Automatic driving difficult example scenario mining method and device, equipment and storage medium

Abstract

The application provides an automatic driving difficult example scene mining method and device, equipment and a storage medium, comprising: obtaining trajectory data of a traffic participant and road map data, performing spatiotemporal alignment processing to obtain standardized traffic flow data, combining road topology of the road map data to obtain motion risk characteristics, generating atomic traffic events, and screening root events from the atomic traffic events; taking the root events as starting nodes to search for candidate traffic participants along the road topology, establishing parent-child propagation relationships after spatiotemporal constraint and response intensity verification, constructing a traffic disturbance propagation tree, determining a scene time boundary, and accordingly, cutting scene data from the standardized traffic flow data and the road map data, combining the traffic disturbance propagation tree to distinguish a complete participant set and a core participant set, and screening an automatic driving difficult example scene set in combination with the traffic disturbance propagation tree; thereby improving the accuracy of difficult example scene mining.

Description

Technical Field

[0001] This application relates to the field of autonomous driving data mining, and in particular to a method, apparatus, device, and storage medium for mining difficult scenarios in autonomous driving. Background Technology

[0002] In the closed loop of autonomous driving data, road test and simulation data are typically spatiotemporally aligned, trajectory matched, and risk event retrieved, and traffic scene segments are extracted accordingly for training and testing.

[0003] However, existing solutions mostly rely on fixed threshold triggering and static time window slicing, which makes it difficult to accurately characterize the propagation range and participation relationships of disturbances in continuous traffic flow. This can easily lead to incomplete event boundaries and interrupted propagation processes. Furthermore, scene filtering and deduplication still cannot reflect the differences in multi-vehicle interaction structures, resulting in insufficient efficiency in identifying high-value complex scenes.

[0004] Therefore, how to improve the accuracy of mining difficult scenarios for autonomous driving from continuous traffic data has become the technical problem that this application needs to solve. Summary of the Invention

[0005] This application provides a method, apparatus, device, and storage medium for mining challenging scenarios in autonomous driving, to solve the aforementioned technical problems. The method is based on traffic participant trajectory data and road map data, which are spatiotemporally aligned to obtain standardized traffic flow data. Motion risk features are extracted by combining road topology. A traffic disturbance propagation tree is constructed by generating atomic traffic events and filtering root events. Based on this, the scene time boundary is determined, scene data is extracted, and the complete participant set and core participant set are distinguished. Finally, a set of challenging scenarios for autonomous driving is obtained through filtering, achieving accurate mining of challenging scenarios.

[0006] Firstly, this application provides a method for mining difficult scenarios in autonomous driving, including:

[0007] Acquire trajectory data of traffic participants and road map data, and perform spatiotemporal alignment processing to obtain standardized traffic flow data;

[0008] Based on the road topology of standardized traffic flow data and road map data, motion risk characteristics are obtained;

[0009] Based on the characteristics of motion risk, atomic traffic events are generated, and root events are obtained by filtering from the atomic traffic events;

[0010] Starting with the root event as the starting node, candidate traffic participants are searched along the road topology. After verifying the spatiotemporal constraints and response strength, parent-child propagation relationships are established to construct a traffic disturbance propagation tree.

[0011] Based on the traffic disturbance propagation tree, the scene time boundary is determined. Based on the scene time boundary, scene data is extracted from standardized traffic flow data and road map data, and the complete participant set and core participant set are distinguished by combining the traffic disturbance propagation tree.

[0012] Based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, a set of challenging scenarios for autonomous driving is obtained.

[0013] In one possible implementation, spatiotemporal alignment processing is performed to obtain standardized traffic flow data, including:

[0014] The trajectory data of traffic participants are synchronized with the time base to obtain synchronized trajectory data;

[0015] Based on the map coordinate system of the road map data, the synchronized trajectory data is subjected to spatial coordinate transformation, followed by noise smoothing and target identity continuity verification to obtain standardized traffic flow data.

[0016] In one possible implementation, motion risk characteristics are obtained based on the road topology of standardized traffic flow data and road map data, including:

[0017] By matching the vehicle trajectory points of standardized traffic flow data to the lane and intersection areas corresponding to the road topology, lane-level positioning data is obtained.

[0018] Based on lane-level positioning data, longitudinal motion parameters, lateral interaction parameters, and collision risk parameters of the vehicle are calculated and integrated to obtain motion risk characteristics.

[0019] In one possible implementation, atomic traffic events are generated based on motion risk characteristics, and root events are obtained by filtering from the atomic traffic events, including:

[0020] Based on the characteristics of sports risks, match the corresponding preset time trigger threshold;

[0021] Search for consecutive intervals exceeding a preset time trigger threshold in the time series corresponding to the motion risk characteristics, merge candidate intervals that are adjacent in time and have the same traffic participants, and generate atomic traffic events based on the candidate intervals.

[0022] Based on the atomic traffic events, the corresponding propagation potential score is determined, and atomic traffic events with propagation potential scores higher than the preset propagation threshold are selected as root events.

[0023] In one possible implementation, candidate traffic participants are searched along the road topology starting from the root event, including:

[0024] The road topology is transformed into a traffic relationship network that includes the physical connectivity foundation and the set of right-of-way rules. Starting from the root event, a breadth-first traversal is performed along the traffic relationship network in multiple directions to determine the search range.

[0025] Identify vehicles whose motion state changes within the search range as candidate traffic participants.

[0026] In one possible implementation, the parent-child propagation relationship is established after spatiotemporal constraints and response strength verification, including:

[0027] Based on the historical movement status of candidate traffic participants before the root event and the road topology, the expected trajectory of the candidate traffic participants is determined, and the expected trajectory deviation is determined based on the deviation between the expected trajectory and the actual trajectory.

[0028] Spatiotemporal constraint verification and propagation direction consistency verification are performed on candidate traffic participants, and false connections are filtered out by combining expected trajectory deviations to obtain real candidate participants;

[0029] Based on the actual candidate participants, calculate the corresponding response strength and propagation edge risk to determine the node inclusion level. Based on the node inclusion level, establish a parent-child propagation relationship between the actual candidate participants as child nodes and their corresponding parent nodes.

[0030] In one possible implementation, based on the node inclusion level, real candidate participants are designated as child nodes and their corresponding parent nodes to establish parent-child propagation relationships, including:

[0031] When the response strength meets the effective threshold and the propagation edge danger is both higher than the first threshold and lower than the second threshold, the node is determined to be a leaf node, and the real candidate participants are used as leaf nodes to establish a parent-child propagation relationship with the corresponding parent nodes.

[0032] When the response strength meets the effective threshold and the propagation edge danger is higher than the second threshold, the node is determined to be included as an intermediate node, and the real candidate participants are used as intermediate nodes to establish a parent-child propagation relationship with the corresponding parent node.

[0033] In one possible implementation, it also includes:

[0034] When the response intensity is lower than the effective threshold, the spatiotemporal constraints are not valid, and / or the propagation edge danger is lower than the first threshold, the node inclusion level is determined to be non-inclusion, and the parent-child propagation relationship between the real candidate participants and their corresponding parent nodes is not established.

[0035] In one possible implementation, determining the scene time boundary based on the traffic disturbance propagation tree includes:

[0036] The difference between the start time and the leading time of the root event in the traffic disturbance propagation tree is used as the scene start time, and the sum of the end time and the delay time of the atomic traffic event corresponding to the latest response node in the traffic disturbance propagation tree is used as the scene end time.

[0037] Determine the scene time boundaries based on the scene start time and scene end time.

[0038] In one possible implementation, scene data is extracted from standardized traffic flow data and road map data based on the scene's time boundary, and combined with a traffic disturbance propagation tree to distinguish between the complete set of participants and the core set of participants, including:

[0039] Based on the scene time boundary, the scene data containing road topology and traffic flow information is extracted from the standardized traffic flow data and the road spatial range corresponding to the road map data.

[0040] Based on traffic flow information and the node coverage of the traffic disturbance propagation tree, the multiple real candidate participants contained in the traffic disturbance propagation tree, as well as the direct participants of the root event and / or atomic traffic events, are defined as the core participant set, and multiple traffic participants within the scene data are defined as the complete participant set.

[0041] In one possible implementation, the road topology is transformed into a traffic relationship network that includes a physical connectivity basis and a set of right-of-way rules, including:

[0042] Extract lane connection relationships, intersection connectivity relationships, and ramp merging and merging relationships from the road topology to form the basis of physical connectivity;

[0043] Extract lane access permissions, intersection turning rules, and traffic light control rules from the road topology to form a set of right-of-way rules;

[0044] By using lanes as the basic unit and linking them with the physical connectivity foundation and the set of right-of-way rules, a traffic relationship network is constructed.

[0045] In one possible implementation, spatiotemporal constraint verification and propagation direction consistency verification are performed on candidate traffic participants, and false connections are filtered in conjunction with expected trajectory deviations, including:

[0046] Calculate the directional deflection angle between the vehicle disturbance direction of the root event and the motion deflection direction of the candidate traffic participants, and verify the consistency of the directional deflection angle with the preset directional deflection angle threshold to obtain the verification result.

[0047] Based on the deviation between the verification results and the expected trajectory, a double filtering process is performed to eliminate candidate traffic participants with false connections.

[0048] In one possible implementation, a set of challenging autonomous driving scenarios is obtained by filtering based on the traffic disturbance propagation tree, the complete set of participants, and the core participant set, including:

[0049] Based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, multiple initial candidate scenarios are generated;

[0050] Each initial candidate scenario is compared with multiple other initial candidate scenarios in a multi-dimensional structure to obtain multiple structure comparison difference results. Initial candidate scenarios with structure comparison difference results greater than a preset threshold are retained to form a set of difficult autonomous driving scenarios.

[0051] In one possible implementation, a set of challenging autonomous driving scenarios is obtained by filtering based on the traffic disturbance propagation tree, the complete set of participants, and the core participant set, including:

[0052] Based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, multiple initial candidate scenarios are generated;

[0053] Based on the preset multi-dimensional segmentation rules, multiple difficult scenarios already included in the pre-built scenario library are divided into multiple coverage units, and the scenario gap ratio of each coverage unit is calculated.

[0054] Multiple candidate scenarios are filtered based on the scenario gap ratio to obtain a set of challenging autonomous driving scenarios.

[0055] In one possible implementation, it also includes:

[0056] The road area corresponding to the traffic disturbance propagation tree is divided into multiple spatiotemporal grid units;

[0057] Density clustering is performed on multiple autonomous driving challenge scenarios within each spatiotemporal grid cell, and the grid boundary of each spatiotemporal grid cell is adjusted based on the clustering results to generate the corresponding scenario index space.

[0058] Based on the scene index space, the traffic disturbance propagation tree is stored in a spatiotemporal fragment and incrementally updated to obtain a spatiotemporally indexed traffic disturbance propagation tree.

[0059] Secondly, this application provides an autonomous driving difficult scenario discovery device, comprising:

[0060] The acquisition module is used to acquire trajectory data of traffic participants and road map data, and perform spatiotemporal alignment processing to obtain standardized traffic flow data;

[0061] The processing module is used to obtain motion risk characteristics based on the road topology of standardized traffic flow data and road map data;

[0062] The processing module is also used to generate atomic traffic events based on the characteristics of motion risks, and to filter out root events from the atomic traffic events;

[0063] The processing module is also used to search for candidate traffic participants along the road topology starting from the root event, and establish parent-child propagation relationships after verifying spatiotemporal constraints and response strength, so as to construct a traffic disturbance propagation tree.

[0064] The processing module is also used to determine the scene time boundary based on the traffic disturbance propagation tree, and extract scene data from standardized traffic flow data and road map data based on the scene time boundary, and combine the traffic disturbance propagation tree to distinguish between the complete participant set and the core participant set.

[0065] The processing module is also used to filter out a set of challenging scenarios for autonomous driving based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants.

[0066] In one possible implementation, the acquisition module is further configured to:

[0067] The trajectory data of traffic participants are synchronized with the time base to obtain synchronized trajectory data;

[0068] Based on the map coordinate system of the road map data, the synchronized trajectory data is subjected to spatial coordinate transformation, followed by noise smoothing and target identity continuity verification to obtain standardized traffic flow data.

[0069] In one possible implementation, the processing module is further configured to:

[0070] By matching the vehicle trajectory points of standardized traffic flow data to the lane and intersection areas corresponding to the road topology, lane-level positioning data is obtained.

[0071] Based on lane-level positioning data, longitudinal motion parameters, lateral interaction parameters, and collision risk parameters of the vehicle are calculated and integrated to obtain motion risk characteristics.

[0072] In one possible implementation, the processing module is further configured to:

[0073] Based on the characteristics of sports risks, match the corresponding preset time trigger threshold;

[0074] Search for consecutive intervals exceeding a preset time trigger threshold in the time series corresponding to the motion risk characteristics, merge candidate intervals that are adjacent in time and have the same traffic participants, and generate atomic traffic events based on the candidate intervals.

