that the road connection relationship, traffic object interaction relationship and accident process time change relationship are modeled in a unified way, avoiding the problem that it is difficult to express complex traffic interactions when using static maps or single trajectory data alone.

[0074] In the specific implementation, the test equipment inputs the attribute vectors of different types of nodes in the traffic semantic graph, the semantic edges between nodes, and the time series into the spatiotemporal graph neural network. Through graph convolution, graph attention, or temporal gating units, the spatial relationships within the same time segment and the dynamic changes between adjacent time segments are encoded to obtain the spatiotemporal correlation features of the road network that reflect the correlation between road structure, traffic object movement, environmental changes, and observation status.

[0075] In some embodiments, the spatiotemporal graph neural network can be trained based on historical traffic scene graphs, accident scene semantic graphs, and manually annotated spatiotemporal association labels; the training objectives include node state prediction, edge relationship prediction, traffic object interaction relationship prediction, or accident scene temporal reconstruction, etc.

[0076] Step S105: Construct a continuous spatial representation model based on the spatiotemporal correlation features of the road network.

[0077] It should be noted that the continuous spatial representation model refers to the model used to continuously express the spatial occupancy relationship and observation presentation relationship of a three-dimensional traffic scene, including implicit occupancy field and neural radiation field.

[0078] Implicit occupancy field refers to a model that expresses the drivable area, obstacle area, and traffic object area in three-dimensional space through the mapping relationship between spatial location and occupancy status; neural radiation field refers to a model that generates viewpoint-related observation results through spatial location, viewing direction, and scene attributes.

[0079] In some embodiments, the implicit occupancy field can be trained using road boundary points, traffic object bounding boxes, drivable area annotations, and obstacle area annotations; the neural radiation field can be trained or fitted using multi-view accident videos, roadside camera images, simulation rendering images, and camera parameters.

[0080] In the specific implementation, the test equipment inputs the spatiotemporal correlation features of the road network into the continuous space modeling module, and generates spatial occupancy information based on road boundaries, traffic object locations, occlusion areas, and environmental conditions. At the same time, it generates scene rendering information based on illumination, viewing angle, object surface attributes, and sensor observation status. The continuous space modeling module expresses the occupancy relationship of roads, vehicles, pedestrians, obstacles, and drivable areas in three-dimensional space through implicit occupancy fields, and expresses the visual observation results under different viewing angles through neural radiation fields, thereby forming a continuous space expression model.

[0081] Step S106: Generate a three-dimensional continuous traffic environment in the digital twin space based on the continuous spatial representation model.

[0082] In the specific implementation, the testing equipment generates road geometry, traffic participant models, environmental state models, sensor observation models, and traffic rule constraints based on the continuous spatial representation model, and loads these models into the digital twin space according to a unified spatial coordinate and a unified time axis. During the loading process, the movement trajectory of traffic objects is configured according to the temporal characteristics of the accident scene, the road connection relationship and traffic rules are configured according to the road network topology map, and the drivable area, occluded area and sensor observation area are configured according to the continuous spatial representation model, thereby forming a three-dimensional continuous traffic environment.

[0083] This embodiment uses a road network topology map to represent the real road structure and traffic rules, uses the temporal features of accident scenarios to represent the dynamic process of accidents, uses a traffic semantic map to integrate the relationships between roads, objects, environment and observation, uses a spatiotemporal graph neural network to extract the spatiotemporal correlation features of the road network, and uses implicit occupancy fields and neural radiation fields to construct a continuous spatial expression model, so that the three-dimensional continuous traffic environment has the ability to simultaneously have road rule constraints, traffic interaction expression and sensor observation generation capabilities.

[0084] refer to Figure 4 , Figure 4 This is a flowchart illustrating the third embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins of the present invention.

[0085] Based on the above embodiments, in this embodiment, step S20 further includes: Step S201: Extract multimodal integrity features from the multi-source accident data.

[0086] It should be noted that multimodal integrity features refer to a set of features used to characterize the completeness of accident data in multiple dimensions such as data source, time, space, object status, and pre-accident process. These features include data source integrity features, temporal continuity features, spatial coverage integrity features, traffic participant object status integrity features, and pre-accident process integrity features.

[0087] Among them, the data source integrity feature refers to the feature used to characterize whether accident data comes from different sources such as accident reports, accident videos, vehicle-mounted sensors, roadside sensors, and vehicle driving records; the temporal continuity feature refers to the feature used to characterize whether there are time discontinuities in the data before and after the accident; the spatial coverage integrity feature refers to the feature used to characterize whether the accident area, conflict area, and object activity area are covered by the accident data; the traffic participant status integrity feature refers to the feature used to characterize whether the position, speed, direction, and behavior status of vehicles, pedestrians, non-motorized vehicles, or obstacles are complete; and the pre-accident process integrity feature refers to the feature used to characterize whether the risk formation process before the accident is recorded.

[0088] It is understandable that this embodiment decomposes the integrity of accident data into multiple dimensions such as source, time, space, object status and pre-accident process, so that subsequent integrity judgment does not rely on a single missing marker, but can identify whether the accident data is sufficient to support the reconstruction of the accident process or the extraction of scene fragments.

[0089] In some embodiments, the testing equipment performs source identification, timestamp parsing, spatial coordinate verification, traffic object status identification, and pre-accident process coverage checks on multi-source accident data. This yields the availability relationship between different accident data sources, the continuity of accident data on the time axis, the coverage of the accident area in the spatial range, the completeness of the status records of vehicles, pedestrians, non-motorized vehicles, or obstacles, and the completeness of the record of the risk evolution process before the accident. These features are then uniformly encoded to form multimodal integrity features.

[0090] Step S202: Input the multimodal integrity features into a masked autoencoder network to reconstruct missing items, thereby obtaining reconstructed accident features and missing pattern features.

[0091] It should be noted that a masked autoencoder network is a neural network that masks some data items in the input features and reconstructs the masked data items based on the unmasked data items.

[0092] Reconstruction incident features refer to the incident features obtained by the masked autoencoder network after estimating missing data items, anomalous data items, or low-confidence data items. Missing pattern features refer to features used to characterize the location, category, correlation, and reconstruction difficulty of incident data.

[0093] It is understood that this embodiment transforms the missing content in the accident data into computable reconstructable features and missing pattern features, so that the missing data is no longer just used as a static basis for judging whether it is complete, but can participate in subsequent integrity judgment and accident scene fragment construction.

[0094] In the specific implementation, the test equipment converts the multimodal integrity features into feature sequences and configures mask tags for missing data items, abnormal data items, or low-confidence data items. Then, the feature sequences with mask tags are input into a mask autoencoder network. The mask autoencoder network uses the contextual relationships between the incident features that are not masked to reconstruct the masked data items, outputs reconstructed incident features, and generates missing mode features based on the location, type, reconstruction difficulty, and reconstruction deviation of the masked data items.

[0095] Step S203: Generate data consistency features based on the reconstructed accident features and the multimodal integrity features.

[0096] It should be noted that data consistency features refer to features used to characterize the degree of matching, support, and conflict between the original accident data and the reconstructed accident features. For example, if the vehicle trajectory in the accident video matches the speed changes in the vehicle driving record, a high degree of consistency can be formed; if the collision location in the accident report does not match the conflict location in the roadside video, a conflict feature can be formed.

[0097] It is understandable that this embodiment compares the reconstruction results with the original integrity information to ensure that the integrity assessment can simultaneously identify both "data missing" and "data contradiction" issues, thus avoiding the need to judge whether accident data can be used to construct an accident triggering chain based solely on the number of data items.

[0098] In the specific implementation, the test equipment will dimensionally align the reconstructed accident features with the multimodal integrity features, and compare the differences, supporting relationships and conflicting relationships between the two in dimensions such as data source, temporal continuity, spatial coverage, state of traffic participants and pre-accident process; then, feature encoding will be performed on the differences, supporting relationships and conflicting relationships to generate data consistency features.

[0099] Step S204: Construct a completeness discrimination vector based on the missing pattern features and the data consistency features.

[0100] It should be noted that the integrity discrimination vector refers to the vectorized feature used as input to the integrity discrimination model and to characterize the integrity of accident data. It can include information such as missing structure, reconstruction credibility, original data credibility, and cross-source consistency.

[0101] It is understood that this embodiment encodes missing patterns and data consistency into a completeness discrimination vector, enabling subsequent discrimination models to utilize both missing structural information and data credibility relationships simultaneously, thereby reducing misjudgments caused by judging solely based on the missing proportion.

[0102] In the specific implementation, the test equipment performs feature concatenation, normalization and weight allocation on the missing pattern features and data consistency features, and encodes the missing location, missing type, reconstruction confidence level, data consistency level and data conflict level into feature representations in the same vector space to obtain the integrity discrimination vector.

[0103] Step S205: Input the integrity discrimination vector into the graph attention discrimination network to perform integrity discrimination and obtain the integrity judgment result.

[0104] It should be noted that the graph attention discriminant network is used to learn the association weights between different nodes or different feature dimensions in a graph structure through the attention mechanism, and output the discriminant result.

[0105] It is understandable that this embodiment assigns weights to different integrity dimensions through a graph attention discriminant network, so that key dimensions such as pre-accident process coverage, traffic participant status and spatial conflict area coverage can be highlighted in integrity discrimination, avoiding the weakening of key deficiencies due to simple averaging of multiple dimensions.

[0106] In the specific implementation, the test equipment inputs the integrity discrimination vector into the graph attention discrimination network. The graph attention discrimination network assigns attention weights according to the correlation between data source, temporal continuity, spatial coverage, object state and pre-accident process, and performs fusion discrimination on features of each dimension to output integrity judgment results.

[0107] Step S206: If the integrity determination result indicates that the multi-source accident data meets the preset integrity conditions, then an accident triggering chain is constructed based on the multi-source accident data.

[0108] It is understood that this embodiment constructs an accident trigger chain when the accident data is sufficient to reconstruct the accident evolution process, so that the risk formation, risk triggering and risk enhancement processes in the complete accident data can be directly used as the basis for subsequent high-risk scenario instantiation, avoiding the complete accident data being simplified into a single accident result.

[0109] In some embodiments, when the integrity determination result indicates that the multi-source accident data meets the preset integrity conditions, the test equipment extracts continuous events before the accident, traffic object interaction process, road environment change process and autonomous driving response process from the multi-source accident data, and organizes the risk events according to the time sequence, spatial conflict relationship and risk enhancement relationship to construct the accident trigger chain.

[0110] Step S207: If the integrity determination result indicates that the multi-source accident data does not meet the preset integrity conditions, then an accident scene fragment is constructed based on the missing pattern feature and the multi-source accident data.

[0111] It is understandable that this embodiment retains reusable local risk elements when accident data cannot form a complete accident triggering chain, so that incomplete accident data can still be transformed into input for subsequent scenario completion and risk consistency splicing, avoiding the discarding of incomplete accident data.

