Automatic driving generalization test scene construction method based on accident report

Abstract

The invention relates to an automatic driving generalization test scene construction method based on an accident report. According to the technology, an accident scene is preliminarily abstracted into three layers according to layer-based scene definition: a road network and traffic guide object, an environmental condition and a dynamic object. And then, information related to accident influence factors is extracted from the accident report systematically by using LLM, and the information is organized layer by layer. Then, constraints are established for the pre-accident driving behavior of the traffic participant, and a constraint solver is used to generate a trajectory conforming to the specified constraints in the accident report. And then, inputting a track generated by the extracted environment and road information into the driving simulator, and reconstructing a traffic accident scene. And finally, in an evaluation stage, the generated simulation scene is converted into test cases containing a test oracle, and the test cases are subjected to an automatic driving simulation test. The objective of the invention is to solve the problem of reappearing accident scenes on different roads so as to help an automatic driving system employee to carry out testing work in accident scenes of different road structures.

Description

Technical Field

[0001] The present invention belongs to the field of safety testing of automated driving systems (ADS), and is particularly suitable for the field of generalized test scenario construction for automated driving. Its purpose is to provide guidance for automated driving system practitioners to carry out testing work in accident scenarios with different road structures. It is a technology that automatically generates general road scenarios from accident reports. Background Art

[0002] The advent of autonomous driving technology has ushered in a new era of transportation, promising improved safety, efficiency, and convenience. However, the occurrence of accidents involving autonomous vehicles, including those resulting in fatalities, demonstrates that autonomous driving is not currently as safe as it is advertised to be. In many cases, these crashes can be attributed to defective software, highlighting the urgent need for enhanced methods for testing autonomous driving software.

[0003] As one of the key quality assurance technologies, ADS testing has received widespread attention from academia and industry. The key testing technologies of ADS can be divided into two categories: road testing and simulation testing. Road testing involves testing specific driving scenarios in a closed autonomous vehicle proving ground or monitoring autonomous vehicles in real traffic, but this method requires a long time and a lot of resources. In addition, obtaining a variety of key test data, including a wide range of actual usage scenarios, poses a major challenge to testers. Therefore, high-fidelity simulation testing methods have become a necessity for the development and verification of autonomous driving technology. ADS testing is performed on simulation platforms such as LGSVL, CARLA, etc. An important aspect of ADS simulation testing is to identify and construct key scenarios that may lead to accidents.

[0004] Software engineering researchers have proposed using real-life accident cases to generate key test scenarios, as real-life crashes pose significant challenges for autonomous vehicles. Therefore, recent research has focused on scene reconstruction from different sources of driving data. This includes generating basic scenarios using text descriptions, accident sketches, sensor data, and video recordings. Compared with sensor data and video recordings of key cases, text descriptions of collisions are easier to obtain and richer. Accident report analysis methods provide more comprehensive insights into collisions, including complex details such as weather, lighting, and road conditions, which are difficult to represent visually. However, current scene reconstruction methods based on accident reports have limited accuracy in extracting information. In addition, current methods can only reproduce accident scenarios on the same road structure described in the accident report. Since the road structure in simulation may be different from that described in the accident report, it limits the application to simulation-based ADS testing. Reproducing accident scenarios on different roads is a challenging task.

[0005] To bridge this gap, we propose an automatic and general method to reconstruct accident scenarios on different road structures from accident reports. The main goal of this method is to automatically reconstruct accident scenarios from accident reports on different roads in a simulated environment, which are then used as initial seeds for ADS testing. The method first utilizes a carefully designed language prompt pattern to guide the LLM to accurately extract the environment, road, and object motion information. Based on the extracted information and the road reconstruction scenario, the method formulates constraints for the accident scenario. These constraints are solved by a constraint solver to generate accident trajectories of the accident participants and finally reconstruct the accident scenario. Therefore, the generated scenario can be executed on the desired mapping. In addition, to use these scenarios for ADS testing, the method converts them into test scenarios by identifying the ego vehicle (i.e., the vehicle controlled by the ADS under test) and the NPC (non-player character) vehicle. The NPC vehicle will follow the calculated trajectory, while the ego vehicle is controlled by the ADS instead of following a predetermined trajectory. Summary of the invention

[0006] The present invention provides a technology for automatically generating common road scenarios from accident reports to effectively solve the current problem of unknown strength requirements for new version software mutation testing, thereby helping testers to automatically generate common road scenarios and carry out testing work in accident scenarios with different road structures.