[0075] Based on the atomic traffic events, the corresponding propagation potential score is determined, and atomic traffic events with propagation potential scores higher than the preset propagation threshold are selected as root events.

[0076] In one possible implementation, the processing module is further configured to:

[0077] The road topology is transformed into a traffic relationship network that includes the physical connectivity foundation and the set of right-of-way rules. Starting from the root event, a breadth-first traversal is performed along the traffic relationship network in multiple directions to determine the search range.

[0078] Identify vehicles whose motion state changes within the search range as candidate traffic participants.

[0079] In one possible implementation, the processing module is further configured to:

[0080] Based on the historical movement status of candidate traffic participants before the root event and the road topology, the expected trajectory of the candidate traffic participants is determined, and the expected trajectory deviation is determined based on the deviation between the expected trajectory and the actual trajectory.

[0081] Spatiotemporal constraint verification and propagation direction consistency verification are performed on candidate traffic participants, and false connections are filtered out by combining expected trajectory deviations to obtain real candidate participants;

[0082] Based on the actual candidate participants, calculate the corresponding response strength and propagation edge risk to determine the node inclusion level. Based on the node inclusion level, establish a parent-child propagation relationship between the actual candidate participants as child nodes and their corresponding parent nodes.

[0083] In one possible implementation, the processing module is further configured to:

[0084] When the response strength meets the effective threshold and the propagation edge danger is both higher than the first threshold and lower than the second threshold, the node is determined to be a leaf node, and the real candidate participants are used as leaf nodes to establish a parent-child propagation relationship with the corresponding parent nodes.

[0085] When the response strength meets the effective threshold and the propagation edge danger is higher than the second threshold, the node is determined to be included as an intermediate node, and the real candidate participants are used as intermediate nodes to establish a parent-child propagation relationship with the corresponding parent node.

[0086] In one possible implementation, the processing module is further configured to:

[0087] When the response intensity is below the effective threshold, the spatiotemporal constraints are not valid, and / or the propagation edge danger is below the first threshold, the node inclusion level is determined to be non-inclusion, and the parent-child propagation relationship between the real candidate participants and the root event is not established.

[0088] In one possible implementation, the processing module is further configured to:

[0089] The difference between the start time and the leading time of the root event in the traffic disturbance propagation tree is used as the scene start time, and the sum of the end time and the delay time of the atomic traffic event corresponding to the latest response node in the traffic disturbance propagation tree is used as the scene end time.

[0090] Determine the scene time boundaries based on the scene start time and scene end time.

[0091] In one possible implementation, the processing module is further configured to:

[0092] Based on the scene time boundary, the scene data containing road topology and traffic flow information is extracted from the standardized traffic flow data and the road spatial range corresponding to the road map data.

[0093] Based on traffic flow information and the node coverage of the traffic disturbance propagation tree, the multiple real candidate participants contained in the traffic disturbance propagation tree, as well as the direct participants of the root event and / or atomic traffic events, are defined as the core participant set, and multiple traffic participants within the scene data are defined as the complete participant set.

[0094] In one possible implementation, the processing module is further configured to:

[0095] Extract lane connection relationships, intersection connectivity relationships, and ramp merging and merging relationships from the road topology to form the basis of physical connectivity;

[0096] Extract lane access permissions, intersection turning rules, and traffic light control rules from the road topology to form a set of right-of-way rules;

[0097] By using lanes as the basic unit and linking them with the physical connectivity foundation and the set of right-of-way rules, a traffic relationship network is constructed.

[0098] In one possible implementation, the processing module is further configured to:

[0099] Calculate the directional deflection angle between the vehicle disturbance direction of the root event and the motion deflection direction of the candidate traffic participants, and verify the consistency of the directional deflection angle with the preset directional deflection angle threshold to obtain the verification result.

[0100] Based on the deviation between the verification results and the expected trajectory, a double filtering process is performed to eliminate candidate traffic participants with false connections.

[0101] In one possible implementation, the processing module is further configured to:

[0102] Based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, multiple initial candidate scenarios are generated;

[0103] Each initial candidate scenario is compared with multiple other initial candidate scenarios in a multi-dimensional structure to obtain multiple structure comparison difference results. Initial candidate scenarios with structure comparison difference results greater than a preset threshold are retained to form a set of difficult autonomous driving scenarios.

[0104] In one possible implementation, the processing module is further configured to:

[0105] Based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, multiple initial candidate scenarios are generated;

[0106] Based on the preset multi-dimensional segmentation rules, multiple difficult scenarios already included in the pre-built scenario library are divided into multiple coverage units, and the scenario gap ratio of each coverage unit is calculated.

[0107] Multiple candidate scenarios are filtered based on the scenario gap ratio to obtain a set of challenging autonomous driving scenarios.

[0108] In one possible implementation, the processing module is further configured to:

[0109] The road area corresponding to the traffic disturbance propagation tree is divided into multiple spatiotemporal grid units;

[0110] Density clustering is performed on multiple autonomous driving challenge scenarios within each spatiotemporal grid cell, and the grid boundary of each spatiotemporal grid cell is adjusted based on the clustering results to generate the corresponding scenario index space.

[0111] Based on the scene index space, the traffic disturbance propagation tree is stored in a spatiotemporal fragment and incrementally updated to obtain a spatiotemporally indexed traffic disturbance propagation tree.

[0112] Thirdly, this application provides an electronic device, including: a memory and a processor;

[0113] The memory stores instructions that the computer executes;

[0114] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0115] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible embodiments of the first aspect.

[0116] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0117] This application provides a method, apparatus, device, and storage medium for mining challenging scenarios in autonomous driving. It obtains standardized traffic flow data by performing spatiotemporal alignment processing on the trajectory data of traffic participants and road map data, and combines this with road topology to obtain motion risk features, enabling accurate identification of corresponding traffic events from continuous traffic flow. By generating atomic traffic events based on the motion risk features and filtering to obtain root events, a traffic disturbance propagation tree is constructed along the road topology using the root events as starting nodes. This fully presents the correlation and transmission relationships of traffic disturbances, avoiding scene truncation and information loss caused by relying on fixed time window slices. By determining the scene time boundary based on the traffic disturbance propagation tree, scene data is extracted from standardized traffic flow data and road map data, and the complete participant set and core participant set are distinguished, ensuring the integrity of the scene scope and the accuracy of participant segmentation. Finally, a set of challenging scenarios for autonomous driving is obtained by comprehensively filtering the traffic disturbance propagation tree, the complete participant set, and the core participant set, effectively improving the overall accuracy of mining challenging scenarios for autonomous driving. Attached Figure Description

[0118] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0119] Figure 1 This application provides a schematic diagram of an application data processing system architecture.

[0120] Figure 2 A flowchart illustrating the autonomous driving difficult scenario mining method provided in this application embodiment. Figure 1 ;

[0121] Figure 3 A complete scene slice diagram based on the traffic disturbance propagation tree propagation chain provided for the embodiments of this application;

[0122] Figure 4 A flowchart illustrating the autonomous driving difficult scenario mining method provided in this application embodiment. Figure 2 ;

[0123] Figure 5 A flowchart illustrating the autonomous driving difficult scenario mining method provided in this application embodiment. Figure 3 ;

[0124] Figure 6 A schematic diagram of the structure of the autonomous driving difficult scenario discovery device provided in the embodiments of this application;

[0125] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0126] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0127] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0128] Autonomous driving data closed-loop and testing verification falls under the field of spatiotemporal data mining, which involves the coordinated operation of vehicle-side data collection, cloud-based processing, and simulation feedback. It is typically deployed across road test fleets, simulation platforms, and data middleware. The system continuously accesses multi-source traffic data, including vehicle trajectories, high-precision maps, perception logs, control logs, and images and videos. It retrieves, segments, and stores abnormal interactions within continuous traffic flows to support model training, scene playback, and testing verification.

[0129] Existing data processing methods often use fixed risk thresholds or static time windows to capture scenes. They first match trajectories with maps, then filter events based on indicators such as collision time, spacing changes, or deceleration, and save segments within a preset time range. While this approach is usable for single-vehicle anomalies or short-term interactions, it often only records data near a single trigger moment when faced with complex traffic disturbances such as sudden braking, lane cutting, non-motorized vehicle crossings, or merging in construction zones. This fails to fully cover the propagation process of disturbances across multi-vehicle queues, adjacent lanes, and cross-paths. Furthermore, scene filtering and deduplication rely heavily on label consistency, temporal overlap, or trajectory similarity, making it difficult to distinguish between scenes with different propagation structures. This can easily lead to the omission of high-value, complex segments, duplicate accumulation, and missing context, impacting the coverage efficiency and reuse quality of the scene library.

[0130] Therefore, improving the accuracy of identifying challenging scenarios in autonomous driving from continuous traffic data has become a pressing technical problem that needs to be solved.

[0131] To address the aforementioned issues, this application provides a method, apparatus, device, and storage medium for mining challenging scenarios in autonomous driving. By performing spatiotemporal alignment processing on the trajectory data of traffic participants and road map data to obtain standardized traffic flow data, and combining this with road topology to obtain motion risk features, corresponding traffic events can be accurately identified from continuous traffic flow. By generating atomic traffic events based on motion risk features and filtering to obtain root events, a traffic disturbance propagation tree is constructed along the road topology using the root events as starting nodes. This fully presents the correlation and transmission relationships of traffic disturbances, avoiding scene truncation and information loss caused by relying on fixed time window slices. By determining the scene time boundary based on the traffic disturbance propagation tree, scene data is extracted from standardized traffic flow data and road map data, and the complete participant set and core participant set are distinguished, ensuring the integrity of the scene scope and the accuracy of participant segmentation. Finally, a set of challenging scenarios for autonomous driving is obtained by comprehensively filtering the traffic disturbance propagation tree, the complete participant set, and the core participant set, effectively improving the overall accuracy of mining challenging scenarios for autonomous driving.

[0132] The technical solution of this application and how it solves the above-mentioned technical problems will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0133] Figure 1 This is a schematic diagram of an application data processing system architecture provided in an embodiment of this application. The application data processing system 100 is a vehicle domain controller. Figure 1 As shown, the above architecture includes a data acquisition unit 101 and a processing unit 102.

[0134] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the architecture of the application data processing system. In other feasible embodiments of this application, the above architecture may include more or fewer components than illustrated, or combine some components, or split some components, or arrange different components, which can be determined according to the actual application scenario and is not limited here. Figure 1 The components shown can be implemented in hardware, software, or a combination of both.

[0135] In the specific implementation process, the data acquisition unit 101 may include an input / output interface or a communication interface, and the data acquisition unit 101 may be connected to the processing unit through the input / output interface or the communication interface.

[0136] The processing unit 102 can take the trajectory data of traffic participants and road map data as input, perform spatiotemporal alignment processing to obtain standardized traffic flow data, combine it with road topology to obtain motion risk characteristics, generate atomic traffic events based on motion risk characteristics and filter out root events, search for candidate traffic participants along the road topology with the root event as the starting node, establish parent-child propagation relationship after verification to construct a traffic disturbance propagation tree, determine the scene time boundary based on the traffic disturbance propagation tree, extract scene data and distinguish between the complete participant set and the core participant set, and finally filter out the set of autonomous driving difficult example scenarios.

[0137] It should be understood that the above processing unit can be implemented by a processor reading instructions from memory and executing those instructions, or it can be implemented by a chip circuit.

[0138] Furthermore, the network architecture and business scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of network architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0139] Figure 2 A flowchart illustrating the autonomous driving difficult scenario mining method provided in this application embodiment. Figure 1 ,like Figure 2 As shown, the autonomous driving difficult scenario mining method provided in this embodiment includes:

[0140] S201. Obtain trajectory data of traffic participants and road map data, and perform spatiotemporal alignment processing to obtain standardized traffic flow data.

[0141] For example, trajectory data can come from vehicle positioning terminals, roadside sensing devices, or data management platforms. Trajectory data includes at least timestamps, location coordinates, speed, heading, and acceleration information. Road map data can come from high-precision maps, road basic geographic information databases, or roadside mapping results. Road map data includes at least road centerlines, lane boundaries, intersection connections, and road attribute information.

[0142] Spatiotemporal alignment processing can include synchronizing trajectory data with different sampling frequencies using a unified time reference and performing spatial coordinate system transformation on position data under different reference systems, so that the trajectory data of traffic participants and road map data are located on the same time axis and under the same spatial reference.

[0143] After the above processing, the resulting standardized traffic flow data can simultaneously include time information, location information, speed information, heading information, and matching status with road map data, providing a unified input for subsequent calculation of motion risk characteristics based on road topology.

[0144] In one possible implementation, spatiotemporal alignment processing is performed to obtain standardized traffic flow data, including:

[0145] The trajectory data of traffic participants are synchronized with the time reference to obtain synchronized trajectory data. Based on the map coordinate system of the road map data, the synchronized trajectory data is transformed into spatial coordinates and then subjected to noise smoothing and target identity continuity verification to obtain standardized traffic flow data.