[0112] In some embodiments, when the integrity determination result indicates that the multi-source accident data does not meet the preset integrity conditions, the test equipment determines the missing parts in the accident data that cannot be continuously restored based on the missing pattern characteristics, and extracts local risk elements such as road structure, traffic objects, object behavior, environmental state, occlusion relationship, conflict relationship and autonomous driving response that can still be used independently from the multi-source accident data, and then organizes the local risk elements into accident scene fragments.

[0113] Step S208: Construct an accident risk expression object based on the accident triggering chain or the accident scene fragments.

[0114] In practice, the testing equipment converts accident trigger chains or accident scene fragments into unified structured objects, and configures risk type, time attribute, spatial attribute, traffic object attribute, environmental attribute, missing state attribute and instantiation strategy mark for the unified structured objects to form an accident risk expression object.

[0115] This embodiment describes the coverage of accident data in terms of source, time, space, object state, and pre-accident process through multimodal integrity features. It extracts and reconstructs accident features and missing pattern features through a masked autoencoder network, identifies matching or conflicting relationships between original data and reconstructed data through data consistency features, outputs integrity judgment results through a graph attention discriminant network, and constructs accident trigger chains or accident scene fragments based on the integrity judgment results. This creates a continuous technical connection between accident data integrity judgment, selection of accident risk expression objects, and subsequent digital twin scene instantiation.

[0116] refer to Figure 5 , Figure 5 This is a flowchart illustrating the fourth embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins of the present invention.

[0117] Based on the third embodiment described above, in this embodiment, step S206 may include: Step S2061: Extract the accident-related time sequence data within a preset time window before the accident occurs from the multi-source accident data.

[0118] It should be noted that accident-related time-series data refers to the data sequence related to risk formation, risk triggering, risk amplification, and risk avoidance response within a preset time window before the accident occurs, including the status of traffic participants, road traffic status, environmental status, obstruction status, traffic signal status, and autonomous driving response status. The pre-set time window before an accident refers to a pre-defined time range that precedes the time of the accident and is used to trace the evolution of accident risk. The status of traffic participants can be the position, speed, direction, acceleration, and behavior of vehicles, pedestrians, non-motorized vehicles, or obstacles. The autonomous driving response status refers to the perception and recognition results, behavior prediction results, planning and decision-making results, and control execution results generated by the autonomous driving system during the evolution of accident risk.

[0119] Step S2062: Divide the accident-related time-series data into time-series segments based on a preset time segment length to obtain multiple accident time-series segments.

[0120] It should be noted that the preset time segment length refers to the pre-set time length used to divide the accident-related time series data; the accident time series segment refers to the data segment obtained from the accident-related time series data according to the time sequence.

[0121] It is understood that this embodiment divides the continuous accident process into multiple data segments with unified time boundaries, so that changes in risk status can be located within specific segments, which facilitates subsequent comparison of changes in traffic objects, environmental changes, and system response changes between adjacent segments.

[0122] In the specific implementation, the test equipment continuously segments the accident-related time-series data according to the preset time segment length, and configures the segment start time, segment end time, traffic participant status set, road traffic status set, environmental status set, occlusion status set, traffic signal status set, and autonomous driving response status set for each accident time-series segment. When the accident data sampling frequency is inconsistent, high-frequency data can be time-aggregated and low-frequency data can be time-merged, so that each accident time-series segment has a comparable data structure.

[0123] Step S2063: Perform risk change factor analysis on each accident time segment to obtain a risk change factor sequence.

[0124] It should be noted that risk change factors refer to scenario factors that cause changes in risk status during the evolution of an accident, including target appearance status, obstruction removal status, deceleration status of traffic participants, lateral intrusion status of traffic participants, traffic signal change status, and autonomous driving response change status.

[0125] The target appearance state refers to the state in which a traffic object enters an observable conflict area from an invisible, partially visible, or non-conflict area; the occlusion removal state refers to the state in which the occlusion relationship between the occluder and the target object is weakened or disappears; the autonomous driving response change state refers to the state in which the autonomous driving system changes in its perception, prediction, planning, or control output.

[0126] It is understood that this embodiment extracts risk change factors through intra-segment analysis and adjacent segment difference analysis, so that the source of risk can be determined by the object movement, environmental changes, occlusion changes, signal changes and system response changes, avoiding the identification of accident triggering factors based solely on a single trajectory change or manual annotation.

[0127] In practical implementation, the testing equipment performs intra-segment analysis and adjacent segment difference analysis on the status of traffic participants, road traffic status, environmental status, occlusion status, traffic signal status, and autonomous driving response status in each accident time segment. It also identifies risk change factors such as target objects entering the visible area from the invisible area, the occlusion area decreasing, the speed of the vehicle in front decreasing, the vehicle or pedestrian entering the main vehicle's driving area laterally, traffic signal phase changes, changes in the perception results of the autonomous driving system, changes in the prediction results, changes in the planned trajectory, or changes in the control output. Then, it generates a sequence of risk change factors according to the order of the accident time segments.

[0128] Step S2064: Based on the risk change factor sequence, construct a trigger node sequence according to the temporal relationship of the multiple accident time segments.

[0129] It is understood that this embodiment transforms risk change factors into trigger nodes with time, space, object, and risk type attributes, so that the continuous accident process is structured into a sequence of trigger nodes that can be mapped, compared, and filtered.

[0130] In its implementation, the testing equipment converts factors in the risk change factor sequence that have the functions of risk formation, risk triggering, risk amplification, or risk avoidance failure into trigger nodes. For each trigger node, it configures the node occurrence time, node occurrence location, participating objects, risk type, and associated segment identifier. Then, it sorts the trigger nodes according to the chronological relationship of multiple accident time segments to obtain the trigger node sequence.

[0131] Step S2065: Map the trigger node sequence to the three-dimensional continuous traffic environment to obtain a node twin scene corresponding to each trigger node.

[0132] It should be noted that a node twin scene refers to a local scene instance constructed around a single trigger node in a three-dimensional continuous traffic environment.

[0133] It is understood that this embodiment transforms the abstract trigger node into a runnable node twin scene, enabling the importance of the trigger node to be verified in a simulation environment with road geometry, object motion, environmental occlusion, and system response constraints.

[0134] In the specific implementation, the test equipment reads the occurrence time, location, participating objects, and risk type of each trigger node in the trigger node sequence, and configures the corresponding road area, traffic participating object model, environmental conditions, occlusion relationship, signal status, and autonomous driving response status in the three-dimensional continuous traffic environment to generate a local simulation scene corresponding to each trigger node. For example, for the trigger node "pedestrian enters the lane after occlusion is removed", the occluding vehicle, pedestrian position, main vehicle position, visibility status, and pedestrian crossing behavior can be configured in the three-dimensional continuous traffic environment to form a node twin scene.

[0135] Step S2066: Perform counterfactual processing on the corresponding trigger node in each node twin scenario to obtain the counterfactual twin scenario.

[0136] It should be noted that counterfactual processing refers to the process of removing or replacing the target trigger node while keeping other scenario conditions unchanged, including removing or replacing the trigger node.

[0137] Counterfactual twin scenarios refer to the contrast scenario formed after counterfactual processing of the target trigger node in a node twin scenario; the contrast node refers to the scenario node used to replace the target trigger node and has a lower risk state.

[0138] It is understood that this embodiment constructs a control scenario that only changes the target trigger node, so that the independent effect of the target trigger node on the change of accident risk can be identified, reducing the interference of road background, environmental background or ordinary traffic flow factors on causal judgment.

[0139] In the specific implementation, the test equipment performs counterfactual processing on the current trigger node in each node twin scenario, while keeping the road structure, basic traffic flow, environmental background and non-target risk factors unchanged. Counterfactual processing may include removing the target trigger node or replacing the target trigger node with a control node with lower risk. For example, replacing "pedestrian suddenly crossing" with "pedestrian waiting on the roadside", replacing "vehicle in front suddenly decelerating" with "vehicle in front moving at a constant speed", and replacing "occlusion suddenly removed" with "occlusion continues to exist", thereby obtaining the counterfactual twin scenario.

[0140] Step S2067: Calculate accident risk indicators based on the node twin scenario and the counterfactual twin scenario respectively, and determine the causal contribution of each triggering node according to the change between each accident risk indicator.

[0141] It should be noted that accident risk indicators refer to indicators used to measure the level of accident risk in a scenario, which can be collision risk, minimum safe distance, collision time, braking demand, trajectory deviation, or takeover risk; causal contribution refers to the degree to which the triggering node affects the changes in accident risk indicators.

[0142] It is understandable that this embodiment transforms the role of triggering nodes into the difference in risk indicators between node twin scenarios and counterfactual twin scenarios, so that the selection of key triggering factors has a quantitative basis and avoids judging the importance of nodes solely based on the time of occurrence or human experience.

[0143] In practice, the test equipment runs node twin scenarios and counterfactual twin scenarios respectively, and collects the perception and recognition results, predicted trajectory, planned trajectory, control output and vehicle operating status of the autonomous driving system in the two scenarios. Then, it calculates accident risk indicators such as collision risk, minimum safe distance, collision time, braking demand, trajectory deviation degree and takeover risk. Subsequently, it compares the changes in accident risk indicators between the node twin scenario and the counterfactual twin scenario corresponding to the same trigger node, and determines the causal contribution of each trigger node based on the changes.

[0144] Step S2068: Select key triggering nodes from the triggering node sequence based on the causal contribution, and construct an accident triggering chain based on the key triggering nodes.

[0145] It should be noted that a key triggering node refers to a triggering node whose causal contribution meets the preset causal contribution conditions and plays a major role in the formation of accident risk; the preset causal contribution conditions refer to the conditions set in advance for screening key triggering nodes.

[0146] It is understood that this embodiment filters key triggering nodes based on causal contribution and constructs an accident triggering chain based on the multidimensional relationship between key triggering nodes, so that the accident triggering chain can retain the core nodes that dominate the formation of accident risks and reduce the interference of irrelevant nodes on the generation of subsequent high-risk test scenarios.

[0147] In the specific implementation, the test equipment sorts each trigger node in the trigger node sequence according to the causal contribution and selects the trigger nodes that meet the preset causal contribution conditions as key trigger nodes; then, it organizes them in a chain according to the temporal sequence, spatial conflict, risk enhancement and system response relationship between the key trigger nodes to construct the accident trigger chain.

[0148] This embodiment extracts accident-related time-series data through a preset time window before the accident occurs, locates the interval of risk change by dividing the time sequence, forms a trigger node sequence by analyzing risk change factors, calculates the causal contribution by comparing node twin scenarios with counterfactual twin scenarios, and constructs an accident trigger chain by screening key trigger nodes based on the causal contribution, so that the accident trigger chain has the basis of time-series evolution, digital twin verification, and causal screening.