[0007] To achieve the above objectives, the present invention proposes a method for constructing a generalized test scenario for autonomous driving based on accident reports, characterized in that, according to a layer-based scenario definition, the accident scenario information is preliminarily abstracted into three layers: road network and traffic guidance objects, environmental conditions, and dynamic objects. Then, a large language model (LLM) is used to systematically extract information about accident influencing factors from the accident report and organize the information layer by layer. Subsequently, constraints are established on the pre-accident driving behavior of traffic participants, and a constraint solver is used to generate trajectories that meet the constraints specified in the accident report. Next, the trajectory generated by the extracted environmental and road information is input into a driving simulator to reconstruct the accident scenario. Finally, in the evaluation phase, the generated simulation scenario is converted into test cases containing a test oracle, and these test cases are input into an autonomous driving system (ADS). The method checks whether the ADS under test successfully reaches the final expected position without encountering any crashes. The method comprises the following steps:

[0008] 1) Information Extraction; To abstractly represent the accident information and organize it into a semantic structure that can be understood by LLM, the extracted information is presented in the form of a hierarchical representation, which is divided into three layers: environment, road, and dynamic objects. The environment layer includes weather and lighting conditions. Weather conditions (Weather) include rain, fog, snow, and other factors. Lighting conditions (Lighting) are related to the lighting conditions on the road, which are usually brighter during the day and darker at night. In addition, street lights can enhance the lighting. The road layer describes the road network and traffic signs used for road guidance. The road represents the geographical context of the collision, including the type of road where the accident occurred (CollisionLocation) and the number of lanes on the relevant road (LaneNum). In addition, the method extracts information about the road speed limit (SpeedLimit) to reconstruct the speed constraint applicable to the accident location. The dynamic object layer contains information about the impactor and the victim, as well as the movement action that caused the collision. The method extracts information about the number of traffic participants (ParticipantsNumber) that collided at the same time and identifies the collision type. The collision type information (CrashType) specifies the angle at which the traffic participants collided, including three types of collisions: rear-end collision, head-on collision, and front-to-back collision. For each vehicle involved in the accident, the method extracts the state of each vehicle, including the vehicle's initial running lane (RunningLanes), the vehicle's initial running direction (DrivingDirections), and the vehicle's behavior before the collision (DrivingActions). Specifically, vehicle behavior describes the normal and abnormal behaviors taken by the vehicle. Regular driving actions include U-turns, stops, entering the road, vehicles crossing the road, turning left, turning right, following lanes, and changing lanes. Abnormal driving behaviors include driving off the road and driving in the opposite direction.

[0009] The proposed method uses prompting patterns to extract accident information layer by layer. For each language pattern, the meaning of each attribute is first explained to help LLM understand the extracted information. In addition, the pattern also includes heuristic rules to guide LLM to produce accurate results. Finally, few-shot learning is used to ensure that the output of LLM meets the expected standards.

[0010] 2) Trajectory planning; This method uses the extracted driving actions and road information to simulate collisions and calculate the trajectories of simulated traffic participants. The inputs are the road information R extracted in step 1), the dynamic object information D extracted in step 1), the given map MAP, and the defined driving action constraints C. First, the given map is parsed into a set of candidate roads, and the road type (e.g., intersection) of each road is noted. Then the candidate road set is iterated, and lanes of different lengths and widths are selected in turn. If the current road type matches the road type where the accident occurred, and the maximum number of lanes occupied by all traffic participants during their movement is less than or equal to the number of lanes on the selected road, the waypoint generation process is started. Subsequently, after the traffic participants adjust their driving direction and initial lane position to match the selected road, the constraint solver parses the constraints based on the driving actions and generates waypoints.

[0011] In order to make the participants perform the corresponding actions and drive to the collision scene, a set of trajectory constraints C is defined for each action from five aspects. Then, the method uses a constraint solver to automatically generate a trajectory for each participant. A trajectory is an ordered sequence of waypoints, i.e., the positions and velocities that a traffic participant must follow. In order to formally represent the waypoints and thus design trajectory constraints, the b of vehicle X th Action th A waypoint can be represented as Where x, y, v are the x-coordinate, y-coordinate, and velocity, respectively. In addition, we introduce pos = (x, y) to represent the position of the waypoint in the plane. In addition, the traffic participants must reach the collision position at the same time when performing the action. To do this, the method introduces a collision area CA, which is automatically calculated by the method.