[0146] For example, time reference synchronization refers to mapping trajectory records from vehicle-mounted devices, roadside devices, or data management platforms to the same time axis, so that trajectory points with different sampling frequencies and different collection times have a consistent time reference.

[0147] Trajectory data can include information such as timestamps, location coordinates, speed, and heading angle of traffic participants. When synchronizing the time base, alignment corrections are made based on the clock offset, sampling period difference, or timestamp error of each data source to form synchronized trajectory data arranged in a unified time series.

[0148] Spatial coordinate transformation refers to converting the synchronized trajectory data from the acquisition coordinate system to the map coordinate system corresponding to the road map data, so that the trajectory position is on the same spatial reference as the road boundary, lane centerline and road topology.

[0149] Noise smoothing can be achieved by using mean filtering based on sliding windows, trajectory fitting based on spline curves, or state estimation based on Kalman filtering to suppress location jumps, velocity jitters, and short-term anomalies.

[0150] Target identity continuity verification refers to checking and correcting the identity association of the same traffic participant at adjacent times based on the spatial proximity, speed continuity, heading continuity, and target identifier consistency between trajectory points. If rebinding or breakage occurs, the continuous identity is restored by combining the association results of previous and subsequent frames, thereby obtaining standardized traffic flow data with continuous identity.

[0151] In this implementation, time base synchronization, spatial coordinate transformation, noise smoothing, and target identity continuity verification sequentially unify the trajectory data in terms of time and space, suppress anomalies, and organize identities, enabling the output data to establish a stable correspondence directly with the road map data. Standardized traffic flow data serves as the foundational input for subsequent motion risk feature calculation, atomic traffic event generation, and traffic disturbance propagation tree construction, providing continuous, unified, and matchable data support for scene boundary determination and participant propagation relationship identification.

[0152] After the above processing, the temporal offset, spatial deviation, and local noise of the trajectory data are corrected synchronously, and the correlation between targets in the time series remains continuous, thereby improving the consistency between standardized traffic flow data and road topology. Based on this data, false and false associations in subsequent risk feature extraction can be reduced, making the data on which traffic event screening and propagation relationship construction are based more complete and stable, thus improving the accuracy and reusability of difficult scenario mining results.

[0153] S202. Based on the road topology of standardized traffic flow data and road map data, motion risk characteristics are obtained.

[0154] For example, topological parsing can be performed on road map data, and trajectory points in standardized traffic flow data can be associated with road topology to establish a spatial correspondence between traffic participants and road topology.

[0155] Based on this, interaction risk indicators can be calculated by combining longitudinal motion state, lateral interaction characteristics, and spatial relationship with other traffic participants. Longitudinal motion state can be characterized by relative speed, deceleration, headway, or braking trend. Lateral interaction characteristics can be characterized by lane deviation, cutting angle, lateral speed, or lane change intention. Spatial relationship can be characterized by minimum distance, collision time, rear approach relationship, or cross-conflict relationship.

[0156] Multiple interactive risk indicators can be combined into a motion representation vector according to preset rules, which is then used as the output of motion risk features.

[0157] By constraining the spatial assignment of trajectory points through road topology and calculating the interaction risk within the corresponding topological structure, the motion risk characteristics can directly reflect the degree of conflict and disturbance intensity of traffic participants in a specific road structure.

[0158] In one possible implementation, the motion risk characteristics are obtained, including:

[0159] Vehicle trajectory points from standardized traffic flow data are matched to lane and intersection areas corresponding to the road topology to obtain lane-level positioning data. Based on the lane-level positioning data, longitudinal motion parameters, lateral interaction parameters, and collision risk parameters of vehicles are calculated and integrated to obtain motion risk features.

[0160] For example, lane-level positioning data is used to represent the correspondence between vehicle trajectory points and specific lanes and intersection areas in the road topology. The process of establishing this data includes parsing the road topology and mapping the trajectory points in the standardized traffic flow data to the corresponding spatial units based on the lane centerline, intersection entrance lanes, exit lanes, and the geometric boundaries of conflict areas.

[0161] Lane-level positioning data can define the actual position of a vehicle within the road structure and serve as the spatial basis for subsequent motion parameter calculations.

[0162] Vehicle longitudinal motion parameters are used to describe the vehicle's speed, acceleration, and deceleration characteristics in the direction of travel. Lateral interaction parameters are used to describe the vehicle's lateral offset relative to the lane centerline, lateral velocity, heading angle changes, and interaction with adjacent lane targets. Collision risk parameters are used to quantify the potential conflict level between the vehicle and forward targets, lateral cutting targets, or intersection conflict targets.

[0163] The above parameters can be calculated from the trajectory increments of adjacent sampling points, vehicle posture information, and relative distance, relative speed, and spatiotemporal overlap with other traffic participants. They are then further normalized and weighted to form motion risk characteristics.

[0164] In practice, lane-level positioning data can be completed by the vehicle's map matching unit. Based on the lane topology primitives, intersection area polygons and connectivity relationships in the high-precision map, this unit performs nearest lane candidate retrieval, direction consistency verification and area affiliation determination for each trajectory point, thereby outputting point-level lane identification, area identification and corresponding confidence level.

[0165] Longitudinal motion parameters can be obtained by combining the position projection difference, velocity difference, and acceleration difference between adjacent moments. Lateral interaction parameters can be obtained by combining the normal distance from the trajectory point to the lane centerline, the heading angle deviation, and the relative displacement of adjacent vehicles. Collision risk parameters can be obtained by combining the time distance, collision time, spacing margin, and the entry sequence of the conflict zone.

[0166] By aligning the above parameters according to a unified time reference and then concatenating them, a motion risk feature vector for atomic traffic event identification can be constructed. This feature vector can simultaneously characterize the vehicle's forward movement along the lane, cross-lane interaction, and conflict trends with other traffic participants, and provide quantitative input for the subsequent generation of atomic traffic events.

[0167] By mapping trajectory points to lane and intersection areas and establishing lane-level positioning data, the spatial position of vehicles is transformed from continuous coordinates to road structure semantic coordinates, enabling longitudinal, lateral, and conflict risks to be calculated under unified topological constraints. The motion risk features formed in this way have clear road semantic associations and can be used to identify complex interaction states such as following, cutting in, crossing, and intersection conflicts, and provide a stable risk representation for the subsequent construction of disturbance propagation relationships.

[0168] S203. Based on the characteristics of motion risk, generate atomic traffic events and filter out root events from the atomic traffic events.

[0169] For example, based on preset detection rules and risk thresholds, candidate intervals that meet the triggering conditions can be searched within the continuous time interval corresponding to the motion risk characteristics. After merging candidate intervals that are temporally adjacent and have the same traffic participants, an atomic traffic event with a clear start and end time can be formed.

[0170] An atomic traffic event can correspond to the smallest traffic abnormality interaction unit, such as a lane change and cut-in, a sudden braking, a crossing, a following and approaching, or a rushing through an intersection. Each atomic traffic event can record the participating objects, the location of occurrence, the duration, and the judgment criteria.

[0171] Subsequently, the propagation potential of each atomic traffic event can be assessed based on the event type, risk level, duration, and subsequent propagation space of the road topology. Atomic traffic events with propagation potential scores higher than a preset propagation threshold are selected as root events, which serve as the starting nodes of the traffic disturbance propagation tree.

[0172] By first discretizing continuous risks into atomic traffic events, and then selecting events with high propagation potential as root events, we can provide clear basic units and starting conditions for subsequent propagation relationship searches.

[0173] In one possible implementation, atomic traffic events are generated based on motion risk characteristics, and root events are obtained by filtering from the atomic traffic events, including:

[0174] Based on the characteristics of movement risks, a corresponding preset time trigger threshold is matched; in the time series corresponding to the movement risk characteristics, continuous intervals exceeding the preset time trigger threshold are searched, and candidate intervals with adjacent times and the same traffic participants are merged. Based on the candidate intervals, atomic traffic events are generated; based on the atomic traffic events, the corresponding propagation potential score is determined, and atomic traffic events with propagation potential scores higher than the preset propagation threshold are selected as root events.

[0175] In this embodiment, the preset time trigger threshold is used to limit the trigger boundary of different motion risk features in the time dimension. It can be pre-configured according to the subsequent spatial range corresponding to the event type, risk level or road topology, and the corresponding threshold is selected according to the type or value range of the motion risk feature during actual matching.

[0176] Atomic traffic events are used to characterize the smallest unit of traffic anomaly interaction, which includes a clear start and end time, participating traffic participants, and the criteria for determining the validity of the event.

[0177] The propagation potential score is used to measure the likelihood that an atomic traffic event will trigger a chain reaction of subsequent vehicle responses, while the preset propagation threshold is used to distinguish between ordinary events and root events with propagation analysis value.

[0178] For example, the system first calculates the longitudinal motion parameters, lateral interaction parameters, and collision risk parameters of vehicles based on standardized traffic flow data, and integrates them into a time-varying motion risk feature sequence. For different feature types in this sequence, the system pre-establishes a threshold mapping table, configures corresponding time trigger thresholds for risk types such as sudden deceleration, cutting-in conflict, and cross-intrusion, and corrects the thresholds by combining the length of the subsequent propagation space in the road topology, the number of lanes, or the range of the intersection conflict zone.

[0179] Subsequently, the system retrieves continuous segments that consistently exceed the threshold in the corresponding time series, merges segments that are temporally adjacent and correspond to the same combination of traffic participants into candidate intervals, and then generates atomic traffic events based on the start and end times of the candidate intervals, the identities of the participants, and risk assessment criteria. For cases where the event boundaries are slightly discontinuous but still belong to the same abnormal interaction, the system can perform continuous merging according to the preset maximum interval duration to ensure that the boundaries of the event units are clear and do not destroy the integrity of the same disturbance.

[0180] After generating an atomic traffic event, the system combines the event type, duration, risk peak, number of participants, and subsequent road topology space to calculate the propagation potential score for each event, and identifies events that exceed a preset propagation threshold as root events, serving as the starting nodes for constructing the traffic disturbance propagation tree.

[0181] Specifically, the system sets up corresponding event detectors for different types of traffic events, and each detector is configured with K. y It can be represented as:

[0182]

[0183] in, Indicates the detector type; It can represent a state or an event. Represents the set of detection parameters; Indicates the time range before and after slicing; This indicates the weight used to calculate the confidence level.

[0184] The detected first Atom traffic incident Represented as:

[0185]

[0186] in, and These represent the start and end times of the event, respectively. For the participants; Event type; To test the confidence level; This represents the risk value of the j-th atomic traffic event; This represents the set of criteria used to prove the validity of the atomic traffic event.

[0187] The detection confidence score is determined based on the reliability of map matching during the event duration, the degree of change in vehicle movement, the intensity of interaction between traffic participants, and the degree to which the event duration is satisfied. The calculation method is as follows:

[0188]

[0189] in, This indicates the reliability of the trajectory-road topology matching during the event; Indicates the degree of change in vehicle motion, such as speed, acceleration, and lateral displacement; Indicates the intensity of interaction between participants; Indicates the degree to which the duration of the event meets the requirement; For the corresponding weights, satisfying And all weights are non-negative.

[0190] The risk value of an atomic traffic incident is calculated as follows:

[0191]

[0192] in, The risk calculation function corresponding to the event type. This is a set of risk characteristics.

[0193] After generating atomic traffic events, the system calculates the propagation potential score for each event. The calculation method is as follows:

[0194]

[0195] in, Indicates whether an event type belongs to the set of propagable events; Indicates the risk of atomic events; This indicates whether the road topology has a subsequent search space; Indicate whether there are any potentially affected entities within the scope of influence; Represents the normalization function; and Non-negative weights; This is a bias term.

[0196] Based on propagation potential and detection confidence, atomic traffic events will be classified. Divided into ordinary direct output event sets and root event queue :

[0197]

[0198] in, Indicates the threshold of transmission potential. This represents the confidence threshold for the root event; for entering... Atomic events are considered normal direct output events and are directly processed in subsequent scene slices; for events entering... The atomic event is the root event, and the construction of the traffic disturbance propagation tree continues.

[0199] The above processing first transforms motion risk characteristics into well-defined atomic traffic events, and then uses propagation potential scoring to select root events with diffusion value, thereby ensuring that subsequent propagation relationship tracing is based on consistent basic event units. Since root events have both time boundary and propagation starting point attributes, the system can use this to identify subsequent response objects along the road topology and retain the complete link from triggering to propagation of abnormal interactions.

[0200] After adopting this implementation method, the time triggering conditions of motion risk characteristics and event generation rules can be matched differently according to risk type, the boundaries of atomic traffic events are clearer, the root event screening is more in line with the disturbance diffusion characteristics, thereby improving the integrity, distinguishability and accuracy of complex traffic anomaly segments and subsequent propagation tree construction.