[0149] refer to Figure 6 , Figure 6This is a flowchart illustrating the fifth embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins of the present invention.

[0150] Based on the third embodiment described above, in this embodiment, step S207 may include: Step S2071: Determine the missing data type and the retained data type from the multi-source accident data based on the missing pattern characteristics.

[0151] It should be noted that missing data types refer to data categories in multi-source accident data that lack complete records, cannot be directly restored, or have insufficient reliability; reserved data types refer to data categories in multi-source accident data that can still be used as the basis for constructing accident scene fragments.

[0152] It is understandable that this embodiment uses missing pattern features to distinguish between missing content and available content, so that incomplete accident data can be first split into "non-recoverable parts" and "parts that can participate in fragment construction", avoiding the abandonment of the entire set of accident data due to the lack of local data.

[0153] In the specific implementation, the test equipment reads the missing data location, missing source, missing category and reconstruction credibility recorded in the missing mode feature, and marks the data categories that cannot continuously restore the accident process or cannot support independent scene instantiation as missing data types, and marks the data categories that still have a credible source, available time stamp, available spatial location or available object status as reserved data types.

[0154] For example, in cases where accident video is missing but accident reports, scene photos, and vehicle braking records are still available, the accident video trajectory can be marked as a missing data type, while road structure, occlusion relationships, and braking response can be marked as a retained data type.

[0155] Step S2072: Extract the accident retention feature set from the multi-source accident data based on the retained data type.

[0156] It should be noted that the accident-preserved feature set refers to the set of risk-related features extracted from incomplete accident data that can still be used to construct accident scene fragments. These features include road structure data, traffic object data, object behavior data, environmental state data, occlusion relationship data, conflict relationship data, and autonomous driving response data.

[0157] In practical implementation, the testing equipment filters multi-source accident data according to the type of data to be retained, and extracts usable accident features from available accident reports, scene photos, vehicle operation records, roadside perception records, and environmental records. For example, lane morphology, intersection structure, and obstruction locations are extracted from scene photos; traffic object categories and conflict descriptions are extracted from accident reports; braking response, steering response, and speed changes are extracted from vehicle operation records; and conditions such as rain / fog, nighttime, backlighting, or slippery road surfaces are extracted from environmental records. All these features are then grouped into a set of retained accident features.

[0158] Step S2073: Input the accident-preserved feature set and the missing data type into the fragment semantic coding network to perform fragment semantic coding and obtain fragment semantic features.

[0159] It should be noted that the fragment semantic coding network refers to a neural network used to jointly encode the set of retained accident features and the missing data type, and output the semantic representation of accident fragments; fragment semantic features refer to features used to characterize the type, meaning and correlation of local risk elements of an accident; and mask constraint vector refers to a vector used to express the constraint relationship of the missing data type on the fragment semantic coding process.

[0160] It is understood that this embodiment will retain both accident features and missing types as inputs to the fragment semantic coding network, so that the fragment semantic features not only reflect existing accident information, but also reflect the constraints of missing information on risk expression, avoiding the ambiguity of fragment risk meaning caused by mechanical classification based solely on data source.

[0161] In the specific implementation, the test equipment converts the accident-preserved feature set into feature vectors and the missing data types into mask constraint vectors. Then, the feature vectors and mask constraint vectors are input into the fragment semantic coding network. The fragment semantic coding network performs semantic encoding on features such as road structure, traffic objects, object behavior, environmental state, occlusion relationship, conflict relationship and autonomous driving response based on the contextual relationship between available accident features and the impact of missing data types on the completeness of risk expression, to obtain fragment semantic features.

[0162] In some embodiments, the testing equipment first extracts training samples from historical accident data, simulated accident data, and manually labeled accident samples. Each training sample includes a set of retained accident features, missing data types, fragment type labels, risk effect labels, and fragment association labels. Among them, fragment type labels can be used to mark road structure fragments, traffic object fragments, object behavior fragments, environmental state fragments, occlusion relationship fragments, conflict relationship fragments, or autonomous driving response fragments. Risk effect labels can be used to mark risk sources, risk triggers, risk amplification, or avoidance failures. Fragment association labels can be used to mark whether there are spatial constraint relationships, temporal constraint relationships, occlusion visibility relationships, or conflict interaction relationships between different fragments.

[0163] The test equipment constructs a fragment semantic coding network, which includes a feature embedding layer, a missing type embedding layer, a semantic fusion layer, a context coding layer, and a semantic output layer. The feature embedding layer is used to convert the accident retained feature set into an accident feature vector. The missing type embedding layer is used to convert the missing data type into a missing constraint vector. The semantic fusion layer is used to fuse the accident feature vector and the missing constraint vector. The context coding layer is used to encode the contextual relationship between different accident features. The semantic output layer is used to output the fragment semantic features.

[0164] During training, the testing equipment inputs training samples into the fragment semantic encoding network and jointly trains the network parameters based on fragment type prediction, risk effect prediction, fragment association prediction, and missing feature reconstruction tasks. Specifically, the fragment type prediction task constrains the network to identify the fragment category to which accident-preserved features belong; the risk effect prediction task constrains the network to identify the role of accident-preserved features in the accident risk formation process; the fragment association prediction task constrains the network to identify the combinable relationships between different fragments; and the missing feature reconstruction task constrains the network to learn the impact of missing information on fragment semantics based on available accident features and missing data types. If complete manual labels are lacking in historical accident samples, weak labels can be generated based on the accident cause description, traffic object trajectory, collision location, and accident liability determination information in the accident report. These weak labels are then used to pre-train the fragment semantic encoding network, which is then fine-tuned using manually verified samples. Through this training method, the fragment semantic encoding network can output fragment semantic features that simultaneously contain fragment category, risk effect, and association constraint information after inputting the accident-preserved feature set and missing data types.

[0165] Step S2074: Based on the fragment semantic features, the accident retained feature set is divided into fragment types to obtain multi-dimensional fragments.

[0166] It should be noted that multi-dimensional fragments refer to a set of local risk elements divided according to dimensions such as road, object, behavior, environment, occlusion, conflict, and system response in an accident scenario. These include road structure fragments, traffic object fragments, object behavior fragments, environmental state fragments, occlusion relationship fragments, conflict relationship fragments, and autonomous driving response fragments.

[0167] It is understandable that this embodiment uses fragment semantic features to divide the accident retained feature set into risk semantic dimensions, so that each type of local risk element can be independently completed, invoked and combined, providing a structured node foundation for subsequent fragment association mapping.

[0168] In practical implementation, the testing equipment classifies the accident retention feature set into fragment types based on the semantic similarity, risk type label, and scene functional attributes of different accident retention features in the fragment semantic features. For example, lane boundaries, intersection shapes, and zebra crossing locations are classified as road structure fragments; vehicle, pedestrian, and non-motorized vehicle categories are classified as traffic object fragments; deceleration, crossing, lane changing, and sudden stopping behaviors are classified as object behavior fragments; and rain, fog, night, backlight, and slippery road surfaces are classified as environmental state fragments, thus obtaining multi-dimensional fragments.

[0169] Step S2075: Based on the missing pattern features, determine the completable association relationships between each fragment in the multi-dimensional fragment to obtain a fragment association graph.

[0170] It should be noted that a completable association refers to an association between different accident fragments that can provide constraints for completing missing information; a fragment association graph refers to graph structure data used to express completable associations between multi-dimensional fragments.

[0171] It is understandable that this embodiment uses missing pattern features as the basis for mapping, so that the connection relationship between fragments is formed around the need to complete the missing information, rather than just based on the similarity of fragment categories, which can enhance the matching between fragment combinations and the missing state of accident data.

[0172] In the specific implementation, the test equipment identifies fragment combinations in multi-dimensional fragments that can provide constraints for missing time, missing space, missing trajectory, or missing occlusion relationship based on missing pattern features. It also establishes graph structure connections based on the spatial constraint relationship between road structure fragments and object behavior fragments, the interaction relationship between traffic object fragments and conflict relationship fragments, the visibility influence relationship between occlusion relationship fragments and environmental state fragments, and the response correspondence relationship between autonomous driving response fragments and object behavior fragments, to obtain a fragment association graph.

[0173] Step S2076: Input the fragment association graph into the graph attention fusion network and assign association weights to obtain fragment association weights.

[0174] It should be noted that graph attention fusion network refers to a neural network used to learn attention weights for nodes and edges in a graph structure and fuse related information; fragment association weight refers to weights used to characterize the association strength and combination priority between fragments of different accidents.

[0175] It is understood that this embodiment assigns weights to different fragment relationships through a graph attention fusion network, so that fragment relationships that contribute more to missing completion and risk representation are given priority in subsequent combinations, thereby reducing the interference of weakly related fragments or background fragments on the construction of accident scene fragments.

[0176] In the specific implementation, the test device inputs the fragment nodes, node attributes and associated edges in the fragment association graph into the graph attention fusion network. The graph attention fusion network assigns attention weights according to the degree of contribution of different fragments to the completion of missing information, the preservation of risk semantics and the expression of conflict relationships, and outputs the fragment association weights between fragment nodes. For example, when road structure fragments and object behavior fragments can jointly constrain the pedestrian crossing trajectory, the two can obtain a higher fragment association weight.

[0177] Step S2077: Combine the multi-dimensional fragments based on the fragment association weights to obtain accident scene fragments.

[0178] It is understandable that this embodiment combines multi-dimensional fragments based on fragment association weights rather than random splicing, so that the generated accident scene fragments simultaneously have data source evidence, missing information completion evidence, and risk association evidence.

[0179] In the specific implementation, the testing equipment filters multi-dimensional fragments that meet the combination requirements based on the fragment association weight, and combines the multi-dimensional fragments according to road constraints, object interaction relationships, behavior time relationships, environmental impact relationships, occlusion visibility relationships, conflict risk relationships, and autonomous driving response relationships to obtain accident scene fragments; for example, multi-dimensional fragments such as "intersection zebra crossing", "bus occlusion", "pedestrian crossing", "low light at night" and "main vehicle braking response delay" can be combined into occlusion crossing accident scene fragments.

[0180] It should be noted that the combination requirements may be that the fragment association weight is greater than the preset association weight threshold, there is a complete association relationship between fragments, the fragments meet the temporal compatibility condition, or the fragments can form at least one complete risk expression among roads, objects, behaviors, environments and conflict relationships after combination.

[0181] This embodiment distinguishes between missing data types and retained data types by using missing pattern features, extracts a set of retained accident features by using retained data types, generates fragment semantic features by using a fragment semantic coding network, divides fragments into multi-dimensional fragments by using fragment semantic features, establishes complete association relationships by using missing pattern features, and determines fragment association weights by using a graph attention fusion network. This enables incomplete accident data to be transformed from "incomplete records" into accident scene fragments with semantic categories, complete relationships, and combination priorities.