[0012] Group 1: Initial position and destination constraints

[0013] The first set has four constraints. Equation 1 restricts arrive Driving direction and lane l m Calculate fd(w i , w j , l) is defined as

[0014] (w j .xw i .x)(l ex .xl en .x)>0∧(w j .yw i .y)(l ex .yl en .y)>0.

[0015] Among them l en is the entrance point of the lane, l exis the exit point of the lane, and both points are known after parsing the map. d When true, it is guaranteed that in the direction of lane 1, waypoint w j Ahead of w i Equation 2 constrains the road locations where the FollowLane behavior starts and ends. and In the same lane m Roadside R i Formulas 3 and 4 define the starting and ending road positions for left turns and passing maneuvers: When the driving maneuver avoids a collision, and and Formula 3 applies when the vehicle is on different lanes of the road; Formula 4 applies when the action results in a collision, and the destination position In CA.

[0016]

[0017] Group 2: Positional restrictions for performing actions

[0018] In this group, Equation 5 restricts the position of the waypoint when executing the left turn action, ensuring that the current waypoint is located on the right side of the line connecting the two adjacent waypoints. Represents the direction vector consisting of the i-th waypoint position and the (i+1)-th waypoint position. represents the right normal vector formed by the positions of the i-th and (i+2)-th waypoints. Equations 6 and 7 impose constraints on the positional relationship between roads visited by executing left-turn behavior and vehicle-crossing behavior, respectively. k1 and k2 represent the position of road l m and l n The slope of .

[0019]

[0020] Group 3: Waypoint speed constraints

[0021] Group 3 limits the speed at all intermediate waypoints. arrive The speed of the waypoint,Equation 9 limits the speed limit during driving.,DX(i,j) and DY(i,j) calculate the distance of the vehicle from the,ith waypoint to the jth waypoint along the X-axis and the Y-axis, respectively. Where vc.x should be the speed of the waypoint on the X axis, and vc.y should be the speed of the waypoint on the Y axis. i The speed is v i .spd limit It is the speed limit on the driving road.

[0022]

[0023] Group 4: Trajectory combination constraints for multiple basic driving actions

[0024] The fourth group constrains the relationship between waypoints when multiple actions are connected. Equations 10 and 11 represent the trajectory constraints for multiple actions of the aggressor and victim, respectively. The end position of the current action of the traffic participant should be consistent with the start position of the subsequent driving action. The aggressor performs n actions, while the victim performs m actions before the collision.

[0025]

[0026] Group 5: Vehicle Collision Constraints

[0027] The fifth set of constraints ensures that the collision information is consistent with the report description. Equation 12 restricts the occurrence of the accident collision to the location described in the collision report, while Equation 13 describes that the attacker and the victim arrive at the collision location at the same time.

[0028]

[0029] 3) Simulation and test generation; The simulator receives the waypoints generated by the trajectory planning module and, combined with the environmental condition information extracted by the LLM, performs simulation to reconstruct the collision scenario described in detail in the accident report. The method then automatically extracts system-level test cases from the output of the trajectory planning module, allowing the autonomous vehicle to choose a different trajectory than described in the accident report to avoid the collision. For the simulation scenarios generated based on the accident report, multiple test scenarios are created by designating each traffic participant as the ego vehicle. The starting point of each autonomous vehicle is set to the starting point of the generated waypoints, while the NPC vehicle follows the waypoints provided by the trajectory planning module. Collision scenarios involving autonomous vehicles are recorded during these runs. To minimize false positives, collision scenarios where the autonomous vehicle's speed is not close to zero are calculated.

[0030] Further, the specific steps of the above step 1) are as follows:

[0031] Step 1)-1: Starting state;

[0032] Step 1)-2: Enter the accident report text;

[0033] Step 1)-3: Design of large language model prompt mode;

[0034] Step 1)-4: Large language model information extraction;

[0035] Step 1)-5: Output the extracted key information (environment, road, dynamic object);

[0036] Step 1)-6: End state.

[0037] Further, the specific steps of the above step 2) are as follows:

[0038] Step 2)-1: Starting state;

[0039] Step 2)-2: Input the extracted road information, the extracted dynamic object information, the simulation map to be tested Map, and the designed accident constraint C;

[0040] Step 2)-3: The parsing tool parses the road information to be tested in the Map;

[0041] Step 2)-4: Determine whether the extracted road type matches the selected parsed road type. If so, execute the next step, otherwise execute step 2)-3;

[0042] Step 2)-5: Determine whether the number of occupied lanes of the extracted information is less than or equal to the total number of lanes of the parsed road. If so, proceed to the next step, otherwise proceed to step 2)-3;

[0043] Step 2)-6: Adjust the driving direction and initial lane position;

[0044] Step 2)-7: Solve the constraints to generate waypoints;

[0045] Step 2)-8: Output the accident trajectory;

[0046] Step 2)-9: End state.