[0201] S204. Starting from the root event, search for candidate traffic participants along the road topology. After verifying the spatiotemporal constraints and response strength, establish parent-child propagation relationships to construct a traffic disturbance propagation tree.

[0202] For example, starting from the spatial location, time of occurrence, and participating objects corresponding to the root event, a search can be conducted along the road topology for candidate traffic participants that may be affected by the disturbance.

[0203] During the search process, the actual response time of the candidate traffic participant can be compared with the occurrence time of the root event. When the response delay meets the preset delay range, it is retained as the object to be verified. Then, the spatial location of the candidate traffic participant and the road topology connectivity are verified to confirm that the target is within the road topology range that the root event disturbance can propagate to.

[0204] Subsequently, the response strength is verified. When the response strength of the candidate traffic participant meets the preset threshold, the parent-child propagation relationship between the root event and the candidate traffic participant is established.

[0205] After establishing the propagation relationship, the newly formed child node can be used as a new propagation starting point to continue searching for subsequent candidate traffic participants along the road topology, and the spatiotemporal constraints and response strength verification can be repeatedly performed to form a traffic disturbance propagation tree. The traffic disturbance propagation tree is used to characterize the diffusion path, hierarchical relationship and propagation range of the disturbance among multiple traffic participants, so that the subsequent scene boundary and participant division no longer depend on a fixed time window, but are determined based on the actual propagation link.

[0206] S205. Based on the traffic disturbance propagation tree, determine the scene time boundary. Based on the scene time boundary, extract scene data from standardized traffic flow data and road map data, and combine the traffic disturbance propagation tree to distinguish between the complete participant set and the core participant set.

[0207] For example, the scene time boundary can be determined based on the start time of the root event in the traffic disturbance propagation tree and the end time corresponding to the latest response node.

[0208] Once the time boundaries of the scene are determined, scene data can be extracted from the standardized traffic flow data, road map data, and associated traffic flow data corresponding to that time range.

[0209] Subsequently, traffic participants within the scene can be labeled based on the node coverage in the traffic disturbance propagation tree. Traffic participants appearing within the scene's time boundary are defined as the complete participant set, while those participating in the root event and subsequent propagation are defined as the core participant set. The complete participant set is used to preserve the scene context, while the core participant set is used to label key objects that have a direct response relationship to the occurrence and spread of the disturbance. Together, they form the participant basis for scene selection.

[0210] In one possible implementation, determining the scene time boundary based on the traffic disturbance propagation tree includes:

[0211] The scene start time is determined by the difference between the start time and the leading time of the root event in the traffic disturbance propagation tree, and the scene end time is determined by the sum of the end time and the delay time of the atomic traffic event corresponding to the latest response node in the traffic disturbance propagation tree. The scene time boundary is determined based on the scene start time and the scene end time.

[0212] In this embodiment, the scene start time is used to define the starting point of the scene slice, and the scene end time is used to define the ending point of the scene slice. Together, they represent the complete propagation process corresponding to the traffic disturbance propagation tree.

[0213] The start time of the root event in the traffic disturbance propagation tree refers to the moment when the atomic traffic event corresponding to the root event is first identified as an abnormal interaction; the end time of the atomic traffic event corresponding to the latest response node refers to the moment when the event corresponding to the last node in the propagation chain that has a valid response ends.

[0214] By taking the start time of the root event as the starting point and the end time of the event corresponding to the latest response node as the ending point, the scene time boundary can cover the entire process of disturbance occurrence, diffusion and convergence.

[0215] In the specific implementation, after the system completes the construction of the traffic disturbance propagation tree, it extracts the timestamps of the atomic traffic events corresponding to each node in the tree, and tracks the response order of each node according to the parent-child propagation relationship.

[0216] The start time of the root event is taken from the start timestamp of its atomic traffic event, and the latest response node is determined based on the valid node at the end of the response link in the propagation tree, with the end timestamp of its corresponding atomic traffic event serving as the scene termination time.

[0217] Subsequently, the system forms a basic time interval based on the start and end times of the scene, and can directly extract the corresponding traffic data segments based on this interval. The time boundaries correspond one-to-one with the propagation structure, without relying on a fixed-length time window.

[0218] Specifically, first, obtain the response end time of the latest valid response node in the traffic disturbance propagation tree:

[0219]

[0220] in, Indicates the first The end time of the latest response node in the traffic disturbance propagation tree; This represents the set of leaf nodes in the traffic disturbance propagation tree; Represents a node The end time of the corresponding response event.

[0221] The time range of the scene is:

[0222]

[0223] in, Indicates the start time of the root event; Indicates the leading time of the scene; Indicates the delay time of the scene.

[0224] In practical applications, the scene time boundary can also be connected with the preceding data buffer and the subsequent tail observation area to completely preserve the background state before the disturbance occurs and the convergence state after the disturbance ends. When there are concurrent response nodes in the propagation tree, the system determines the scene termination time based on the node that last completes a valid response, thereby ensuring that the boundary covers the farthest time range of multi-vehicle chain propagation.

[0225] By adopting the aforementioned time boundary, the truncation results can be accurately matched with the traffic disturbance propagation process, reducing fragment truncation or background redundancy caused by static time windows, and improving the consistency and reusability of scene data and propagation structure.

[0226] Optionally, based on the above embodiments, scene data is extracted from road map data according to the scene time boundary and the traffic disturbance propagation tree is used to distinguish between the complete participant set and the core participant set, including:

[0227] Based on the scene time boundary and the corresponding road spatial range of standardized traffic flow data and road map data, scene data containing road topology and traffic flow information is extracted; based on the traffic flow information and the node coverage of the traffic disturbance propagation tree, the multiple real candidate participants contained in the traffic disturbance propagation tree, as well as the direct participants of the root event and / or atomic traffic events, are defined as the core participant set, and the multiple traffic participants in the scene data are defined as the complete participant set.

[0228] In this embodiment, the road spatial range is used to define the spatial boundary for extracting scene data from standardized traffic flow data and road map data, and the scene time boundary is used to define the effective time interval corresponding to the disturbance propagation process. The two together constrain the scope of scene data acquisition.

[0229] Road topology is used to characterize lanes, intersections, ramps and their connectivity, while traffic flow information is used to characterize the motion state, speed changes, relative positional relationships and interactive responses of each traffic participant within the time boundary.

[0230] The node coverage range is used to represent the range of participants actually affected by the nodes in the traffic disturbance propagation tree. The core participant set corresponds to the real candidate participants directly associated with the propagation chain, while the complete participant set corresponds to all traffic participants and their background context retained from the scene data.

[0231] In the specific implementation, road elements corresponding to the time period are first extracted from standardized traffic flow data and road map data according to the scene time boundary, and the spatial area is trimmed according to the road spatial range. The lane center line, lane boundary, intersection connection relationship and traffic participant trajectory records are uniformly organized into scene data.

[0232] Scene data can consist of a rasterized map layer, a lane-level topology layer, and a participant trajectory layer. The rasterized map layer is used to maintain spatial continuity, the lane-level topology layer is used to maintain traversable relationships, and the participant trajectory layer is used to record dynamic traffic conditions.

[0233] Subsequently, by combining the trajectory changes, response times and interaction intensity in the traffic flow information, as well as the coverage of each node in the traffic disturbance propagation tree, the objects identified as real candidate participants in the tree are summarized, and the core participant set is formed by combining the direct participants of the root event and / or atomic traffic event.

[0234] At the same time, all traffic participants appearing within the spatial and temporal window in the scene data are uniformly retained as a complete set of participants, so as to preserve both the key interaction chain and the surrounding traffic background.

[0235] If the scene data is in vector map format, it can be directly extracted through road geometric boundaries and time index; if multi-level index storage is used, it can be read through the combined effect of spatial index and time index. In practical applications, other forms of data organization can also be selected, and this application does not limit them.

[0236] This method first locks the target spatiotemporal window based on the scene's temporal boundaries and road spatial range. Then, it extracts corresponding road topology and traffic flow information from standardized traffic flow data and road map data, thus forming replayable and analyzable scene data. Subsequently, based on traffic flow information and node coverage, it identifies real candidate participants who directly contribute to disturbance propagation, as well as direct participants of root events and / or atomic traffic events, as core participants, while retaining all participants within the scene as a complete participant set. Through this distinction, the scene can highlight the main disturbance propagation link while preserving the surrounding interaction background.

[0237] With this implementation method, the scene data can simultaneously include road structure information and dynamic traffic information, and the separation between core participants and complete participants is clearer. This enables subsequent scene retrieval, verification, and simulation re-feedback to accurately locate key interaction objects and maintain the necessary contextual integrity.

[0238] In summary, Figure 3 A complete scene slice diagram based on the traffic disturbance propagation tree propagation chain provided in the embodiments of this application is shown below. Figure 3 As shown in the diagram, this schematic diagram uses a timeline as a reference to illustrate the logic for defining scene slices and the rules for dividing participants.

[0239] The diagram uses the root event of the traffic disturbance propagation tree as the starting source of the disturbance propagation. The root event has a corresponding duration, and direct response events and subsequent response events are generated sequentially along the direction of disturbance propagation, which together constitute a complete multi-vehicle disturbance propagation link.

[0240] The start time of a scene slice is the start time of the root event minus a preset lead time. The end time of a scene slice is the end time of the atomic traffic event corresponding to the latest effective response node in the traffic disturbance propagation tree plus a preset delay time. The start and end times together define the time range of the complete scene slice, i.e., the scene time boundary.

[0241] All traffic participants appearing within the time frame of a complete scene slice constitute a complete participant set, which can retain the background traffic flow information required for scene playback and simulation reconstruction; the traffic participants corresponding to the root event and the traffic participants corresponding to each node of the traffic disturbance propagation tree constitute a core participant set, which is used to identify the key objects participating in the disturbance propagation process.

[0242] S206. Based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, a set of challenging scenarios for autonomous driving is obtained.

[0243] For example, the extracted scene data can be filtered based on the propagation structure of the traffic disturbance propagation tree;

[0244] When a scenario corresponds to a relatively complex propagation chain, a large number of concurrent response branches, or obvious cross-lane and cross-participant disturbance diffusion characteristics, it can be regarded as a candidate difficult scenario.

[0245] At the same time, the results of distinguishing between the complete set of participants and the core set of participants can be combined to make a judgment on deduplication and retention of the scene, retaining the scene that contains both complete interaction context and can reflect the key propagation structure, and eliminating the segment that only contains a single trigger fragment but lacks propagation context.

[0246] The set of challenging autonomous driving scenarios obtained after the above screening can be used for scenario import, model training, simulation backfeeding, or testing and verification. Distinguishing propagation structures based on traffic disturbance propagation trees and jointly screening with the complete participant set and the core participant set ensures that the output scenario simultaneously possesses complete context, clearly defined core interaction objects, and a reusable propagation structure expression.

[0247] In one possible implementation, a set of challenging autonomous driving scenarios is obtained by filtering based on the traffic disturbance propagation tree, the complete set of participants, and the core participant set, including:

[0248] Based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, multiple initial candidate scenarios are generated. Each initial candidate scenario is then compared with the other multiple initial candidate scenarios in a multi-dimensional structure to obtain multiple structure comparison difference results. Initial candidate scenarios with structure comparison difference results greater than a preset threshold are retained to form a set of autonomous driving difficult scenario scenarios.

[0249] In this embodiment, the traffic disturbance propagation tree is used to represent the parent-child propagation relationship from the root event to multiple traffic participants, the complete participant set is used to cover all participants within the scene time boundary, and the core participant set is used to identify participants who have a direct response to the disturbance propagation.

[0250] The initial candidate scenarios are generated based on the combination of the above three factors, and can correspond to different combinations of participant coverage, time truncation range and structural expression granularity under the same propagation tree.

[0251] The structural comparison difference results are used to represent the degree of difference between different initial candidate scenarios in terms of the number of propagation layers, node coverage, parent-child relationship patterns, key response timing, and participant role distribution. The preset threshold is used to limit the retention conditions.

[0252] For example, the root event, leaf nodes, intermediate nodes and the response time of each node are first extracted based on the traffic disturbance propagation tree. Then, multiple initial candidate scenarios are generated by combining the complete set of participants and the core set of participants. Each initial candidate scenario includes at least the scenario time boundary, a subset of participants, a propagation link representation and node role labeling.

[0253] Subsequently, each initial candidate scenario is compared with the other initial candidate scenarios in a multi-dimensional structure. The multi-dimensional structure comparison includes the joint calculation of propagation tree depth, number of branches, number of concurrent nodes in the same layer, coverage of core participants, coverage of complete participants, and node time offset. The differences in each dimension are normalized to form the structure comparison difference results.

[0254] When the difference result is greater than the preset threshold, it indicates that the initial candidate scenario has independence and distinctiveness in the propagation structure, and it is retained and included in the set of difficult autonomous driving scenarios; when the difference result is not greater than the preset threshold, it is treated as a scenario with similar structure and is eliminated or merged.