[0182] refer to Figure 7 , Figure 7 This is a flowchart illustrating the sixth embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins of the present invention.

[0183] Based on the above embodiments, in this embodiment, step S30 may include: Step S301: When the accident risk expression object is an accident triggering chain, the key triggering nodes in the accident triggering chain are mapped to the three-dimensional continuous traffic environment to obtain the first basic accident scenario.

[0184] It should be noted that the first accident basic scenario refers to the basic simulation scenario formed by mapping the accident trigger chain to a three-dimensional continuous traffic environment, which is used to carry the risk evolution process corresponding to the complete accident data.

[0185] In practical implementation, when the accident risk expression object is an accident trigger chain, the test equipment reads the occurrence time, occurrence location, participating objects, risk type, and sequential relationship between key trigger nodes in the accident trigger chain, and maps the above information to the road area, traffic object model, environmental state, occlusion relationship, and autonomous driving response state in the three-dimensional continuous traffic environment. For example, for the accident trigger chain of "bus occlusion - pedestrian appearance - driver vehicle braking delay", the bus position, pedestrian appearance position, driver vehicle driving trajectory, occlusion area, and braking response process can be configured into the corresponding intersection scene to form the basic scenario of the first accident.

[0186] Step S302: Generate a risk-equivalent disturbance scenario based on the first accident basic scenario and the key triggering node to obtain the first candidate high-risk scenario.

[0187] It should be noted that the risk equivalent perturbation scenario refers to the simulation scenario formed by perturbing the scenario parameters while keeping the accident risk transmission mechanism the same or similar; the first candidate high-risk scenario refers to the candidate scenario generated by the accident trigger chain corresponding to the complete accident data and can enter the subsequent high-risk screening process.

[0188] It is understandable that this embodiment perturbs parameters around key triggering nodes, so that the scene changes focus on the core factors that drive the increase in accident risk, and by maintaining the risk triggering order and spatial conflict relationship, the generated scene is different from the original accident reproduction, while retaining the original accident risk mechanism, thereby improving the scene expansion capability of complete accident data.

[0189] In some embodiments, the testing equipment determines the adjustable scenario parameters corresponding to the key triggering nodes based on the first accident basic scenario, and while keeping the risk triggering order, spatial conflict relationship and main participating objects unchanged, it perturbs the target appearance time, object speed, object spacing, occlusion release time, braking delay, road surface adhesion state or environmental visibility to generate multiple risk equivalent perturbation scenarios; it performs basic executability checks on the risk equivalent perturbation scenarios, and identifies the scenarios that meet the vehicle motion constraints, road traffic constraints and time continuity constraints as the first candidate high-risk scenarios.

[0190] Step S303: When the accident risk expression object is an accident scene fragment, the accident scene fragment is filled with missing information and risk consistency splicing is performed based on the three-dimensional continuous traffic environment to obtain a second basic accident scene. The second basic accident scene is screened based on the preset high-risk scene judgment conditions to obtain a second candidate high-risk scene.

[0191] It should be noted that missing information completion refers to supplementing the missing temporal, spatial, trajectory, occlusion, or environmental information in accident scene fragments based on roads, objects, rules, and visibility constraints in a three-dimensional continuous traffic environment. Risk consistency stitching refers to combining multiple accident scene fragments into a continuous accident scene according to the logical relationship between risk formation, risk triggering, risk amplification, and risk avoidance response.

[0192] It should be noted that the second basic accident scenario refers to the basic simulation scenario formed by splicing together fragments of the complete accident scenario through risk consistency, which is used to carry the continuous risk evolution process corresponding to incomplete accident data; the preset high-risk scenario judgment conditions may be that the minimum safe distance is lower than the preset safe distance threshold, the collision time is lower than the preset collision time threshold, the available braking distance is less than the preset braking distance threshold, the traffic object conflict area overlaps with the main vehicle trajectory, or the risk level reaches the preset risk level; the second candidate high-risk scenario refers to the candidate scenario generated from accident scenario fragments that can enter the subsequent high-risk screening process.

[0193] It is understandable that this embodiment uses a three-dimensional continuous traffic environment to constrain and complete accident scene fragments, so that local risk elements in incomplete accident data can form a basic scene that is temporally continuous, spatially reasonable, and has executable motion; at the same time, by risk consistency splicing, the combination of fragments is restricted to conflict relationships and risk action relationships, avoiding scene distortion caused by random combination of fragments.

[0194] In its implementation, when the accident risk representation object is an accident scene fragment, the testing equipment first completes the missing time information, spatial location, movement trajectory, occlusion relationship, or environmental state in the accident scene fragment based on the road geometry constraints, traffic object motion constraints, traffic rule constraints, and sensor visibility constraints in the three-dimensional continuous traffic environment. Then, it splices the completed accident scene fragment according to the conflict geometry relationship, temporal compatibility relationship, and risk effect complementarity relationship to obtain the second accident basic scene. Finally, it judges the second accident basic scene according to the preset high-risk scene judgment conditions and determines the scene that meets the judgment conditions as the second candidate high-risk scene.

[0195] Step S304: Based on preset high-risk screening conditions, the first candidate high-risk scenario or the second candidate high-risk scenario is screened to obtain high-risk test scenarios.

[0196] It should be noted that the preset high-risk screening conditions can be that the candidate scenario meets at least one of the following: vehicle dynamics executable conditions, road traffic rule constraints, risk index threshold, scenario integrity threshold, module injection interface requirements, and scenario repeatability threshold.

[0197] Understandably, this implementation incorporates candidate scenarios from accident triggering chain paths and accident scenario fragment paths into a unified screening process, ensuring that the final high-risk test scenarios simultaneously meet the requirements for risk intensity, simulation executability, and functional module injection, thus avoiding situations where generated scenarios, although containing risk factors, cannot run stably or be used for tiered testing.

[0198] In practice, the testing equipment uniformly screens the first or second candidate high-risk scenarios and sorts and filters the candidate scenarios based on scenario executability, risk intensity, scenario completeness, module injection adaptability, and scenario repetition.

[0199] For example, candidate scenarios that are discontinuous in the movement of traffic objects, conflict with road traffic rules, have risk indicators lower than the screening requirements, or are highly repetitive with existing scenarios can be eliminated. The remaining candidate scenarios can be converted into high-risk test scenarios that include scenario execution scripts, initial states of traffic objects, environmental states, trigger event configurations, and evaluation indicator configurations.

[0200] This embodiment maps key triggering nodes to the first basic accident scenario in the accident triggering chain path and generates risk-equivalent perturbation scenarios around the key triggering nodes, enabling complete accident data to be expanded into candidate high-risk scenarios with the same risk mechanism; by completing missing information and splicing risk consistency in the fragmented accident scenario path, incomplete accident data can be transformed into candidate high-risk scenarios that are spatiotemporally reasonable and have continuous risk relationships; by pre-setting high-risk screening conditions, the two types of candidate scenarios are uniformly screened, so that the final high-risk test scenario simultaneously possesses the accident risk source, digital twin executability, and module injection adaptability.

[0201] refer to Figure 8 , Figure 8 This is a flowchart illustrating the seventh embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins of the present invention.

[0202] Based on the above six embodiments, in this embodiment, step S302 may include: Step S3021: Extract key trigger node features from the key trigger node.

[0203] It should be noted that key trigger node characteristics refer to structured features used to describe the time, location, participating objects, and risk category of a key trigger node, including the time of occurrence, location, participating objects, and risk type of the node.

[0204] Among them, the node occurrence time refers to the time and location of the key triggering node during the accident evolution process; the node occurrence location refers to the road area or spatial coordinates corresponding to the key triggering node in the three-dimensional continuous traffic environment; the node participants refer to the traffic objects involved in the formation of the key triggering node, such as the main vehicle, the vehicle in front, pedestrians, non-motorized vehicles, or obstructing vehicles; the node risk type can be obstruction crossing risk, rear-end collision risk, lane change conflict risk, oncoming conflict risk, or insufficient braking risk.

[0205] In practical implementation, the testing equipment reads the key trigger nodes selected in the accident trigger chain and parses the node occurrence time, node location, participating objects, and node risk type from each key trigger node. For example, for the key trigger node "pedestrian crossing after obstruction is removed", the time of obstruction removal, the location of the pedestrian, the pedestrian and the main vehicle as participating objects, and the risk of obstruction crossing as the node risk type can be parsed. For the key trigger node "vehicle in front decelerates suddenly", the time when the vehicle in front begins to decelerate, the lane position of the vehicle in front, the vehicle in front and the main vehicle as participating objects, and the risk of rear-end collision as the node risk type can be parsed.

[0206] Step S3022: Determine the perturbed scenario parameters based on the key trigger node characteristics and the first accident basic scenario.

[0207] It should be noted that the perturbed scenario parameters refer to the scenario parameters that can be adjusted without damaging the basic road structure and risk mechanism of the first accident scenario. These parameters include the speed of traffic participants, the distance between traffic participants, the time of target appearance, the position of obstructions, road surface adhesion parameters, and braking delay parameters.

[0208] Among them, the distance between traffic participants refers to the longitudinal distance, lateral distance or conflict zone distance between the main vehicle and other traffic participants; the road surface adhesion parameter refers to the parameter used to characterize the adhesion ability between the tire and the road surface; the braking delay parameter refers to the delay parameter between the vehicle generating a braking request and forming a braking response.

[0209] It is understood that this embodiment determines the perturbed scenario parameters based on the characteristics of key triggering nodes and the basic scenario of the first accident, so that the perturbed parameters correspond to the accident risk type, avoids perturbing irrelevant environmental elements, and improves the correlation between the perturbed scenario and the original accident risk mechanism.

[0210] In practical implementation, the testing equipment matches the key trigger node features with the road structure, traffic object trajectory, obstruction area, environmental state, and vehicle response state in the basic scenario of the first accident, and determines the perturbed scenario parameters according to the node risk type. For example, in the risk of obstruction crossing, the time of target appearance, the position of the obstruction, the speed of the main vehicle, and the speed of the pedestrian crossing can be determined as perturbed scenario parameters; in the risk of rear-end collision, the deceleration of the vehicle in front, the following distance of the main vehicle, the braking delay of the main vehicle, and the road surface adhesion parameters can be determined as perturbed scenario parameters; in the risk of lane change conflict, the lateral speed of the adjacent vehicle, the distance between the main vehicles, and the time of target appearance can be determined as perturbed scenario parameters.

[0211] Step S3023: Determine the original risk transmission characteristics based on the first accident basic scenario, and construct a perturbation constraint vector based on the original risk transmission characteristics and the perturbable scenario parameters.

[0212] It should be noted that the original risk transmission characteristics refer to the transmission process characteristics of risk from formation, triggering, enhancement to response failure in the basic scenario of the first accident; the disturbance constraint vector refers to the vectorized characteristics used to constrain the generation process of disturbance parameters, which can characterize the disturbance range, risk transmission maintenance requirements, spatial conflict maintenance requirements, and motion feasibility requirements.