[0047] Further, the specific steps of the above step 3) are as follows:

[0048] Step 3)-1: Starting state;

[0049] Step 3)-2: Input the extracted environmental information and all accident participant waypoints;

[0050] Step 3)-3: Use a simulator to simulate an accident;

[0051] Step 3)-4: Select the starting point and end point of the autonomous driving vehicle to be tested in the trajectory in the accident;

[0052] Step 3)-5: Generate simulation test cases;

[0053] Step 3)-6: Test the autonomous driving system;

[0054] Step 3)-7: Output collision test results;

[0055] Step 3)-8: End state. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1The present invention is a flowchart of a method for constructing a generalization test scenario for autonomous driving based on accident reports in an implementation of the present invention.

[0057] Figure 2 for Figure 1 Flowchart of information extraction.

[0058] Figure 3 for Figure 1 Flowchart of trajectory planning.

[0059] Figure 4 for Figure 1 Flowchart of simulation and test generation. DETAILED DESCRIPTION

[0060] In order to better understand the technical content of the present invention, specific embodiments are given and described as follows in conjunction with the accompanying drawings.

[0061] Figure 1 The present invention is a flowchart of a method for constructing a generalization test scenario for autonomous driving based on accident reports.

[0062] A method for constructing a generalization test scenario for autonomous driving based on accident reports, characterized in that it includes the following steps.

[0063] S1 information extraction uses a large language model to extract accident report text information, and designs the same language prompt mode to extract accurate accident information.

[0064] S2 trajectory planning, plans a trajectory consistent with the accident report information, inputs the extracted information, and outputs the waypoints of all traffic participants by solving the accident constraints.

[0065] S3 simulation and test generation uses waypoints of traffic participants to reproduce simulated accident scenarios and convert them into test cases to test the autonomous driving system.

[0066] Figure 2 This is a flowchart for information extraction, which parses and extracts accident text information. The specific steps are as follows:

[0067] Step 1: Starting state; Step 2: Input accident report text; Step 3: Design of large language model prompt mode; Step 4: Extraction of large language model information; Step 5: Output of extracted key information (environment, road, dynamic objects); Step 6: End state.

[0068] Figure 3 This is a flow chart of trajectory planning. Analyze the simulated map to be tested, match the analyzed road to be tested with the accident road information, and use 5 sets of constraints to solve the trajectory of traffic participants. The specific steps are as follows:

[0069] Step 1: Starting state; Step 2: Input the extracted road information, the extracted dynamic object information, the simulation map to be tested Map, and the designed accident constraint C; Step 3: The parsing tool parses the road information to be tested in the Map; Step 4: Whether the extracted road type matches the selected parsed road type, if so, execute the next step, otherwise execute step 3; Step 5: The number of occupied lanes of the extracted information is less than or equal to the total number of lanes of the parsed road, if so, execute the next step, otherwise execute step 3; Step 6: Adjust the driving direction and initial lane position; Step 7: Perform constraint solving to generate waypoints; Step 8: Output the accident trajectory; Step 9: End state.

[0070] Figure 4 Flowchart generated for simulation and testing. Based on the waypoint information generated previously, the traffic accident scene is reproduced and converted into a simulation test scene to achieve the purpose of testing the autonomous driving system. The specific steps are as follows:

[0071] Step 1: Starting state; Step 2: Input the extracted environmental information and all accident participant waypoints; Step 3: Use the simulator to simulate the accident; Step 4: Select the starting and end points of the autonomous driving vehicle to be tested in the trajectory in the accident; Step 5: Generate simulation test cases; Step 6: Test the autonomous driving system; Step 7: Output the collision test results; Step 8: End state.

[0072] In summary, the present invention introduces a tool designed to automatically reconstruct accident scenarios from accident reports and test ADS. The method uses LLM to extract detailed accident information, and significantly improves the text understanding and parsing capabilities of LLM by designing a dedicated language pattern for extracting prompts. In order to enhance the generalization ability of the method in accident scene reconstruction, the method generates accident-related trajectories by solving a set of predefined trajectory specifications. The method can successfully replicate accident reports on different map structures, and the reconstructed simulation scenarios can identify various types of safety violations based on industrial-grade ADS. This method plays a key role in ensuring the quality and reliability of ADS.