[0255] Through the above methods, initial candidate scenarios can form multiple structural representations under the same disturbance source. By comparing multi-dimensional structures, scenarios with significant differences in propagation patterns are retained, enabling the scenario set to possess both structural distinctiveness and coverage integrity. This allows for the selection of complex traffic interaction segments with similar labels but different propagation paths, reducing the accumulation of duplicate scenarios and improving the ability of the autonomous driving scenario set to represent complex disturbance propagation patterns.

[0256] In another possible implementation, based on the traffic disturbance propagation tree, the complete set of participants, and the core participant set, a set of challenging autonomous driving scenarios is obtained, including:

[0257] Based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, multiple initial candidate scenarios are generated. According to the preset multi-dimensional division rules, multiple difficult scenarios already included in the pre-built scenario library are divided into multiple coverage units, and the scenario gap ratio of each coverage unit is calculated. Based on the scenario gap ratio, multiple candidate scenarios are screened to obtain a set of autonomous driving difficult scenarios.

[0258] In this embodiment, the traffic disturbance propagation tree is used to represent the response chain propagating outward from the root event, the complete participant set is used to represent all traffic participants within the scene time boundary, and the core participant set is used to represent the participants that play a decisive role in the propagation of disturbances.

[0259] The coverage unit is a statistical unit formed according to multi-dimensional division rules. The multi-dimensional division rules include at least the root event type, propagation topology, risk range and environmental semantics, so that the difficult scenarios already included in the scenario library can be clustered and statistically analyzed by category.

[0260] The scenario gap ratio is used to measure the gap between the number of scenarios already included in the database within a certain coverage unit and the target coverage requirement. The larger the value, the higher the degree of missingness of that type of scenario in the scenario database.

[0261] In practical implementation, the propagation level, number of branches and termination node information of the root event can be extracted based on the traffic disturbance propagation tree. Then, by combining all traffic participants in the complete participant set and key participants in the core participant set, several initial candidate scenarios with different scenario boundary expression forms can be generated.

[0262] Subsequently, the features of the difficult scenarios already included in the pre-built scenario library are merged, and mapped to the corresponding coverage units according to the root event type, propagation topology, risk range and environmental semantics. Then, the number of covered, the number of targets or the target weight of each coverage unit are counted, and the scenario gap ratio is calculated accordingly.

[0263] Then, each initial candidate scenario is associated with the gap ratio of its corresponding coverage unit. The candidate scenarios are assigned a screening priority according to the gap ratio, and the candidate scenarios that can fill the gap coverage units are retained to form a set of difficult autonomous driving scenarios.

[0264] If there are differences in the encoding method, statistical caliber, or gap ratio calculation model of the coverage unit in practical applications, this application does not impose any restrictions on these aspects.

[0265] This approach generates initial candidate scenarios by propagation trees, complete participant sets, and core participant sets, ensuring that the candidate scenarios are consistent with the actual perturbation links. Then, it filters the candidate scenarios by using the ratio of coverage units to scenario gaps, so that the objects added to the database correspond to the currently missing categories in the scenario database, thereby making the final set of difficult scenario examples more concentrated in supplementing complex interaction types.

[0266] Specifically, the system first evaluates the data completeness of the candidate scenarios, and then... The data completeness of each candidate scenario is represented as follows:

[0267]

[0268] in, Indicates the first The completeness components of each candidate scenario include trajectory completeness rate, map matching stability, target identity continuity, and semantic tag effectiveness. Indicates the first The values ​​of each completeness component; This indicates the corresponding weight.

[0269] Different risk calculation methods are used for ordinary atomic event scenarios and traffic perturbation spread tree (TPST) scenarios:

[0270]

[0271] in, Indicates the risk of the root event; This represents the maximum risk of all propagation edges in TPST; r represents the average risk of propagation. uv E represents the danger level of the propagation edge between parent node u and child node v in the traffic disturbance propagation tree. i This represents the set of parent-child propagation edges of the i-th traffic disturbance propagation tree; All are non-negative weights; This means limiting the calculation result to the range of 0 to 1. For the i-th candidate scene, the j-th... (i) The risk of individual atomic traffic incidents.

[0272] The impact of disturbance propagation is calculated based on propagation depth, propagation width, number of key participants, and propagation duration:

[0273]

[0274] in, Indicates the impact of disturbance propagation. and These represent the normalized propagation depth, maximum propagation width, number of core participants, and propagation duration, respectively. The weights are non-negative.

[0275] Furthermore, the long-tail degree of scene semantics is calculated based on the existing number of scene semantic combinations in the current scene library. Let the smooth occurrence probability of the semantic combinations corresponding to candidate scenes in the current scene library be denoted as... for:

[0276]

[0277] in, This indicates the number of existing semantic combinations in the scene library; Indicates the total number of scenes in the library; This indicates the probability of occurrence of a coarser-grained parent combination; This represents the smoothing coefficient.

[0278] Semantic long tail degree N i Represented as:

[0279]

[0280] in, The minimum reference probability lower bound used for long-tail normalization; To prevent abnormal numerical values, a very small positive number is set; the first During the evaluation round, the scenario library for that round is used. Recalculate.

[0281] The current scene library is divided into multiple coverage units according to a preset multi-dimensional partitioning rule. Let the first... The target number of coverage units is Current scene library The existing number of this coverage unit is Then the first Coverage gap ratio of each coverage unit Represented as:

[0282]

[0283] The first scene in the scene library When the number of coverage units has reached the target number, the coverage gap is zero; when the number of existing units is lower than the target number, the smaller the number of existing units, the larger the corresponding coverage gap.

[0284] Candidate scenarios Coverage gain relative to the current scene library Represented as:

[0285]

[0286] in, Indicates the number of coverage units; Indicates the first The importance of each coverage unit; Representing candidate scenarios For the first The coverage level of each coverage unit.

[0287] Maximum overlap between candidate scenes and existing scenes in the current scene library Represented as:

[0288]

[0289] in, Represents the similarity function; It should include at least the main event type, TPST structure, dissemination depth, dissemination breadth, key risk indicators, road tags, environmental tags, and participant types.

[0290] Dynamic incremental value of candidate scenes relative to the current scene library Represented as:

[0291]

[0292] in, Indicates the difficulty level of the optional model; and All are non-negative parameters.

[0293] The scenario is divided into outputs based on dynamic incremental value and completeness:

[0294]

[0295] in, Indicates the value selection threshold; This represents the minimum available completeness threshold; This represents the threshold for the completeness of the simulation reconstruction, and satisfies... .

[0296] After adopting the above method, the selection of candidate scenes no longer relies solely on a single label or local similarity, but is dynamically allocated based on the scene library coverage status. This can improve the category balance and missing completion capabilities of the scene library, and enhance the support of output difficult examples for subsequent training, playback, and testing.

[0297] Optionally, before the formal screening is executed, the system first performs basic gating checks on the initial candidate scenarios to filter out scenarios that do not meet the minimum quality requirements:

[0298]

[0299] in, The detection confidence level of the atomic events corresponding to the candidate scene; The minimum detection confidence threshold; For the completeness of data for candidate scenarios; The minimum usable completeness threshold is used; candidate scenarios that meet both conditions enter the subsequent screening process to form a candidate set. .

[0300] The selection process employs an iterative selection mechanism. From the candidate scenarios that have not yet been added to the database, select the scenario with the greatest current dynamic incremental value. :

[0301]

[0302] in, Indicates the first The scene library after the selection round is completed. Representing candidate scenarios The dynamic incremental value relative to the current scenario library.

[0303] A selected scene is added to the scene library only when its dynamic incremental value reaches a preset selection threshold.

[0304]

[0305] in, Select a threshold for the value of the scenario. This is the scene library after the rth round of updates.

[0306] After each round of scenario entry, the system synchronously updates the existing quantity and scenario duplication of the corresponding coverage unit, and recalculates the dynamic incremental value of the remaining candidate scenarios; when the maximum dynamic incremental value of the remaining scenarios is lower than The iterative selection will stop when the maximum number of items to be added or the maximum storage capacity set for this round is reached.

[0307] The above approach enables the pre-filtering of low-quality candidate scenarios through basic gating to ensure consistency of the screening benchmark. The iterative selection mechanism, combined with the real-time coverage status of the scenario library, dynamically adjusts the priority and determines whether to include the scenario in the library. This improves the accuracy of difficult scenario screening, while dynamically filling gaps in the scenario library coverage, reducing redundancy of duplicate scenarios, and improving the efficiency and category balance of scenario library construction.

[0308] This application provides a method for mining challenging scenarios in autonomous driving. It obtains standardized traffic flow data by performing spatiotemporal alignment processing on the trajectory data of traffic participants and road map data. Combined with road topology, motion risk features are obtained, enabling accurate identification of corresponding traffic events from continuous traffic flow. Atomic traffic events are generated based on the motion risk features, and root events are selected. A traffic disturbance propagation tree is constructed along the road topology starting from the root events, fully presenting the correlation and transmission relationships of traffic disturbances and avoiding scene truncation and information loss caused by relying on fixed time window slices. The scene time boundary is determined based on the traffic disturbance propagation tree, and scene data is extracted from standardized traffic flow data and road map data accordingly, distinguishing between the complete participant set and the core participant set, ensuring the integrity of the scene scope and the accuracy of participant segmentation. Finally, a set of challenging scenarios for autonomous driving is obtained by comprehensively filtering the traffic disturbance propagation tree, the complete participant set, and the core participant set, effectively improving the overall accuracy of mining challenging scenarios for autonomous driving.

[0309] Figure 4 A flowchart illustrating the autonomous driving difficult scenario mining method provided in this application embodiment. Figure 2 ,like Figure 4 As shown, this embodiment, based on the above embodiments, details the search process for candidate traffic participants and the specific process for establishing parent-child propagation relationships, including:

[0310] S401. Transform the road topology into a traffic relationship network that includes the physical connectivity foundation and the right-of-way rule set. Starting from the root event, perform a breadth-first traversal along the traffic relationship network in multiple directions to determine the search range.

[0311] For example, a traffic relationship network is used to map road topology into a searchable propagation analysis structure, where the physical connectivity basis is used to characterize spatial reachability relationships such as lane connections, intersection connections, and ramp merging and merging, and the right-of-way rule set is used to characterize lane access rights, intersection turning rules, and traffic light control rules.

[0312] The road topology can be generated from the lane centerline, connecting segments, intersection geometric boundaries and control attributes in the high-precision map. The relationship between nodes and edges is established with lanes as the basic unit, so that each lane connection corresponds to the passability conditions and passability constraints, thereby forming a traffic relationship network that can be used for propagation and tracking.

[0313] After starting with the root event as the starting node, a breadth-first traversal is performed along the traffic relationship network in multiple directions, and the search boundary is limited by combining the physical connectivity basis and right-of-way rules.

[0314] For urban intersections, roundabouts, or highway merging zones, the traversal process can simultaneously cover the downstream direction of the lane where the root event is located, the conflict direction of adjacent lanes, the merging direction, and the merging and merging directions. During the traversal process, road units that are connected to the root event and meet traffic constraints are recorded to determine the search range, which is the spatial set of traffic disturbances that the root event may cause to propagate.

[0315] In one possible implementation, the road topology is transformed into a traffic relationship network that includes a physical connectivity basis and a set of right-of-way rules, including:

[0316] The road topology extracts lane connection relationships, intersection connection relationships, and ramp merging and merging relationships to form the physical connectivity foundation; it also extracts lane access permissions, intersection turning rules, and traffic light control rules to form the right-of-way rule set; and by using lanes as the basic unit and associating the physical connectivity foundation with the right-of-way rule set, a traffic relationship network is constructed.

[0317] In this embodiment, lane connection relationship is used to characterize the passable connection state between adjacent lanes, intersection connection relationship is used to characterize the connection state within the intersection and between the intersection and the connecting road, and ramp merging and merging relationship is used to characterize the entry and exit connection state between the ramp and the main road.

[0318] The physical connectivity foundation consists of the relationships described above and is used to describe spatial accessibility in road networks.

[0319] Lane access rights are used to characterize traffic constraints at the lane level, intersection turning rules are used to characterize legal turning constraints at intersections, traffic light control rules are used to characterize traffic priority and timing constraints controlled by traffic lights, and the right-of-way rule set is composed of the above rules and is used to describe the traffic constraint relationships in the road network.

[0320] In the specific implementation, firstly, based on the road topology in the road map data, the lane centerline, intersection node and ramp connection section are parsed to identify the connection edge between adjacent lanes in the same direction, the connection edge that can be turned into inside the intersection, and the merging edge and exit edge between the main road and the ramp, and write these connection edges into the physical connectivity table.