[0213] It is understood that in this embodiment, the risk transmission process in the basic scenario of the first accident is coupled and encoded with the perturbed scenario parameters into a perturbation constraint vector, so that the generation of subsequent perturbation parameters is constrained by the original accident risk path, thus avoiding the generation of scenarios that only have parameter changes but no longer maintain the accident risk transmission relationship.

[0214] In the specific implementation, the test equipment runs the first accident basic scenario and extracts the risk triggering sequence, risk propagation path, traffic object conflict relationship, minimum safe distance change, collision time change, braking demand change and autonomous driving response change from the operation process to form the original risk transmission characteristics. The original risk transmission characteristics are associated with the perturbed scenario parameters and encoded as perturbation constraint vectors, which parameters can be perturbed, the perturbation range, the risk transmission sequence maintenance requirement, the spatial conflict relationship maintenance requirement, and the vehicle motion feasibility requirement.

[0215] Step S3024: Input the perturbation constraint vector into the constraint-based generative adversarial network to generate perturbation parameters and obtain candidate perturbation parameter combinations.

[0216] It should be noted that constrained generative adversarial networks refer to generative adversarial networks that introduce scene constraint evaluation mechanisms in addition to the generator and discriminator; candidate perturbation parameter combinations refer to the set of perturbation parameters generated by constrained generative adversarial networks that have not yet completed risk transmission similarity screening.

[0217] It is understood that this embodiment generates candidate perturbation parameter combinations through a constrained generative adversarial network, so that the perturbation parameters no longer rely on manual enumeration or random search, but are generated under the joint constraints of risk similarity, vehicle motion constraints and road traffic constraints, thereby improving the matching degree between candidate perturbation parameters and accident risk scenarios.

[0218] In a specific implementation, the test equipment can pre-build a constrained generative adversarial network, which includes a parameter generator, a scene discriminator, and a constraint evaluator. The parameter generator is used to generate candidate disturbance parameters based on the disturbance constraint vector. The scene discriminator is used to determine whether the scene after loading the candidate disturbance parameters still belongs to the same type of risk scenario as the accident. The constraint evaluator is used to determine whether the candidate disturbance parameters meet the vehicle motion constraints, road traffic constraints, and parameter value constraints.

[0219] During training, training samples can be extracted from historical accident variant scenarios, simulated perturbation scenarios, and manually labeled high-risk scenarios. The perturbation constraint vector is used as input, and combinations of perturbation parameters that satisfy risk similarity and executability are used as positive samples, while combinations of perturbation parameters that violate risk mechanisms or physical constraints are used as negative samples. This process jointly trains the parameter generator, scenario discriminator, and constraint evaluator. After training, the current perturbation constraint vector is input into a constraint-based generative adversarial network, which outputs multiple candidate perturbation parameter combinations.

[0220] Step S3025: Filter the candidate perturbation parameter combinations based on a preset risk transmission similarity threshold to obtain risk equivalent perturbation parameter combinations.

[0221] It should be noted that the preset risk transmission similarity threshold can be a pre-set lower limit of comprehensive similarity, such as the risk triggering sequence similarity reaching a preset sequence similarity threshold, the spatial conflict relationship similarity reaching a preset spatial similarity threshold, the key object interaction similarity reaching a preset interaction similarity threshold, or the comprehensive similarity reaching a preset comprehensive similarity threshold; the risk equivalent disturbance parameter combination refers to the set of disturbance parameters that can make the disturbance scenario maintain the same or similar risk transmission mechanism as the basic scenario of the first accident.

[0222] It is understood that this embodiment filters candidate perturbation parameter combinations by setting a risk transmission similarity threshold, so that the retained perturbation parameters not only generate high-risk states, but also maintain the risk triggering order, spatial conflict relationship and key object interaction relationship, thereby reducing the problem of the scene deviating from the original accident mechanism after perturbation.

[0223] In the specific implementation, the test equipment loads each set of candidate disturbance parameter combinations into the first accident basic scenario to generate corresponding candidate disturbance scenarios, and calculates the similarity of risk triggering order, spatial conflict relationship, key object interaction, and risk indicator change trend between the candidate disturbance scenarios and the first accident basic scenario respectively; then, the comprehensive similarity is compared with the preset risk transmission similarity threshold, and the candidate disturbance parameter combinations that meet the threshold requirements are determined as risk equivalent disturbance parameter combinations.

[0224] Step S3026: Load the risk equivalent disturbance parameter combination into the first accident basic scenario to obtain the risk equivalent disturbance scenario.

[0225] It should be noted that the risk equivalent disturbance scenario refers to the simulation scenario formed after loading the combination of risk equivalent disturbance parameters, and which maintains the same or similar risk transmission mechanism as the basic scenario of the first accident.

[0226] It is understandable that in this embodiment, the selected disturbance parameters are loaded into the first accident scenario, so that the original accident scenario is expanded into multiple scenarios with the same risk mechanism but different parameter combinations, providing boundary-based and diversified test inputs for the autonomous driving system.

[0227] In practice, the testing equipment modifies the corresponding scenario parameters in the basic scenario of the first accident based on the combination of risk equivalent disturbance parameters, and regenerates the initial state of the traffic object, the trajectory of the traffic object, the change of the occlusion area, the environmental state configuration, and the vehicle response configuration. For example, while maintaining the risk sequence of "occlusion removal - pedestrian crossing - insufficient braking of the main vehicle", the pedestrian appearance time, the speed of the main vehicle, the position of the occluding vehicle, and the braking delay parameters can be adjusted to form a risk equivalent disturbance scenario.

[0228] Step S3027: Perform executability verification on the risk equivalent disturbance scenario based on the three-dimensional continuous traffic environment to obtain the first candidate high-risk scenario.

[0229] It should be noted that executability verification refers to the verification process used to determine whether a risk equivalent disturbance scenario can operate stably in a three-dimensional continuous traffic environment. It can include vehicle dynamics verification, road boundary verification, traffic rule verification, time continuity verification, and sensor observation generation verification.

[0230] It is understandable that this embodiment puts the risk-equivalent disturbance scenario back into the three-dimensional continuous traffic environment for executability verification, so that the final output first candidate high-risk scenario not only retains the accident risk transmission mechanism, but also meets the requirements of vehicle motion, road rules and sensor observation generation, avoiding the generation of scenarios that cannot run or cannot be injected into the test.

[0231] In the specific implementation, the test equipment runs a risk equivalent disturbance scenario in a three-dimensional continuous traffic environment and checks whether the movement of traffic objects is continuous, whether the vehicle dynamics are executable, whether the traffic objects violate the road boundaries, whether there are conflicts with traffic rules, whether sensor observations can be generated, and whether key triggering nodes occur as configured. If the risk equivalent disturbance scenario meets the executability verification requirements, the risk equivalent disturbance scenario is identified as the first candidate high-risk scenario.

[0232] This embodiment locates the disturbance object by identifying key trigger node features, limits the disturbance range by perturbed scenario parameters, constructs a disturbance constraint vector by using original risk transmission features and perturbed scenario parameters, generates candidate disturbance parameter combinations by using a constraint-based generative adversarial network, filters risk-equivalent disturbance parameter combinations by using a preset risk transmission similarity threshold, and performs executability verification through a three-dimensional continuous traffic environment. This enables the first basic accident scenario corresponding to complete accident data to be expanded into multiple first candidate high-risk scenarios with consistent risk mechanisms, different parameter forms, and operability.

[0233] refer to Figure 9 , Figure 9 This is a flowchart illustrating the eighth embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins of the present invention.

[0234] Based on the above six embodiments, in this embodiment, step S303 may include: Step S3031: Construct a spatiotemporal mask map of the fragments based on the key fragments in the accident scene fragments.

[0235] It should be noted that critical fragments refer to fragments in an accident scenario that play a major role in the formation, triggering, amplification, or failure of an accident risk, including road structure fragments, traffic object fragments, object behavior fragments, environmental state fragments, occlusion relationship fragments, conflict relationship fragments, and autonomous driving response fragments.

[0236] A fragmented spatiotemporal mask map is a graph structure data that uses key fragments as graph nodes, and the temporal, spatial, interactive, and missing relationships between key fragments as graph edges, and uses mask markers to represent missing information; mask markers are location markers used to identify missing time, missing space, missing trajectory, missing occlusion relationship, or missing response status.

[0237] It is understood that this embodiment transforms incomplete accident scene fragments into a graph structure with spatiotemporal relationships and missing location markers, so that the subsequent completion process can clearly identify "which risk fragments need to be retained" and "which missing information needs to be completed", avoiding unconstrained completion of accident fragments.

[0238] In practical implementation, the testing equipment filters key fragments from accident scene fragments that can characterize the source, trigger, amplification, or avoidance response of the accident risk, and converts these key fragments into graph nodes. Then, based on the temporal sequence, spatial adjacency, object interaction, and observation gap relationships among the key fragments, graph edges are constructed, and mask markers are configured for missing time, missing spatial location, missing motion trajectory, missing occlusion relationship, or missing response state to obtain a spatiotemporal mask map of the fragments. For example, for accident fragments such as "nighttime intersection, bus obstruction, pedestrian crossing, and insufficient braking of the main vehicle," the nighttime environment, intersection road, bus obstruction, pedestrian behavior, and main vehicle response can be used as graph nodes, and mask markers can be configured for the missing complete pedestrian trajectory.

[0239] Step S3032: Construct a complete condition vector based on the multidimensional constraints in the three-dimensional continuous traffic environment.

[0240] It should be noted that the completion condition vector refers to the condition input formed after vectorizing and encoding multidimensional constraints. Multidimensional constraints refer to the set of constraints extracted from the three-dimensional continuous traffic environment to limit the scope and form of missing information completion, including road geometry constraints, vehicle kinematics constraints, traffic rule constraints, traffic participant behavior constraints, and sensor visibility constraints.

[0241] Among them, road geometric constraints refer to spatial constraints formed by lane boundaries, intersection shapes, road width, zebra crossing positions, or curb positions; vehicle kinematic constraints refer to kinematic constraints formed by vehicle speed, acceleration, turning radius, braking distance, or lateral movement capability; traffic participant behavior constraints refer to the range of behaviors that vehicles, pedestrians, or non-motorized vehicles may take in traffic scenarios; and sensor visibility constraints refer to observation constraints formed by sensor field of view, detection distance, occlusion relationships, and target size.

[0242] It is understood that this embodiment encodes the road, vehicle, rule, behavior and sensor constraints in the three-dimensional continuous traffic environment into a completion condition vector, so that the completion of missing information is subject to the constraints of the real road structure and the simulation executable conditions, avoiding problems such as trajectory crossing road boundaries, behavior violating traffic rules or unreasonable unobservable relationships of the target in the completion results.