Claims

1. A method for constructing a generalization test scenario for autonomous driving based on accident reports, characterized in that: According to the layer-based scenario definition, the accident scenario information is initially abstracted into three layers: road network and traffic guidance objects, environmental conditions, and dynamic objects. Then, a large language model (LLM) is used to systematically extract information about accident influencing factors from the accident report and organize the information layer by layer. Subsequently, constraints are established on the pre-accident driving behavior of traffic participants, and a constraint solver is used to generate trajectories that meet the constraints specified in the accident report. Next, the trajectories generated by the extracted environmental and road information are input into the driving simulator to reconstruct the car accident scenario. Finally, in the evaluation phase, the generated simulation scenarios are converted into test cases containing test oracles, and these test cases are input into the automated driving system (ADS). The method checks whether the ADS under test successfully reaches the final expected position without encountering any crashes. The method comprises the following steps: 1) Information Extraction; To abstractly represent the accident information and organize it into a semantic structure that can be understood by LLM, the extracted information is presented in the form of a hierarchical representation, which is divided into three layers: environment, road, and dynamic objects. The environment layer includes weather and lighting conditions. Weather conditions (Weather) include rain, fog, snow, and other factors. Lighting conditions (Lighting) are related to the lighting conditions on the road, which are usually brighter during the day and darker at night. In addition, street lights can enhance the lighting. The road layer describes the road network and traffic signs used for road guidance. The road represents the geographical context of the collision, including the type of road where the accident occurred (CollisionLocation) and the number of lanes on the relevant road (LaneNum). In addition, the method extracts information about the road speed limit (SpeedLimit) to reconstruct the speed constraint applicable to the accident location. The dynamic object layer contains information about the impactor and the victim, as well as the movement action that caused the collision. The method extracts information about the number of traffic participants (ParticipantsNumber) that collided at the same time and identifies the collision type. The collision type information (CrashType) specifies the angle at which the traffic participants collided, including three types of collisions: rear-end collision, head-on collision, and front-to-back collision. For each vehicle involved in the accident, the method extracts the state of each vehicle, including the vehicle's initial running lane (RunningLanes), the vehicle's initial running direction (DrivingDirections), and the vehicle's behavior before the collision (DrivingActions). Specifically, vehicle behavior describes the normal and abnormal behaviors taken by the vehicle. Regular driving actions include U-turns, stops, entering the road, vehicles crossing the road, turning left, turning right, following lanes, and changing lanes. Abnormal driving behaviors include driving off the road and driving in the opposite direction. The method uses prompt patterns to extract accident information layer by layer. For each language pattern, the meaning of each attribute is first explained to help LLM understand the extracted information. In addition, the model also includes heuristic rules to guide LLM to produce accurate results. Finally, few-shot learning is used to ensure that the output of LLM meets the expected standards. 2) Trajectory planning; This method uses the extracted driving actions and road information to simulate collisions and calculate the trajectories of simulated traffic participants. The inputs are the road information R extracted in step 1), the dynamic object information D extracted in step 1), the given map MAP, and the defined driving action constraints C. First, the given map is parsed into a set of candidate roads, and the road type (e.g., intersection) of each road is noted. Then the candidate road set is iterated, and lanes of different lengths and widths are selected in turn. If the current road type matches the road type where the accident occurred, and the maximum number of lanes occupied by all traffic participants during their movement is less than or equal to the number of lanes on the selected road, the waypoint generation process is started. Subsequently, after the traffic participants adjust their driving direction and initial lane position to match the selected road, the constraint solver parses the constraints based on the driving actions and generates waypoints. In order to make the participants perform the corresponding actions and drive to the collision scene, a set of trajectory constraints C is defined for each action from five aspects. Then, the method uses a constraint solver to automatically generate a trajectory for each participant. A trajectory is an ordered sequence of waypoints, i.e., the positions and velocities that a traffic participant must follow. In order to formally represent the waypoints and thus design trajectory constraints, the b of vehicle X th Action th A waypoint can be represented as in, x, y, v are the x-coordinate, y-coordinate, and velocity, respectively. In addition, we introduce pos = (x, y) to represent the position of the waypoint in the plan view. In addition, the traffic participants must reach the collision position at the same time when performing the maneuver. To do this, the method introduces a collision area CA, which is automatically calculated by the method. Group 1: Initial position and destination constraints The first set has four constraints. Equation 1 restricts arrive Driving direction and lane l m Calculate fd(w i , w j , l) is defined as (w j .x-w i .x)(l ex .x-l en .x)>0∧(w j .y-w i .y)(l ex .y-l en .y)>0。 Among them l en is the entrance point of the lane, l ex is the exit point of the lane, and both points are known after parsing the map. d When true, it is guaranteed that in the direction of lane 1, waypoint w j Ahead of w i Equation 2 constrains the road locations where the FollowLane behavior starts and ends. and In the same lane m Roadside R i Formulas 3 and 4 define the starting and ending road positions for left turns and passing maneuvers: When the driving maneuver avoids a collision, and and Formula 3 applies when the vehicle is on different lanes of the road; Formula 4 applies when the action results in a collision, and the destination position In CA. Group 2: Positional restrictions for performing actions In this group, Equation 5 restricts the position of the waypoint when executing the left turn action, ensuring that the current waypoint is located on the right side of the line connecting the two adjacent waypoints. Represents the direction vector consisting of the i-th waypoint position and the (i+1)-th waypoint position. represents the right normal vector formed by the positions of the i-th and (i+2)-th waypoints. Equations 6 and 7 impose constraints on the positional relationship between roads visited by executing left-turn behavior and vehicle-crossing behavior, respectively. k1 and k2 represent the position of road l m and l n The slope of . Group 3: Waypoint speed constraints Group 3 limits the speed at all intermediate waypoints. arrive The speed of the waypoint,Equation 9 limits the speed limit during driving.,DX(i,j) and DY(i,j) calculate the distance of the vehicle from the,ith waypoint to the jth waypoint along the X-axis and the Y-axis, respectively. Where vc.x should be the speed of the waypoint on the X axis, and vc.y should be the speed of the waypoint on the Y axis. i The speed is v i .spd limit It is the speed limit on the driving road. Group 4: Trajectory combination constraints for multiple basic driving actions The fourth group constrains the relationship between waypoints when multiple actions are connected. Equations 10 and 11 represent the trajectory constraints for multiple actions of the aggressor and victim, respectively. The end position of the current action of the traffic participant should be consistent with the start position of the subsequent driving action. The aggressor performs n actions, while the victim performs m actions before the collision. Group 5: Vehicle Collision Constraints The fifth set of constraints ensures that the collision information is consistent with the report description. Equation 12 restricts the occurrence of the accident collision to the location described in the collision report, while Equation 13 describes that the attacker and the victim arrive at the collision location at the same time. 3) Simulation and test generation; The simulator receives the waypoints generated by the trajectory planning module and, combined with the environmental condition information extracted by the LLM, performs simulation to reconstruct the collision scenario described in detail in the accident report. The method then automatically extracts system-level test cases from the output of the trajectory planning module, allowing the autonomous vehicle to choose a different trajectory than described in the accident report to avoid the collision. For the simulation scenarios generated based on the accident report, multiple test scenarios are created by designating each traffic participant as the ego vehicle. The starting point of each autonomous vehicle is set to the starting point of the generated waypoints, while the NPC vehicle follows the waypoints provided by the trajectory planning module. Collision scenarios involving autonomous vehicles are recorded during these runs. To minimize false positives, collision scenarios where the autonomous vehicle's speed is not close to zero are calculated.