[0321] Subsequently, road attribute fields, lane identification fields, and traffic control fields are extracted to generate lane access permissions, intersection turning rules, and traffic light control rules, which are then written into the right-of-way rule set table. The traffic relationship network uses lanes as nodes, connecting edges in the physical connectivity foundation as reachable edges, and constraint information in the right-of-way rule set as edge attributes or node attributes. It also unifies and associates the reachability of the same lane under different access permissions and signal states, forming a traversable lane-level network structure.

[0322] To facilitate subsequent breadth-first traversal search, nodes in the network record lane number, road segment, intersection and ramp attributes, while edges record connection type, traffic conditions and turning restrictions. The graph structure is stored in a relational database or in-memory graph. In practical applications, other storage formats can also be used for this graph structure, and this application does not limit it.

[0323] When this traffic relationship network is in operation, it incorporates the spatial connectivity and traffic constraints in the road topology into the same lane-level graph model. This allows root events to be searched along real reachable paths, and the discovery process of candidate traffic participants is constrained by lane connectivity, turning rules, and signal control, thus forming a propagation search basis with clear boundaries and well-defined directions.

[0324] With this implementation method, the reachability relationships and right-of-way constraints between lanes are uniformly mapped to the same network, which can avoid false connectivity caused by relying solely on geometric adjacency, improve the accuracy of subsequent participant searches, and make the propagation path expression more complete in complex intersections, ramp merging and merging, and signal control scenarios.

[0325] S402. Identify vehicles whose motion state changes within the search range as candidate traffic participants.

[0326] For example, the speed, acceleration, heading angle and lateral displacement of vehicle trajectories within the search range are continuously compared to identify vehicles whose motion states change as candidate traffic participants.

[0327] Changes in motion state can manifest as a sudden increase in deceleration, a change in heading, an increase in lateral offset, or a fluctuation in vehicle speed exceeding a threshold. Once a vehicle exhibits such changes within the search range, it is retained as a subsequent verification object. This can distinguish vehicles that have an actual response association with the root event from the surrounding background traffic, providing candidate objects for establishing parent-child propagation relationships.

[0328] S403. Based on the historical movement status of candidate traffic participants before the root event and the road topology, determine the expected trajectory of the candidate traffic participants, and determine the expected trajectory deviation based on the deviation between the expected trajectory and the actual trajectory.

[0329] For example, the expected trajectory deviation is used to characterize the degree of deviation between the undisturbed expected trajectory of a candidate traffic participant based on the historical motion state and road topology before the root event and the actual trajectory after the root event.

[0330] Specifically, when determining the expected trajectory deviation, the system constructs an intention baseline trajectory for each candidate traffic participant. This intention baseline trajectory is generated based on the candidate traffic participant's navigation path planning, lane keeping expectations, or average behavior of the upstream traffic flow in the lane before the root event occurs, representing the vehicle's expected movement path under the influence of no external disturbances.

[0331] After the root event occurs, the deviation envelope between the candidate vehicle's actual trajectory and the intended reference trajectory in the spatiotemporal domain is calculated. The lateral offset increment, longitudinal velocity residual, and heading change rate are extracted and fused with the distance difference, angle difference, and time difference to form the expected trajectory deviation value.

[0332] The system only considers a response valid when the direction of the actual trajectory deviation is consistent with the direction of the root event disturbance's physical action, and the deviation exceeds a reasonable threshold range for intentional autonomous adjustment. Simultaneously, a predictive trajectory smoothing constraint is introduced to filter out non-disturbance motion fluctuations caused by instantaneous sensor noise or local road condition fine-tuning, ensuring that propagation edges connect only to true causal response nodes. This intention deviation perception mechanism effectively isolates the kinematic similarity between vehicle autonomous planning behavior and external disturbance responses, significantly reducing the false positive connection rate in traffic disturbance propagation tree construction.

[0333] S404. Perform spatiotemporal constraint verification and propagation direction consistency verification on candidate traffic participants, and filter out false connections by combining expected trajectory deviation to obtain real candidate participants.

[0334] For example, spatiotemporal constraint verification is used to verify whether candidate traffic participants meet the time delay constraint after the root event at the response time and to confirm that their spatial location falls within the propagation range defined by the road topology.

[0335] The propagation direction consistency check is used to confirm whether the motion deflection direction of the candidate traffic participant is consistent with the propagation direction of the root event disturbance.

[0336] The difference between the actual response time of the candidate traffic participant and the occurrence time of the root event is verified, and it is verified whether the spatial location is located in the lane connected to the root event, the adjacent lane, or the intersection associated area. At the same time, the directional deviation angle between the direction of its movement deflection and the direction of the root event disturbance is calculated. When the deviation angle and the time delay both meet the threshold requirements and the expected trajectory deviation exceeds the preset deviation threshold, it is determined as a real candidate participant.

[0337] In one possible implementation, spatiotemporal constraint verification and propagation direction consistency verification are performed on candidate traffic participants, and false connections are filtered in conjunction with expected trajectory deviations, including:

[0338] The direction angle between the vehicle disturbance direction of the root event and the motion deflection direction of the candidate traffic participants is calculated. The direction angle is then checked for consistency with a preset direction angle threshold to obtain the check result. Based on the check result and the deviation from the expected trajectory, a double filter is performed to remove candidate traffic participants with false connections.

[0339] In this embodiment, the directional deviation angle is used to characterize the angle difference between the driving direction of the vehicle disturbance corresponding to the root event and the movement deflection direction of the candidate traffic participant. The preset directional deviation angle threshold is used to limit the judgment boundary of the consistency of the propagation direction, and the verification result is used to indicate whether the directional deviation angle meets the judgment boundary.

[0340] Expected trajectory deviation is used to characterize the degree of deviation between the expected motion path of a candidate traffic participant under road topology constraints and the actual deflection trajectory. False connection refers to the situation where a candidate traffic participant falls within the search range in time and space, but there is no real propagation association between its motion response and the root event.

[0341] In practical implementation, after obtaining the initial disturbance direction of the root event, the system determines the vehicle disturbance direction based on the vehicle's heading, lateral offset change, or instantaneous velocity vector associated with the root event; then, it performs tangent vector fitting on the trajectory points of the candidate traffic participants at several consecutive sampling times to obtain the motion deflection direction, and uses the vector angle calculation method to obtain the direction deflection angle.

[0342] When the directional deviation angle is less than or equal to the preset directional deviation angle threshold, the output is a verification result that meets the consistency condition; when the directional deviation angle is greater than the threshold, the output is a verification result that does not meet the consistency condition.

[0343] Subsequently, the verification results are jointly judged with the expected trajectory deviation calculated based on historical motion status, lane connectivity, and right-of-way constraints. If the propagation correlation represented by either of the two is insufficient, the corresponding candidate traffic participants are marked as false connections and eliminated, thus retaining only the real candidate participants whose propagation direction is consistent and whose trajectory deviation meets expectations.

[0344] This processing method first constrains the propagation direction by verifying the consistency of the directional deviation angle, and then performs a second screening based on the expected trajectory deviation. This ensures that the determination of candidate traffic participants simultaneously meets the requirements of directional association and trajectory association, which can reduce the probability of false connectivity introduced by relying solely on a single spatiotemporal proximity relationship, and provide more reliable node input for the establishment of parent-child relationships in the subsequent traffic disturbance propagation tree.

[0345] By adopting this specific implementation, it is possible to identify candidate traffic participants with real propagation relationships to the root event in complex traffic disturbance scenarios, reduce the inclusion of false connections in the propagation tree, improve the accuracy of propagation relationship construction and the completeness of scene slices, and enhance the structural credibility of subsequent difficult scenario screening results.

[0346] S405. Based on the actual candidate participants, calculate the corresponding response strength and propagation edge risk to determine the node inclusion level, and establish a parent-child propagation relationship between the actual candidate participants as child nodes and their corresponding parent nodes according to the node inclusion level.

[0347] For example, response intensity is used to quantify the magnitude of the change in motion state of candidate traffic participants after being disturbed; propagation edge hazard is used to quantify the degree of conflict risk accumulation on the propagation link; node inclusion level is used to determine whether and at what level real candidate participants join the traffic disturbance propagation tree; parent-child propagation relationship is used to characterize the hierarchical propagation connection between the root event and subsequent response participants.

[0348] The response intensity is calculated based on the velocity change rate, braking intensity, lateral displacement increment, and trajectory mutation amplitude of the actual candidate participants. The propagation edge hazard is calculated by combining the spatial distance between the candidate and its corresponding parent node, relative speed, road conflict relationship, and risk accumulation of nodes already included in the propagation chain.

[0349] The response strength and the risk level of the propagation edge are jointly mapped to the node inclusion level, thereby allowing real candidate participants to establish parent-child propagation relationships with their corresponding parent nodes, forming a traffic disturbance propagation tree that reflects the actual disturbance diffusion path.

[0350] In one possible implementation, based on the node inclusion level, real candidate participants are designated as child nodes and their corresponding parent nodes to establish parent-child propagation relationships, including:

[0351] When the response strength meets the effective threshold and the propagation edge danger is higher than the first threshold and lower than the second threshold, the decision node is classified as a leaf node, and the real candidate participant is used as a leaf node to establish a parent-child propagation relationship with the corresponding parent node; when the response strength meets the effective threshold and the propagation edge danger is higher than the second threshold, the decision node is classified as an intermediate node, and the real candidate participant is used as an intermediate node to establish a parent-child propagation relationship with the corresponding parent node.

[0352] In this embodiment, leaf nodes are used to represent real candidate participants whose response strength has reached the effective threshold and whose propagation edge danger is at a low level, and whose propagation terminates at that participant. In the traffic disturbance propagation tree, leaf nodes are only retained as direct parent-child associations with the corresponding parent node and are no longer extended for further search.

[0353] Intermediate nodes are used to represent real candidate participants whose response strength has reached the effective threshold and whose propagation edge danger is at a high level, and who may still continue to propagate backward. In the traffic disturbance propagation tree, they not only establish a parent-child propagation relationship with the corresponding parent node, but also serve as a new starting point for subsequent propagation tracking.

[0354] For example, after obtaining real candidate participants, the system calculates the corresponding response intensity and propagation edge hazard based on the perturbation type of the root event, the spatiotemporal correlation between the participant and the corresponding parent node, and the connectivity direction on the road topology.

[0355] The response intensity can be obtained by combining the velocity change rate, acceleration change amplitude, lateral displacement response amplitude, or braking response characteristics, and compared with a pre-configured effective threshold. Only when the value reaches the preset response level is it allowed to enter the node for inclusion in the judgment.

[0356] The propagation edge hazard is used to characterize the risk persistence of the root event propagating to the candidate participant. It can be calculated by combining the intensity of adjacent lane conflict, the degree of path intersection, the change in the distance between vehicles, the lateral approach trend, and collision risk parameters, and compared with the first threshold and the second threshold respectively.

[0357] For candidate participants whose risk level is higher than the first threshold but lower than the second threshold, the system marks them as leaf nodes and records them as the terminal branch of the root event in the propagation tree; for candidate participants whose risk level is higher than the second threshold, the system marks them as intermediate nodes and retains these nodes as the starting point for further propagation.

[0358] The threshold configuration can be obtained based on historical traffic disturbance samples. In practical applications, the threshold combination can also be adjusted according to different road types, speed ranges or scenario categories. This application does not limit this.

[0359] By marking low-risk participants whose responses have terminated as leaf nodes and high-risk participants whose responses can still propagate as intermediate nodes, the traffic disturbance propagation tree can distinguish between propagation termination points and propagation continuation points, making the response chain associated with the root event more complete.

[0360] The above method enables the scene boundary to correspond synchronously with the participant hierarchy, facilitating the subsequent distinction between the complete participant set and the core participant set, and improving the accuracy of the structured representation of complex perturbation scenes.

[0361] In another possible implementation, it also includes:

[0362] When the response intensity is below the effective threshold, the spatiotemporal constraints are not valid, and / or the propagation edge danger is below the first threshold, the node inclusion level is determined to be non-inclusion, and the parent-child propagation relationship between the real candidate participants and the root event is not established.

[0363] In this embodiment, the effective threshold is a pre-set lower limit of response strength, used to distinguish between participants who can form an effective propagation relationship and those with insufficient response. A node inclusion level of "not included" indicates the processing result of candidate participants whose response strength is insufficient and cannot form an effective propagation relationship; their corresponding nodes do not participate in the generation of parent-child propagation relationships and subsequent propagation chain expansion.

[0364] In practice, after identifying real candidate participants, the system calculates the response intensity based on the participant's trajectory change, velocity change, and spatiotemporal synchronization relationship relative to the root event, and compares the calculation result with the effective threshold.

[0365] When the response intensity is less than the effective threshold, the node inclusion level is directly determined to be excluded. The system will no longer generate a parent node association pointing to the root event for the real candidate participant, nor will it generate a child node record pointing to the participant from the root event. As a result, the participant is only retained in the candidate retrieval results and will not enter the structured tree construction process of the traffic disturbance propagation tree.