[0243] In practical implementation, the testing equipment reads the road area, lane boundary, intersection structure, speed limit rules, traffic signal rules, vehicle movement boundary, pedestrian or non-motorized vehicle movement pattern, obstruction position and sensor field of view corresponding to the accident scene fragments from the three-dimensional continuous traffic environment, and converts these constraints into vectorized conditions. For example, in the intersection obstruction crossing scenario, the zebra crossing position, the vehicle's passable direction, the area where pedestrians may cross, the bus obstruction range and the visible area of ​​the main vehicle sensor can be encoded into a completion condition vector.

[0244] Step S3033: Input the fragment spatiotemporal mask image and the completion condition vector into the pre-constructed spatiotemporal diffusion completion network to complete the missing information and obtain the completed accident scene fragments.

[0245] It should be noted that the spatiotemporal diffusion completion network refers to a neural network that gradually recovers missing spatiotemporal information based on a diffusion generation mechanism; and accident scene fragment completion refers to the fragmented data obtained after supplementing the missing time, space, trajectory, occlusion or response information in accident scene fragments.

[0246] It is understandable that in this embodiment, the missing positions in the spatiotemporal mask image of the fragments and the environmental constraints in the completion condition vector are input into the spatiotemporal diffusion completion network. This allows the missing information completion to utilize the spatiotemporal relationship between the accident fragments while being constrained by road geometry, kinematics, traffic rules, and visibility constraints, thereby obtaining complete accident scene fragments that are more suitable for subsequent scene stitching and simulation execution.

[0247] In the specific implementation, the test equipment inputs the spatiotemporal mask image of the fragments and the completion condition vector into the pre-constructed spatiotemporal diffusion completion network. The spatiotemporal diffusion completion network performs step-by-step denoising and completion of the missing time, missing spatial location, missing motion trajectory, missing occlusion relationship or missing response state corresponding to the mask mark, and outputs the completed accident scene fragments.

[0248] In one embodiment, the spatiotemporal diffusion completion network can be trained using historical complete accident samples, simulated accident samples, and manually verified accident samples. During training, a training fragment spatiotemporal map is first constructed from the complete accident samples. Then, some time, space, trajectory, occlusion, or response information is randomly occluded to form a training fragment spatiotemporal mask map. The occluded information in the original complete accident samples is used as the supervision target. The network structure may include a graph coding layer, a conditional fusion layer, a diffusion denoising layer, and a completion output layer. The graph coding layer is used to encode the fragment spatiotemporal mask map, the conditional fusion layer is used to fuse and complete the conditional vector, the diffusion denoising layer is used to gradually recover the occluded information, and the completion output layer is used to output the completed time, space, trajectory, occlusion, and response states.

[0249] Step S3034: Determine the fragment splicing constraints based on the completed accident scene fragments.

[0250] It should be noted that fragment splicing constraints refer to the constraints used to limit whether and how fragments in a completed accident scene can be combined, including conflict geometric relationships, temporal compatibility relationships, and complementary risk effects.

[0251] Among them, conflict geometry relationship refers to the conflict relationship formed in spatial geometry between traffic participants, road areas and conflict areas; temporal compatibility relationship refers to the relationship in which different fragments can form a continuous accident process in terms of occurrence sequence and duration; risk role complementarity relationship refers to the relationship in which different fragments have complementary roles in risk formation, risk triggering, risk amplification or risk avoidance response.

[0252] It is understood that this embodiment determines the fragment splicing constraints through three dimensions: conflict geometry, temporal compatibility, and risk effect, so that the subsequent splicing is not a simple superposition of fragments, but a combination of conditions based on the correspondence between spatial conflict, temporal continuity, and risk mechanism.

[0253] In practical implementation, the testing equipment performs relationship analysis on the fragments of the completed accident scene to determine whether there are splicable conflict geometric relationships, temporal compatibility relationships, and complementary risk effects between different fragments. For example, "the obstructing vehicle is located in the obstructed area" and "the pedestrian enters the conflict area from the obstructed area" are identified as spatially compatible conflict geometric relationships, "obstruction continues - obstruction is lifted - target appears - main vehicle brakes" are identified as temporal compatibility relationships, and "low illumination at night, obstruction crossing, and delayed braking response" are identified as complementary relationships between risk formation, risk triggering, and insufficient avoidance response.

[0254] Step S3035: Based on the fragment splicing constraints, perform risk consistency splicing on the complete accident scene fragments to obtain the second accident basic scene.

[0255] In practical implementation, the testing equipment selects complete accident scene fragments that meet the splicing conditions based on fragment splicing constraints, and loads these fragments into a three-dimensional continuous traffic environment in the order of risk formation, risk triggering, risk amplification, and hazard avoidance response. During the loading process, the time axis, spatial coordinates, traffic object state, environmental state, and autonomous driving response state between fragments are aligned to form a continuous high-risk accident evolution process, thus obtaining the basic scenario for the second accident. For example, "low visibility in rainy nights," "large vehicle obstruction," "pedestrian crossing trajectory," and "insufficient braking response of the main vehicle" can be spliced ​​together to form a basic scenario for a second accident involving obstruction and crossing in rainy nights.

[0256] This embodiment constructs a spatiotemporal mask map of fragments through key fragments, enabling the structured expression of missing locations and spatiotemporal relationships in accident scene fragments; it constructs a completion condition vector through multi-dimensional constraints, subjecting missing completion to road geometry, vehicle motion, traffic rules, behavioral patterns, and sensor visibility limitations; it recovers missing information through a spatiotemporal diffusion completion network, transforming incomplete fragments into complete accident scene fragments; and it forms fragment splicing constraints through conflict geometric relationships, temporal compatibility relationships, and complementary relationships of risk effects, and performs risk consistency splicing based on fragment splicing constraints, enabling incomplete accident data to be transformed into a continuous, reasonable, and simulable second accident base scenario.

[0257] refer to Figure 10 , Figure 10 This is a flowchart illustrating the ninth embodiment of the simulation testing method for high-risk autonomous driving scenarios based on digital twins of the present invention.

[0258] Based on the above embodiments, in this embodiment, step S40 may include: Step S401: Analyze the scenario execution script, scenario parameter configuration, risk triggering node, and traffic participant status in the high-risk test scenario to obtain causal description data of the high-risk scenario.

[0259] In its implementation, the testing equipment reads the scenario execution script and scenario parameter configuration corresponding to the high-risk test scenario, and parses the scenario execution timeline, road area, initial state of traffic objects, object movement trajectory, environmental conditions, trigger event configuration, and risk trigger nodes. Simultaneously, it identifies the temporal, spatial, and conflict relationships between the main vehicle, target vehicle, pedestrians, non-motorized vehicles, obstructions, and static obstacles based on the states of traffic participants, and organizes the correlation between risk trigger nodes and the states of traffic participants into causal descriptive data for high-risk scenarios. For example, in an obstruction-crossing scenario, content such as "bus obstruction," "pedestrian entering the conflict area," "main vehicle perception delay," and "main vehicle insufficient braking response" can be organized into descriptive data with temporal order and causal orientation.

[0260] Step S402: Obtain the interface description information corresponding to the perception module, prediction module, planning module and control module in the autonomous driving system to be tested, and generate module input constraint data based on the interface description information.

[0261] In practical implementation, the test equipment reads the interface description information of each functional module in the autonomous driving system under test, and parses the data type, field format, timestamp rules, coordinate system rules, data frequency, message synchronization requirements, and input boundary requirements corresponding to each functional module. Then, based on the interface description information, it generates module input constraint data to limit the data structure and transmission conditions that the perception test data, prediction test data, planning test data, and control test data must meet during generation. For example, the perception module may require that the input image frames, point cloud frames, and radar target lists have a unified timestamp, and the planning module may require that the input dynamic obstacle data and drivable area data have the same coordinate reference.

[0262] Step S403: Construct a cross-module causal injection graph based on the high-risk scenario causal description data and the module input constraint data.

[0263] It should be noted that cross-module causal injection graph refers to graph structure data used to express the causal relationship between risk factors in high-risk scenarios and the input-output relationship between different functional modules of the autonomous driving system. It includes scenario risk nodes, module input nodes, module output nodes, and module dependency edges.

[0264] Scenario risk nodes refer to nodes that represent risk-triggered events, risk objects, or risk states in high-risk test scenarios; module input nodes refer to nodes that represent the input data of functional modules; module output nodes refer to nodes that represent the output results of functional modules; and module dependency edges refer to edges that represent data dependency relationships or causal transmission relationships between different functional modules.

[0265] It is understood that this embodiment organizes the scenario risk nodes and the input and output nodes of the functional modules into a cross-module causal injection graph, so that the same risk factor can establish an injection relationship along the perception, prediction, planning and control links, avoiding the problem of the risk causal relationship being broken due to the independent generation of test data of each module.

[0266] In the specific implementation, the test equipment uses the risk triggering nodes, traffic participant states, and risk transmission relationships in the causal description data of high-risk scenarios as the basis for scenario risk nodes. Based on the module input constraint data, module input nodes and module output nodes are established for the perception module, prediction module, planning module, and control module, respectively. Subsequently, module dependency edges are established according to the processing links in the autonomous driving system where perception results affect prediction results, prediction results affect planning results, and planning results affect control results. Scenario risk nodes are then associated with their corresponding module input nodes to form a cross-module causal injection graph.

[0267] Step S404: Based on the cross-module causal injection graph, the high-risk test scenario is sliced ​​to obtain a perception scenario slice, a prediction scenario slice, a planning scenario slice, and a control scenario slice.

[0268] It should be noted that the perception scene slice, prediction scene slice, planning scene slice, and control scene slice refer to the local scene data extracted from the high-risk test scenario according to the input requirements of each functional module.

[0269] In its implementation, the testing equipment performs module-level slicing of high-risk test scenarios based on the correlation between scenario risk nodes and module input nodes in the cross-module causal injection graph. For the perception module, scene content related to sensor observations, target appearance, occlusion relationships, and environmental conditions is extracted to form perception scene slices. For the prediction module, scene content related to the trajectory, behavioral intentions, and future motion uncertainties of traffic participants is extracted to form prediction scene slices. For the planning module, scene content related to road boundaries, dynamic obstacles, traffic rules, and drivable areas is extracted to form planning scene slices. For the control module, scene content related to target trajectory, target speed, braking response, and steering response is extracted to form control scene slices.

[0270] Step S405: Generate perception test data based on the perception scene slices, and generate prediction test data based on the prediction scene slices.

[0271] It should be noted that the perception test data includes image data, point cloud data, and radar target data, while the prediction test data includes traffic participant trajectory data, traffic participant behavioral intention data, and traffic participant movement probability data.