2. The method for constructing a generalizable test scenario for autonomous driving based on accident reports according to claim 1 is characterized in that, in step 1), information extraction is performed, and valuable information about the road environment and vehicle trajectory contained in the accident report is extracted using LLM. During the information extraction process, the method incrementally parses the description of the car accident and accumulates information about weather, lighting, roads, and vehicles into a data structure to form a summary of the car accident. An appropriate prompt is designed to accurately describe the content of the query or request to improve the extraction accuracy and effectiveness of the LLM.

3. The method for constructing a generalization test scenario for autonomous driving based on accident reports according to claim 1, characterized in that: In step 2), trajectory planning is performed, using the extracted driving maneuvers and road information to simulate collisions and calculate the trajectories of simulated traffic participants. The constraint solver parses the constraints according to the set driving actions and generates a sequence of waypoints for the trajectory.

4. The method for constructing a generalization test scenario for autonomous driving based on accident reports according to claim 1, characterized in that: In step 3), simulation and test generation are performed, and the simulator uses the waypoints of the trajectory planning module and the environmental information provided by the LLM to reproduce the collision scenario in the accident report. This method automatically creates system-level test cases by setting one of the traffic participants as an autonomous vehicle, setting the start and end points according to the trajectory planning, solving the waypoint sequence, generating test scenarios and recording collision situations.

Images