[0366] Furthermore, when the spatiotemporal constraints are not met, and / or the danger level of the propagation edge is lower than the first threshold, the system will also determine the node's inclusion level as not included, and it will not enter the structured tree construction process of the traffic disturbance propagation tree.

[0367] The above determination can be completed by the propagation relationship construction unit on the server side. This unit can be implemented using a rule engine or a threshold determination unit. The threshold parameter can be calibrated offline based on historical scene samples and called during online processing. In practical applications, other models of this component can also be selected, and this application does not limit this.

[0368] This approach excludes real candidate participants with insufficient responses from the parent-child relationship during propagation tree construction, ensuring that the propagation chain retains only nodes with valid response evidence, thereby maintaining the coherence and causal consistency of the propagation structure. By identifying low-response nodes as excluded, it reduces the interference of invalid connections on propagation path determination and makes subsequent scene boundary determination and participant set division more closely resemble the actual traffic disturbance diffusion process.

[0369] After adopting the above implementation method, the node connection relationship in the traffic disturbance propagation tree is more accurate, and low-intensity response participants will not be mistakenly added to the propagation chain. Therefore, the causal purity and structural stability of the propagation tree can be improved, and the reliability of the results when using the propagation tree for scene interception and hard case screening can be improved.

[0370] In conjunction with the above embodiments, specifically, the first... The traffic disturbance propagation tree is represented as follows:

[0371]

[0372] in, Indicates the first A traffic disturbance propagation tree, Represents the TPST node set. This represents the set of parent-child propagation edges. This represents the root event node.

[0373] For the parent node and candidate child nodes First, perform time constraint verification:

[0374]

[0375] in, Indicates the time when the parent event occurred. Indicates the time when a candidate vehicle's valid response ends. Indicates the maximum response latency corresponding to the parent event type.

[0376] After satisfying the time constraints and road spatial topology constraints, the response intensity of candidate traffic participants is calculated:

[0377]

[0378] in, and These represent the normalized deceleration, heading change, lateral drift or lane change, speed decrease, and the intensity of planned trajectory correction or safety function triggering, respectively. and The weights are non-negative and their sum is 1.

[0379] For candidate vehicles that meet the time, space, and response conditions, calculate the propagation edge hazard. :

[0380]

[0381] in, Represents the longitudinal risk function. Represents the horizontal risk function; For collision time, For the headway, For deceleration; For post-intrusion time, For the angle of conflict, This refers to the time of lateral intrusion. For the corresponding weights.

[0382] The system considers time constraints, road space topology constraints, actual vehicle response intensity, and propagation edge hazard to determine whether candidate nodes should be added to TPST and whether to continue expanding.

[0383]

[0384] in, and These indicate whether the time constraint and the road space topology constraint are valid, respectively. This indicates the actual response strength of the candidate vehicle; Indicates the minimum effective response strength; and These represent the low-risk threshold (first threshold) and the high-risk threshold (second threshold), respectively. This represents the inclusion rank value for nodes between parent node u and candidate child node v in the traffic disturbance propagation tree. Candidate nodes are not added to TPST. When added as a leaf node to TPST, It is added to TPST as an intermediate node and continues to expand.

[0385] The autonomous driving difficult scenario mining method provided in this application constructs a traffic relationship network integrating physical connectivity and right-of-way rules, and defines the search range through breadth-first traversal. This allows for accurate matching of the actual propagation space of traffic disturbances, ensuring the spatial relevance of candidate objects. Identifying vehicles with changing motion states within the search range allows for filtering out potential responding vehicles from background traffic flow, narrowing the scope of subsequent verification. Determining the expected trajectory deviation corresponding to the root event provides a unified benchmark for filtering false connections. By combining spatiotemporal constraint verification, propagation direction consistency verification, and expected trajectory deviation filtering, irrelevant vehicles can be eliminated, resulting in genuine candidate participants with real response associations. Calculating response intensity and propagation edge hazard and classifying node inclusion levels establishes a clear hierarchical parent-child propagation relationship, fully reconstructing the actual diffusion link of traffic disturbances.

[0386] Figure 5 A flowchart illustrating the autonomous driving difficult scenario mining method provided in this application embodiment. Figure 3 ,like Figure 5 As shown, this embodiment, based on the above embodiments, provides supplementary explanations of the subsequent process for obtaining the set of difficult autonomous driving scenarios, including:

[0387] S501. Divide the road range corresponding to the traffic disturbance propagation tree into multiple spatiotemporal grid units.

[0388] For example, spatiotemporal grid cells are used to divide the corresponding road area into multiple local processing units so as to perform partitioned statistics and index management of the scene in the local area.

[0389] In practice, the road space covered by the traffic disturbance propagation tree can be divided into multiple basic grids according to lane centerlines, intersection areas, or road mileage, and a time window can be added to each grid to form a spatiotemporal grid unit with both spatiotemporal dimensions.

[0390] S502. Perform density clustering on multiple autonomous driving challenge scenarios within each spatiotemporal grid unit, and adjust the grid boundary of each spatiotemporal grid unit based on the clustering results to generate the corresponding scenario index space.

[0391] For example, density clustering is used to identify high-density clustered areas and sparse areas based on the spatial and temporal distribution of autonomous driving challenge scenarios within each unit.

[0392] Specifically, multiple related autonomous driving challenge scenarios can be mapped to corresponding grids, and feature vectors can be constructed according to the start and end times of the scenarios, spatial locations, participant coverage, and propagation chain structure. Density clustering can then be used to identify scenario clusters.

[0393] For densely populated areas with edges spanning multiple units, the grid boundary is expanded outward based on the outer circumference of the clusters. For sparsely populated areas, the boundary is shrunk inward to match the unit range with the scene distribution. The adjusted grid structure together constitutes the scene index space.

[0394] S503. Based on the scene index space, the traffic disturbance propagation tree is stored in a spatiotemporal fragmentation and incrementally updated to obtain a spatiotemporally indexed traffic disturbance propagation tree.

[0395] For example, based on the division of the scene index space, the nodes, edge relationships, and scene fragments of the traffic disturbance propagation tree can be stored in segments.

[0396] The system is deployed based on a distributed processing architecture with spatiotemporal partitioning and incremental state caching. The operating roads are divided into spatially non-overlapping local grid units, and an incremental state machine is built in memory to record the state of traffic disturbance propagation tree nodes, the active state of propagation edges, and the count of covered units in each grid unit in real time.

[0397] New data streams are routed to the corresponding grids according to spatial coordinates, triggering only topology updates and state refreshes for the affected local subgraphs. The global dynamic incremental value assessment adopts an incremental difference mechanism, using the incremental difference (Diff) algorithm to calculate and refresh only the rate of change of coverage gaps.

[0398] The system employs a multi-level caching architecture. A hot cache stores frequently accessed propagation topology features and scene embedding vectors; a warm cache saves intermediate local evaluation results; and a cold cache archives historical full snapshots. After receiving the incremental differential signal, the central value evaluator uses an incremental sorting algorithm to quickly rearrange the candidate queue, avoiding full recalculation. When a new difficult scenario arrives, only the corresponding local index shards are updated through insertion, merging, or reallocation, without needing to rebuild the entire propagation tree, ultimately forming a spatiotemporally indexed traffic disturbance propagation tree.

[0399] This architecture replaces global synchronous computation with spatial decoupling and state localization mechanisms, so that new data only triggers topology reconstruction of local graph segments, greatly reducing redundant computational overhead and realizing a high-throughput, low-latency real-time difficult example mining pipeline. It can maintain the consistency between the propagation tree and the scene distribution, and improve the processing efficiency when storing and maintaining large-scale difficult example scenes.

[0400] The autonomous driving difficult scenario mining method provided in this application divides the road area corresponding to the traffic disturbance propagation tree into multiple spatiotemporal grid units, providing a basic processing unit for the partitioning statistics and index management of local scenarios. Based on the autonomous driving difficult scenario within the unit, density clustering is performed and the grid boundaries are dynamically adjusted, generating a scenario index space that accurately adapts to the actual distribution of the scenario. By relying on the scenario index space for spatiotemporal sharding storage and incremental updates of the traffic disturbance propagation tree, the update granularity can be reduced, improving the processing efficiency of large-scale scenario import and propagation tree maintenance.

[0401] Figure 6 This is a schematic diagram of the structure of the autonomous driving difficult scenario discovery device provided in this embodiment of the application. The device in this embodiment can be in the form of software and / or hardware. Figure 6 As shown in the embodiment of this application, the autonomous driving difficult scenario mining device 600 includes: an acquisition module 601 and a processing module 602.

[0402] The acquisition module 601 is used to acquire trajectory data of traffic participants and road map data, and perform spatiotemporal alignment processing to obtain standardized traffic flow data;

[0403] Processing module 602 is used to obtain motion risk characteristics based on the road topology of standardized traffic flow data and road map data;

[0404] The processing module 602 is also used to generate atomic traffic events based on the characteristics of motion risks, and to filter out root events from the atomic traffic events;

[0405] The processing module 602 is also used to search for candidate traffic participants along the road topology starting from the root event, and establish parent-child propagation relationship after verifying the spatiotemporal constraints and response intensity, so as to construct a traffic disturbance propagation tree;

[0406] The processing module 602 is also used to determine the scene time boundary based on the traffic disturbance propagation tree, and extract scene data from standardized traffic flow data and road map data based on the scene time boundary, and distinguish between the complete participant set and the core participant set in combination with the traffic disturbance propagation tree;

[0407] The processing module 602 is also used to filter out a set of difficult autonomous driving scenarios based on the traffic disturbance propagation tree, the complete set of participants and the core set of participants.

[0408] In one possible implementation, the acquisition module 601 is further configured to:

[0409] The trajectory data of traffic participants are synchronized with the time base to obtain synchronized trajectory data;

[0410] Based on the map coordinate system of the road map data, the synchronized trajectory data is subjected to spatial coordinate transformation, followed by noise smoothing and target identity continuity verification to obtain standardized traffic flow data.

[0411] In one possible implementation, the processing module 602 is further configured to:

[0412] By matching the vehicle trajectory points of standardized traffic flow data to the lane and intersection areas corresponding to the road topology, lane-level positioning data is obtained.

[0413] Based on lane-level positioning data, longitudinal motion parameters, lateral interaction parameters, and collision risk parameters of the vehicle are calculated and integrated to obtain motion risk characteristics.

[0414] In one possible implementation, the processing module 602 is further configured to:

[0415] Based on the characteristics of sports risks, match the corresponding preset time trigger threshold;

[0416] Search for consecutive intervals exceeding a preset time trigger threshold in the time series corresponding to the motion risk characteristics, merge candidate intervals that are adjacent in time and have the same traffic participants, and generate atomic traffic events based on the candidate intervals.

[0417] Based on the atomic traffic events, the corresponding propagation potential score is determined, and atomic traffic events with propagation potential scores higher than the preset propagation threshold are selected as root events.

[0418] In one possible implementation, the processing module 602 is further configured to:

[0419] The road topology is transformed into a traffic relationship network that includes the physical connectivity foundation and the set of right-of-way rules. Starting from the root event, a breadth-first traversal is performed along the traffic relationship network in multiple directions to determine the search range.

[0420] Identify vehicles whose motion state changes within the search range as candidate traffic participants.

[0421] In one possible implementation, the processing module 602 is further configured to:

[0422] Based on the historical movement status of candidate traffic participants before the root event and the road topology, the expected trajectory of the candidate traffic participants is determined, and the expected trajectory deviation is determined based on the deviation between the expected trajectory and the actual trajectory.

[0423] Spatiotemporal constraint verification and propagation direction consistency verification are performed on candidate traffic participants, and false connections are filtered out by combining expected trajectory deviations to obtain real candidate participants;

[0424] Based on the actual candidate participants, calculate the corresponding response strength and propagation edge risk to determine the node inclusion level. Based on the node inclusion level, establish a parent-child propagation relationship between the actual candidate participants as child nodes and their corresponding parent nodes.

[0425] In one possible implementation, the processing module 602 is further configured to:

[0426] When the response strength meets the effective threshold and the propagation edge danger is both higher than the first threshold and lower than the second threshold, the node is determined to be a leaf node, and the real candidate participants are used as leaf nodes to establish a parent-child propagation relationship with the corresponding parent nodes.

[0427] When the response strength meets the effective threshold and the propagation edge danger is higher than the second threshold, the node is determined to be included as an intermediate node, and the real candidate participants are used as intermediate nodes to establish a parent-child propagation relationship with the corresponding parent node.

[0428] In one possible implementation, the processing module 602 is further configured to:

[0429] When the response intensity is below the effective threshold, the spatiotemporal constraints are not valid, and / or the propagation edge danger is below the first threshold, the node inclusion level is determined to be non-inclusion, and the parent-child propagation relationship between the real candidate participants and the root event is not established.