[0272] In the specific implementation, the test equipment renders and generates image data, point cloud data, and radar target data based on the road structure, appearance of traffic objects, occlusion status, ambient lighting, and sensor configuration in the perceived scene slice, and configures the timestamp, coordinate system, and sensor identifier according to the requirements of the perception module interface. At the same time, based on the historical trajectory, speed change, lateral movement trend, and interaction status of traffic participants in the predicted scene slice, it generates trajectory data, behavioral intent data, and motion probability data of traffic participants, and configures the object identifier, trajectory time range, and motion uncertainty information according to the requirements of the prediction module interface.

[0273] Step S406: Generate planning test data based on the planning scene slice, and generate control test data based on the control scene slice.

[0274] It should be noted that planning test data refers to the data used as input to the planning module to describe road boundaries, obstacles, rule constraints, and drivable space, including road boundary data, dynamic obstacle data, traffic rule constraint data, and drivable area data. Control test data refers to the data used as input to the control module to describe the vehicle's target motion and execution response, including target trajectory data, target speed data, braking response data, and steering response data.

[0275] Traffic rule constraint data refers to data used to indicate speed limits, yielding, traffic light compliance, lane change restrictions, or turning restrictions.

[0276] It is understood that this embodiment converts the content related to decision constraints in high-risk testing scenarios into planning test data, and the content related to vehicle execution into control test data, so that the planning module and the control module can be tested at the risk decision layer and the execution layer respectively, avoiding the problem that the vehicle-level results cannot distinguish between planning infeasibility and control execution inadequacy.

[0277] In practical implementation, the testing equipment generates road boundary data, dynamic obstacle data, traffic rule constraint data, and drivable area data based on the road geometry, traffic rules, dynamic obstacle positions, object predicted trajectories, and drivable area of ​​the main vehicle in the planned scenario slice, and configures a unified coordinate reference and time range according to the requirements of the planning module interface. At the same time, based on the planned trajectory, speed target, braking requirements, steering requirements, and actuator response conditions in the control scenario slice, it generates target trajectory data, target speed data, braking response data, and steering response data, and configures the control cycle, vehicle status, and execution parameters according to the requirements of the control module interface.

[0278] Step S407: Based on the cross-module causal injection graph, perform causal consistency verification on the perception test data, the prediction test data, the planning test data, and the control test data to obtain causal consistent hierarchical injection test data.

[0279] It should be noted that causal consistency verification refers to the process of checking whether test data from different functional modules maintain the same risk trigger node, the same traffic participant, the same timeline, and the same spatial reference; causal consistency layered injection test data refers to test data that, after causal consistency verification, can be input into different functional modules and maintain the same high-risk scenario risk causal relationship.

[0280] In the specific implementation, the test equipment verifies the temporal consistency, object consistency, spatial consistency, and risk node consistency among the perception test data, prediction test data, planning test data, and control test data based on the cross-module causal injection graph, thereby avoiding the problem of risk node misalignment or inconsistent object states after the module data is generated separately.

[0281] For example, the time of pedestrian appearance in the perception test data needs to correspond to the starting point of the pedestrian trajectory in the prediction test data; the location of dynamic obstacles in the planning test data needs to correspond to the trajectory of traffic participants in the prediction test data; and the target trajectory in the control test data needs to correspond to the drivable area in the planning test data. If there are time deviations, inconsistent object identifications, inconsistent coordinates, or missing risk nodes, alignment corrections can be performed based on the cross-module causal injection graph to obtain causally consistent hierarchical injection test data.

[0282] Step S408: According to the injection order corresponding to the module dependency edges, inject the test data of the causal consistent hierarchical injection into the perception module, the prediction module, the planning module and the control module respectively.

[0283] In the specific implementation, the test equipment determines the test data injection order based on the module dependency edges in the cross-module causal injection graph, and injects the perception test data, prediction test data, planning test data, and control test data into the corresponding functional modules according to the data dependency relationships between perception input, prediction input, planning input, and control input. During the injection process, serial injection or bypass injection can be selected according to the test purpose. For example, serial injection is used to test the complete functional link, while bypass injection is used to bypass the preceding module and test the prediction module, planning module, or control module separately.

[0284] This embodiment obtains causal description data of high-risk test scenarios by parsing them, generates input constraint data for the interface description information generation module, constructs a cross-module causal injection graph using the causal description data of high-risk scenarios and the module input constraint data, and performs scene slicing, test data generation, causal consistency verification, and dependency-ordered injection based on the cross-module causal injection graph. This enables the same high-risk test scenario to be tested in a layered manner using four types of module inputs: perception, prediction, planning, and control, while maintaining the consistency of risk triggering nodes, traffic object states, and module data dependencies.

[0285] Furthermore, this embodiment of the invention also proposes a computer-readable storage medium storing a high-risk autonomous driving scenario simulation test program. When the high-risk autonomous driving scenario simulation test program is executed by a processor, it implements the steps of the digital twin-based high-risk autonomous driving scenario simulation test method described above.

[0286] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0287] The aforementioned computer-readable storage medium may be included in a digital twin-based autonomous driving high-risk scenario simulation test device; or it may exist independently and not be assembled into a digital twin-based autonomous driving high-risk scenario simulation test device.

[0288] Furthermore, this invention also proposes a computer program product, including an autonomous driving high-risk scenario simulation test program, which, when executed by a processor, implements the steps of the autonomous driving high-risk scenario simulation test method based on digital twins as described above.

[0289] The specific implementation of the computer program product of the present invention is basically the same as the embodiments of the above-mentioned simulation test method for high-risk autonomous driving scenarios based on digital twins, and will not be repeated here.

[0290] Reference Figure 11 , Figure 11This is a structural block diagram of the first embodiment of the high-risk scenario simulation and testing system for autonomous driving based on digital twins of the present invention.

[0291] like Figure 11 As shown in the figure, the high-risk autonomous driving scenario simulation and testing system based on digital twin proposed in this embodiment of the invention includes: Traffic environment construction module 10 is used to acquire the topology of the real road network and multi-source accident data, and construct a three-dimensional continuous traffic environment in the digital twin space based on the topology and the multi-source accident data. The risk object construction module 20 is used to construct an accident risk expression object based on the completeness of the multi-source accident data. When the multi-source accident data meets the preset completeness conditions, the accident risk expression object includes an accident trigger chain. When the multi-source accident data does not meet the preset completeness conditions, the accident risk expression object includes accident scene fragments. The test scenario construction module 30 is used to generate a high-risk test scenario by performing a matching digital twin scenario instantiation strategy on the accident risk expression object based on the three-dimensional continuous traffic environment and the type of the accident risk expression object. The test data injection module 40 is used to convert the high-risk test scenarios into test data corresponding to different functional modules in the autonomous driving system to be tested, and inject the test data into the corresponding functional modules. The functional modules include a perception module, a prediction module, a planning module, and a control module. The test evaluation module 50 is used to output the test evaluation results of high-risk scenarios based on the response results of the autonomous driving system in the injected scenario.

[0292] This embodiment maps the real road network topology and multi-source accident data to a digital twin space, enabling a continuous three-dimensional traffic environment that simultaneously includes road connectivity, traffic participant status, accident environmental conditions, and accident occurrence process information. This avoids the lack of real accident constraints when generating scenarios solely based on static maps or manual scripts, thus providing subsequent high-risk test scenarios with a foundation in real road structures and accident risk sources. Accident trigger chains and accident scenario fragments are constructed based on the completeness of the multi-source accident data, allowing complete accident data to express the gradual accumulation of risk in the accident evolution process, while incomplete accident data can also participate in scenario generation through local risk elements, preventing the direct discarding of incomplete accident data. A matching digital twin scenario instantiation strategy is executed based on the type of accident risk expression object, enabling accident trigger chains to be transformed into high-risk scenarios with continuous risk transmission relationships, and accident scenario fragments to be transformed into combined scenarios with consistent risks. High-risk test scenarios are converted into test data corresponding to each functional module, allowing the same accident risk mechanism to be injected and verified at different functional levels, thereby achieving synergy between high-risk scenario generation, module-level testing, and anomaly source localization.

[0293] The digital twin-based high-risk autonomous driving scenario simulation testing system provided in this application, employing the digital twin-based high-risk autonomous driving scenario simulation testing method described in the above embodiments, can solve the technical problems of digital twin-based high-risk autonomous driving scenario simulation testing. Compared with the prior art, the beneficial effects of the digital twin-based high-risk autonomous driving scenario simulation testing system provided in this application are the same as those of the digital twin-based high-risk autonomous driving scenario simulation testing method provided in the above embodiments, and other technical features of the digital twin-based high-risk autonomous driving scenario simulation testing system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0294] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0295] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0296] In addition, for technical details not described in detail in this embodiment, please refer to the simulation test method for high-risk autonomous driving scenarios based on digital twins provided in any embodiment of the present invention, which will not be repeated here.

[0297] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0298] It should be noted that the user information (including but not limited to user device information, user personal information, user location information, user behavior information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0299] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0300] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0301] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A simulation testing method for high-risk scenarios of autonomous driving based on digital twins, characterized in that, The method includes: The topology and multi-source accident data of the real road network are obtained, and a three-dimensional continuous traffic environment is constructed in the digital twin space based on the topology and the multi-source accident data. An accident risk expression object is constructed based on the completeness of the multi-source accident data. When the multi-source accident data meets the preset completeness conditions, the accident risk expression object includes an accident trigger chain. When the multi-source accident data does not meet the preset completeness conditions, the accident risk expression object includes accident scene fragments. Based on the three-dimensional continuous traffic environment and the type of the accident risk expression object, a matching digital twin scenario instantiation strategy is executed on the accident risk expression object to generate a high-risk test scenario. The high-risk test scenarios are converted into test data corresponding to different functional modules in the autonomous driving system to be tested, and the test data is injected into the corresponding functional modules, including a perception module, a prediction module, a planning module, and a control module. Based on the response results of the autonomous driving system in the injected scenario, the test evaluation results of the high-risk scenario are output.

2. The simulation testing method for high-risk autonomous driving scenarios based on digital twins as described in claim 1, characterized in that, The process of acquiring the topology and multi-source accident data of a real road network, and constructing a three-dimensional continuous traffic environment in a digital twin space based on the topology and the multi-source accident data, includes: The topology of the real road network and multi-source accident data are obtained, and a road network topology map is generated based on the topology of the real road network. The road network topology map includes lane nodes, road connecting edges, and traffic signal nodes. The nodes and edges in the road network topology map are associated with road boundary attributes and road traffic rule attributes. The accident scene temporal features are extracted from the multi-source accident data. The accident scene temporal features include the trajectory of the accident object, the state of the accident environment, the road visibility state, and the sensor observation state. A traffic semantic graph is constructed based on the road network topology map and the temporal features of the accident scenario, wherein the traffic semantic graph includes road nodes, traffic participant nodes, environment nodes, and observation nodes; The traffic semantic graph is input into a spatiotemporal graph neural network for node relationship encoding to obtain the spatiotemporal correlation features of the road network. A continuous spatial representation model is constructed based on the spatiotemporal correlation features of the road network. The continuous spatial representation model includes an implicit occupancy field and a neural radiation field. A three-dimensional continuous traffic environment in a digital twin space is generated based on the continuous spatial representation model.

3. The simulation testing method for high-risk autonomous driving scenarios based on digital twins as described in claim 1, characterized in that, The construction of an accident risk representation object based on the completeness of the multi-source accident data includes: Multimodal integrity features are extracted from the multi-source accident data. These multimodal integrity features include data source integrity features, temporal continuity features, spatial coverage integrity features, traffic participant status integrity features, and pre-accident process integrity features. The multimodal integrity features are input into a masked autoencoder network for missing item reconstruction to obtain reconstructed accident features and missing pattern features; Data consistency features are generated based on the reconstructed accident features and the multimodal integrity features; Construct a completeness discrimination vector based on the missing pattern features and the data consistency features; The integrity discrimination vector is input into the graph attention discrimination network for integrity discrimination to obtain the integrity judgment result; If the integrity determination result indicates that the multi-source accident data meets the preset integrity conditions, then an accident triggering chain is constructed based on the multi-source accident data; If the integrity determination result indicates that the multi-source accident data does not meet the preset integrity conditions, then accident scene fragments are constructed based on the missing pattern features and the multi-source accident data. An accident risk representation object is constructed based on the accident triggering chain or the accident scene fragments.

4. The simulation testing method for high-risk autonomous driving scenarios based on digital twins as described in claim 3, characterized in that, The construction of the accident triggering chain based on the multi-source accident data includes: The accident-related time-series data within a preset time window before the accident occurs is extracted from the multi-source accident data. The accident-related time-series data includes the status of traffic participants, road traffic status, environmental status, obstruction status, traffic signal status, and autonomous driving response status. The accident-related time-series data is divided into multiple accident time-series segments based on a preset time segment length. For each accident time segment, risk change factors are analyzed to obtain a risk change factor sequence. The risk change factors include target appearance state, obstruction removal state, deceleration state of traffic participants, lateral intrusion state of traffic participants, traffic signal change state, and autonomous driving response change state. Based on the sequence of risk change factors, a sequence of trigger nodes is constructed according to the temporal relationship of the multiple accident time segments; The trigger node sequence is mapped to the three-dimensional continuous traffic environment to obtain a node twin scene corresponding to each trigger node; Perform counterfactual processing on the corresponding trigger node in each node twin scenario to obtain a counterfactual twin scenario, wherein the counterfactual processing includes removing or replacing the trigger node; Accident risk indicators are calculated based on the node twin scenario and the counterfactual twin scenario, and the causal contribution of each triggering node is determined based on the change between the accident risk indicators. Based on the causal contribution, key triggering nodes are selected from the triggering node sequence, and an accident triggering chain is constructed based on the key triggering nodes.

5. The simulation testing method for high-risk autonomous driving scenarios based on digital twins as described in claim 3, characterized in that, The construction of accident scene fragments based on the missing pattern features and the multi-source accident data includes: Based on the missing pattern characteristics, the missing data types and retained data types are determined from the multi-source accident data; Based on the data type of the retained accident data, an accident retention feature set is extracted from the multi-source accident data. The accident retention feature set includes road structure data, traffic object data, object behavior data, environmental state data, occlusion relationship data, conflict relationship data, and autonomous driving response data. The accident-preserved feature set and the missing data type are input into the fragment semantic coding network for fragment semantic coding to obtain fragment semantic features. Based on the fragment semantic features, the accident-preserved feature set is divided into fragment types to obtain multi-dimensional fragments. The multi-dimensional fragments include road structure fragments, traffic object fragments, object behavior fragments, environmental state fragments, occlusion relationship fragments, conflict relationship fragments, and autonomous driving response fragments. Based on the missing pattern features, the completable association relationships between each fragment in the multi-dimensional fragment are determined to obtain a fragment association graph; The fragment association graph is input into the graph attention fusion network for association weight allocation to obtain the fragment association weights; Based on the fragment association weights, the multi-dimensional fragments are combined to obtain accident scene fragments.

6. The simulation testing method for high-risk autonomous driving scenarios based on digital twins as described in claim 1, characterized in that, The step of generating high-risk test scenarios by performing a matching digital twin scenario instantiation strategy on the accident risk expression object based on the three-dimensional continuous traffic environment and the type of the accident risk expression object includes: When the accident risk expression object is an accident triggering chain, the key triggering nodes in the accident triggering chain are mapped to the three-dimensional continuous traffic environment to obtain the first basic accident scenario; Based on the first accident scenario and the key triggering node, a risk-equivalent disturbance scenario is generated to obtain the first candidate high-risk scenario; When the accident risk expression object is an accident scene fragment, the accident scene fragment is filled with missing information and risk consistency splicing is performed based on the three-dimensional continuous traffic environment to obtain a second basic accident scene. The second basic accident scene is then screened based on preset high-risk scene judgment conditions to obtain a second candidate high-risk scene. High-risk test scenarios are obtained by filtering the first candidate high-risk scenario or the second candidate high-risk scenario based on preset high-risk screening conditions.

7. The simulation testing method for high-risk autonomous driving scenarios based on digital twins as described in claim 6, characterized in that, The process of generating risk-equivalent disturbance scenarios based on the first accident baseline scenario and the key triggering nodes to obtain the first candidate high-risk scenario includes: Key trigger node features are extracted from the key trigger nodes, including the time of node occurrence, the location of node occurrence, the participating objects of node occurrence, and the risk type of node occurrence. Based on the key trigger node characteristics and the first accident basic scenario, perturbed scenario parameters are determined. The perturbed scenario parameters include the speed of traffic participants, the distance between traffic participants, the time of target appearance, the position of obstruction, road surface adhesion parameters, and braking delay parameters. Based on the first accident scenario, the original risk transmission characteristics are determined, and a perturbation constraint vector is constructed based on the original risk transmission characteristics and the perturbable scenario parameters. The perturbation constraint vector is input into a constraint-based generative adversarial network to generate perturbation parameters, thereby obtaining candidate perturbation parameter combinations. The candidate perturbation parameter combinations are filtered based on a preset risk transmission similarity threshold to obtain risk-equivalent perturbation parameter combinations; The risk equivalent disturbance parameters are combined and loaded into the first accident base scenario to obtain the risk equivalent disturbance scenario; Based on the three-dimensional continuous traffic environment, the feasibility of the risk-equivalent disturbance scenario is verified to obtain the first candidate high-risk scenario.

8. The simulation testing method for high-risk autonomous driving scenarios based on digital twins as described in claim 6, characterized in that, The process of completing missing information and stitching together risk consistency data from the accident scene fragments based on the three-dimensional continuous traffic environment to obtain a second basic accident scene includes: Based on the key fragments in the accident scene fragments, a spatiotemporal mask map of fragments is constructed. The key fragments include road structure fragments, traffic object fragments, object behavior fragments, environmental state fragments, occlusion relationship fragments, conflict relationship fragments, and autonomous driving response fragments. Based on the multidimensional constraints in the three-dimensional continuous traffic environment, a completion condition vector is constructed. The multidimensional constraints include road geometry constraints, vehicle kinematic constraints, traffic rule constraints, traffic participant behavior constraints, and sensor visibility constraints. The fragment spatiotemporal mask image and the completion condition vector are input into a pre-constructed spatiotemporal diffusion completion network to complete the missing information and obtain the completed accident scene fragments; Based on the completed accident scene fragments, fragment splicing constraints are determined, including conflict geometric relationships, temporal compatibility relationships, and complementary risk effects relationships; Based on the fragment splicing constraints, the fragments of the completed accident scene are spliced ​​together with risk consistency to obtain the second basic accident scene.

9. The simulation testing method for high-risk autonomous driving scenarios based on digital twins as described in claim 1, characterized in that, The step of converting the high-risk test scenarios into test data corresponding to different functional modules in the autonomous driving system to be tested, and injecting the test data into the corresponding functional modules, includes: By analyzing the scenario execution scripts, scenario parameter configurations, risk triggering nodes, and traffic participant status in high-risk test scenarios, causal description data of high-risk scenarios can be obtained. Obtain the interface description information corresponding to the perception module, prediction module, planning module and control module in the autonomous driving system to be tested, and generate module input constraint data based on the interface description information; Based on the high-risk scenario causal description data and the module input constraint data, a cross-module causal injection graph is constructed. The cross-module causal injection graph includes scenario risk nodes, module input nodes, module output nodes, and module dependency edges. Based on the cross-module causal injection graph, the high-risk test scenario is sliced ​​to obtain a perception scenario slice, a prediction scenario slice, a planning scenario slice, and a control scenario slice. Perception test data is generated based on the perception scene slices, and prediction test data is generated based on the prediction scene slices. The perception test data includes image data, point cloud data, and radar target data. The prediction test data includes traffic participant trajectory data, traffic participant behavioral intention data, and traffic participant movement probability data. Planning test data is generated based on the planning scene slices, and control test data is generated based on the control scene slices. The planning test data includes road boundary data, dynamic obstacle data, traffic rule constraint data, and drivable area data. The control test data includes target trajectory data, target speed data, braking response data, and steering response data. Based on the cross-module causal injection graph, the causal consistency of the perception test data, the prediction test data, the planning test data, and the control test data is verified to obtain causal consistent hierarchical injection test data. According to the injection order corresponding to the module dependency edges, the test data of the causal consistent hierarchical injection is injected into the perception module, the prediction module, the planning module and the control module respectively.

10. A simulation testing system for high-risk autonomous driving scenarios based on digital twins, characterized in that, The system applies the high-risk autonomous driving scenario simulation testing method based on digital twins as described in any one of claims 1 to 9, and the system comprises: The traffic environment construction module is used to acquire the topology of the real road network and multi-source accident data, and to construct a three-dimensional continuous traffic environment in the digital twin space based on the topology and the multi-source accident data. The risk object construction module is used to construct an accident risk expression object based on the completeness of the multi-source accident data. When the multi-source accident data meets the preset completeness conditions, the accident risk expression object includes an accident trigger chain. When the multi-source accident data does not meet the preset completeness conditions, the accident risk expression object includes accident scene fragments. The test scenario construction module is used to generate high-risk test scenarios by executing a matching digital twin scenario instantiation strategy on the accident risk expression object based on the three-dimensional continuous traffic environment and the type of the accident risk expression object. The test data injection module is used to convert the high-risk test scenarios into test data corresponding to different functional modules in the autonomous driving system to be tested, and inject the test data into the corresponding functional modules. The functional modules include a perception module, a prediction module, a planning module, and a control module. The test evaluation module is used to output the test evaluation results of high-risk scenarios based on the response results of the autonomous driving system in the injected scenarios.

Images