[0430] In one possible implementation, the processing module 602 is further configured to:

[0431] The difference between the start time and the leading time of the root event in the traffic disturbance propagation tree is used as the scene start time, and the sum of the end time and the delay time of the atomic traffic event corresponding to the latest response node in the traffic disturbance propagation tree is used as the scene end time.

[0432] Determine the scene time boundaries based on the scene start time and scene end time.

[0433] In one possible implementation, the processing module 602 is further configured to:

[0434] Based on the scene time boundary, the scene data containing road topology and traffic flow information is extracted from the standardized traffic flow data and the road spatial range corresponding to the road map data.

[0435] Based on traffic flow information and the node coverage of the traffic disturbance propagation tree, the multiple real candidate participants contained in the traffic disturbance propagation tree, as well as the direct parameters of the root event and / or atomic traffic events, are defined as the core participant set, and multiple traffic participants within the scene data are defined as the complete participant set.

[0436] In one possible implementation, the processing module 602 is further configured to:

[0437] Extract lane connection relationships, intersection connectivity relationships, and ramp merging and merging relationships from the road topology to form the basis of physical connectivity;

[0438] Extract lane access permissions, intersection turning rules, and traffic light control rules from the road topology to form a set of right-of-way rules;

[0439] By using lanes as the basic unit and linking them with the physical connectivity foundation and the set of right-of-way rules, a traffic relationship network is constructed.

[0440] In one possible implementation, the processing module 602 is further configured to:

[0441] Calculate the directional deflection angle between the vehicle disturbance direction of the root event and the motion deflection direction of the candidate traffic participants, and verify the consistency of the directional deflection angle with the preset directional deflection angle threshold to obtain the verification result.

[0442] Based on the deviation between the verification results and the expected trajectory, a double filtering process is performed to eliminate candidate traffic participants with false connections.

[0443] In one possible implementation, the processing module 602 is further configured to:

[0444] Based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, multiple initial candidate scenarios are generated;

[0445] Each initial candidate scenario is compared with multiple other initial candidate scenarios in a multi-dimensional structure to obtain multiple structure comparison difference results. Initial candidate scenarios with structure comparison difference results greater than a preset threshold are retained to form a set of difficult autonomous driving scenarios.

[0446] In one possible implementation, the processing module 602 is further configured to:

[0447] Based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, multiple initial candidate scenarios are generated;

[0448] Based on the preset multi-dimensional segmentation rules, multiple difficult scenarios already included in the pre-built scenario library are divided into multiple coverage units, and the scenario gap ratio of each coverage unit is calculated.

[0449] Multiple candidate scenarios are filtered based on the scenario gap ratio to obtain a set of challenging autonomous driving scenarios.

[0450] In one possible implementation, the processing module 602 is further configured to:

[0451] The road area corresponding to the traffic disturbance propagation tree is divided into multiple spatiotemporal grid units;

[0452] Density clustering is performed on multiple autonomous driving challenge scenarios within each spatiotemporal grid cell, and the grid boundary of each spatiotemporal grid cell is adjusted based on the clustering results to generate the corresponding scenario index space.

[0453] Based on the scene index space, the traffic disturbance propagation tree is stored in a spatiotemporal fragment and incrementally updated to obtain a spatiotemporally indexed traffic disturbance propagation tree.

[0454] The autonomous driving difficult scenario mining device 600 provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0455] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 As shown, the electronic device 700 provided in this embodiment includes at least one processor 701 and a memory 702. Optionally, the electronic device 700 further includes a communication component 703. The processor 701, memory 702, and communication component 703 are connected via a bus.

[0456] In a specific implementation, at least one processor 701 executes computer execution instructions stored in memory 702, causing at least one processor 701 to perform the above-described method.

[0457] The specific implementation process of processor 701 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0458] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0459] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0460] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0461] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0462] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0463] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0464] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0465] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0466] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0467] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0468] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0469] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0470] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for mining difficult scenarios in autonomous driving, characterized in that, include: Acquire trajectory data of traffic participants and road map data, and perform spatiotemporal alignment processing to obtain standardized traffic flow data; Based on the road topology of the standardized traffic flow data and the road map data, motion risk characteristics are obtained; Based on the motion risk characteristics, atomic traffic events are generated, and root events are obtained by filtering from the atomic traffic events; Starting with the root event as the starting node, candidate traffic participants are searched along the road topology. After verifying the spatiotemporal constraints and response intensity, parent-child propagation relationships are established to construct a traffic disturbance propagation tree. Based on the traffic disturbance propagation tree, the scene time boundary is determined. Based on the scene time boundary, scene data is extracted from the standardized traffic flow data and the road map data, and combined with the traffic disturbance propagation tree, the complete participant set and the core participant set are distinguished. Based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, a set of challenging scenarios for autonomous driving is obtained.

2. The method according to claim 1, characterized in that, The spatiotemporal alignment process yields standardized traffic flow data, including: The trajectory data of the traffic participants are synchronized with a time reference to obtain synchronized trajectory data; Based on the map coordinate system of the road map data, the synchronized trajectory data is subjected to spatial coordinate transformation, followed by noise smoothing and target identity continuity verification to obtain the standardized traffic flow data.

3. The method according to claim 2, characterized in that, The process of obtaining motion risk characteristics based on the road topology of the standardized traffic flow data and the road map data includes: The vehicle trajectory points of the standardized traffic flow data are matched to the lane and intersection areas corresponding to the road topology to obtain lane-level positioning data. Based on the lane-level positioning data, the vehicle's longitudinal motion parameters, lateral interaction parameters, and collision risk parameters are calculated and integrated to obtain the motion risk characteristics.

4. The method according to claim 1, characterized in that, The process of generating atomic traffic events based on the motion risk characteristics and filtering out root events from these atomic traffic events includes: Based on the described motion risk characteristics, a corresponding preset time trigger threshold is matched; Search for consecutive intervals exceeding the preset time trigger threshold in the time series corresponding to the motion risk characteristics, merge candidate intervals that are adjacent in time and have the same traffic participants, and generate the atomic traffic event based on the candidate intervals; Based on the atomic traffic events, a corresponding propagation potential score is determined, and atomic traffic events with propagation potential scores higher than a preset propagation threshold are selected as the root events.

5. The method according to claim 1, characterized in that, The process of searching for candidate traffic participants along the road topology starting from the root event includes: The road topology is converted into a traffic relationship network that includes a physical connectivity foundation and a set of right-of-way rules. Starting from the root event, a breadth-first traversal is performed along the traffic relationship network in multiple directions to determine the search range. Vehicles whose motion state changes are identified within the search range as candidate traffic participants.

6. The method of claim 5, wherein, The establishment of the parent-child propagation relationship after spatiotemporal constraints and response strength verification includes: Based on the historical movement state of the candidate traffic participants before the root event and the road topology, the expected trajectory of the candidate traffic participants is determined, and the expected trajectory deviation is determined based on the deviation between the expected trajectory and the actual trajectory. Spatiotemporal constraint verification and propagation direction consistency verification are performed on the candidate traffic participants, and false connections are filtered out by combining the expected trajectory deviation to obtain the real candidate participants; Based on the actual candidate participants, the corresponding response strength and propagation edge risk are calculated to determine the node inclusion level. Based on the node inclusion level, the actual candidate participants are established as child nodes and their corresponding parent nodes to establish the parent-child propagation relationship.

7. The method of claim 6, wherein, The step of establishing the parent-child propagation relationship between the real candidate participants as child nodes and their corresponding parent nodes based on the node inclusion level includes: When the response strength meets the effective threshold and the propagation edge danger is both higher than the first threshold and lower than the second threshold, the node is determined to be a leaf node, and the real candidate participant is used as a leaf node to establish a parent-child propagation relationship with the corresponding parent node. When the response strength meets the effective threshold and the propagation edge danger is higher than the second threshold, the node is determined to be an intermediate node, and the real candidate participant is used as an intermediate node to establish a parent-child propagation relationship with the corresponding parent node.

8. The method of claim 7, wherein, Also includes: When the response intensity is lower than the effective threshold, the spatiotemporal constraint is not valid, and / or the propagation edge danger is lower than the first threshold, the node inclusion level is determined to be non-inclusion, and the parent-child propagation relationship between the real candidate participant and the corresponding parent node is not established.

9. The method of claim 7, wherein, Determining the scene time boundary based on the traffic disturbance propagation tree includes: The difference between the start time and the leading time of the root event in the traffic disturbance propagation tree is used as the scene start time, and the sum of the end time and the delay time of the atomic traffic event corresponding to the latest response node in the traffic disturbance propagation tree is used as the scene end time. The time boundary of the scene is determined based on the start time and end time of the scene.

10. The method of claim 9, wherein, The step of extracting scene data from the standardized traffic flow data and the road map data based on the scene time boundary and distinguishing between the complete participant set and the core participant set by combining the traffic disturbance propagation tree includes: Based on the scene time boundary and the road spatial range corresponding to the standardized traffic flow data and the road map data, scene data containing the road topology and traffic flow information is extracted; Based on the traffic flow information and the node coverage of the traffic disturbance propagation tree, the multiple real candidate participants included in the traffic disturbance propagation tree, as well as the direct participants of the root event and / or the atomic traffic event, are defined as the core participant set, and the multiple traffic participants within the scene data are defined as the complete participant set.

11. The method according to claim 5, characterized in that, The process of converting the road topology into a traffic relationship network that includes a physical connectivity foundation and a set of right-of-way rules includes: The lane connection relationships, intersection connectivity relationships, and ramp merging and merging relationships are extracted from the road topology to form the basis of the physical connectivity. The right-of-way rule set is constructed by extracting lane access permissions, intersection turning rules, and traffic light control rules from the road topology. Using lanes as the basic unit, the traffic relationship network is constructed by associating the physical connectivity basis with the right-of-way rule set.

12. The method according to claim 6, characterized in that, The step of performing spatiotemporal constraint verification and propagation direction consistency verification on the candidate traffic participants, combined with filtering false connections based on the expected trajectory deviation, includes: Calculate the directional angle between the vehicle disturbance driving direction of the root event and the motion deflection direction of the candidate traffic participant, and verify the consistency of the directional angle with a preset directional angle threshold to obtain the verification result. Based on the verification results and the deviation from the expected trajectory, a dual filtering process is performed to eliminate candidate traffic participants with false connections.

13. The method according to claim 1, characterized in that, The set of challenging autonomous driving scenarios is obtained by filtering based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, including: Based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, multiple initial candidate scenarios are generated. Each initial candidate scene is compared with the other initial candidate scenes in a multi-dimensional structure to obtain multiple structure comparison difference results. The initial candidate scenes with structure comparison difference results greater than a preset threshold are retained to form the set of autonomous driving difficult scenarios.

14. The method according to claim 1, characterized in that, The set of challenging autonomous driving scenarios is obtained by filtering based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, including: Based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants, multiple initial candidate scenarios are generated. According to the preset multi-dimensional division rules, multiple difficult scenarios that have been included in the pre-built scenario library are divided into multiple coverage units, and the scenario gap ratio of each coverage unit is calculated. The set of challenging autonomous driving scenarios is obtained by filtering multiple candidate scenarios based on the scenario gap ratio.

15. The method according to claim 1, characterized in that, Also includes: The road area corresponding to the traffic disturbance propagation tree is divided into multiple spatiotemporal grid units; Density clustering is performed on multiple autonomous driving challenge scenarios within each spatiotemporal grid unit, and the grid boundary of each spatiotemporal grid unit is adjusted based on the clustering results to generate a corresponding scene index space; Based on the scenario index space, the traffic disturbance propagation tree is spatiotemporally sharded and incrementally updated to obtain a spatiotemporally indexed traffic disturbance propagation tree.

16. A device for identifying challenging scenarios in autonomous driving, characterized in that, include: The acquisition module is used to acquire trajectory data of traffic participants and road map data, and perform spatiotemporal alignment processing to obtain standardized traffic flow data; The processing module is used to obtain motion risk characteristics based on the road topology of the standardized traffic flow data and the road map data; The processing module is also used to generate atomic traffic events based on the motion risk characteristics, and to filter out root events from the atomic traffic events; The processing module is also used to search for candidate traffic participants along the road topology with the root event as the starting node, and establish parent-child propagation relationship after verifying spatiotemporal constraints and response intensity, so as to construct a traffic disturbance propagation tree; The processing module is further configured to determine the scene time boundary based on the traffic disturbance propagation tree, and extract scene data from the standardized traffic flow data and the road map data based on the scene time boundary, and distinguish between the complete participant set and the core participant set in combination with the traffic disturbance propagation tree; The processing module is further configured to filter and obtain a set of challenging autonomous driving scenarios based on the traffic disturbance propagation tree, the complete set of participants, and the core set of participants.

17. An electronic device, characterized in that, include: Memory, processor; The memory stores instructions that the computer executes; The processor executes computer execution instructions stored in memory, causing the processor to perform the method as described in any one of claims 1-15.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-15.

Images