Key scene generation method and system for autonomous driving system simulation test

Abstract

This invention discloses a method and system for generating key scenarios for simulation testing of autonomous driving systems. The method comprises: 1) constructing a corresponding abstract scenario based on each influential behavioral pattern and defining the behaviors of the participants; the master vehicle is an autonomous driving vehicle; 2) generating multiple different trajectories for each participant in the abstract scenario according to the defined behavioral constraints, combining the trajectories of each participant to obtain multiple abstract scenario instances and instantiating them as test scenarios; 3) for each test scenario, calculating the minimum expected collision time (METTC) of the master vehicle based on the expected driving route of the master vehicle and the trajectories of each participant; if the METTC is greater than a set threshold, adding new participants and their behavioral trajectories to the specific test scenario to generate a new specific test scenario, making its METTC less than or equal to the set threshold; and using the specific test scenarios with a METTC less than or equal to the set threshold as key scenarios.

Description

Technical Field

[0001] This invention relates to a method and system for generating key scenarios for simulation testing of autonomous driving systems, belonging to the field of computer software technology. Background Technology

[0002] Autonomous driving systems are highly complex hardware and software systems operating in highly complex and variable environments, making safety paramount. Therefore, comprehensive and thorough testing and evaluation are essential before deploying autonomous driving systems in real-world traffic. Testing autonomous driving systems requires long-distance driving in various traffic scenarios to observe the system's performance and identify internal problems. However, testing autonomous driving systems in the real world presents challenges such as high scenario construction costs, significant time consumption, and potential safety hazards. In contrast, simulation testing based on virtual scenarios offers advantages such as low construction costs, high flexibility, reproducible scenarios, and no safety risks.

[0003] Current methods for generating autonomous driving simulation scenarios mainly fall into two categories: scenario reproduction driven by real-world traffic data and scenario generation using search techniques. Scenario reproduction driven by real-world traffic data (see: Shuo Feng, Xintao Yan, Haowei Sun, Yiheng Feng, and Henry X Liu. 2021. Intelligent driving intelligencetest for autonomous vehicles with naturalistic and adversaria environment. Nature communications (2021), 1–14); or Xintao Yan, Shuo Feng, Haowei Sun, and Henry X Liu. 2021. Distributionally Consistent Simulation of Naturalistic Driving Environment for Autonomous Vehicle Testing. arXiv preprint arXiv:2101.02828 (2021)) simulates natural driving environments for testing autonomous driving systems. However, because safety incidents are relatively rare in real-world traffic scenarios under natural conditions, most scenarios lack sufficient challenge for autonomous driving systems, requiring hundreds of millions of miles to prove the safety performance of autonomous vehicles, which is highly inefficient.To improve the efficiency of testing autonomous driving systems, a research team has proposed a method for generating test scenarios based on traffic accidents (see: Wassim G Najm, Raja Ranganathan, Gowrishankar Srinivasan, John D Smith, Samuel Toma, Elizabeth Swanson, August Burgett, et al. 2013. Description of light-vehicle pre-crash scenarios for safety applications based on vehicle-to-vehicle communications. Technical Report. United States. National Highway Traffic Safety Administration; or Alessio Gambi, Tri Huynh, and Gordon Fraser. 2019. Generating Effective Test Cases for Self-Driving Cars from Police Reports. In Proceedings of the ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering). Based on a traffic accident database, a set of pre-collision scenarios is developed, describing the motion and dynamics of the vehicles, as well as the key events that occur immediately before a collision involving at least one other vehicle. However, these test scenarios are fixed, and while autonomous driving systems may perform well in these scenarios, their performance under broader conditions has not been adequately tested and evaluated.

[0004] Key scenarios are challenging scenarios for autonomous driving systems. The key to autonomous driving system simulation testing lies in how to generate and identify key scenarios. Scene generation techniques based on search methods, considering the complex input and high-dimensional space of traffic scenes, have employed search techniques to guide the generation of key scenes in order to effectively generate them (see: Alessio Gambi, Marc Mueller, and Gordon Fraser. 2019. Automatically testing self-driving cars with search-based procedural content generation. In Proceedings of ACM SIGSOFT International Symposium on Software Testing and Analysis. 318–328; or Moritz Klischat and Matthias Althoff. 2020. A multi-step approach to accelerate the computation of reachable sets for road vehicles. In Proceeding of IEEE International Conference on Intelligent Transportation Systems (ITSC). IEEE, 1–7). By searching within the scene space, these techniques guide scene elements to generate stronger interference for the autonomous driving system. During the scene generation process, feedback from simulation tests is used to optimize the next step of scene generation (see: Tuncali CE, Fainekos G. Rapidly-exploring random trees-based test generation for autonomous driving). vehicles[J].arXiv preprint arXiv:1903.10629,2019); For example, scene elements can be set as agents for reinforcement learning, and activation functions can be designed to enhance the interference of scene elements on the autonomous driving system during scene generation.

[0005] Due to the high-dimensional space and infinite range of parameters in traffic scenarios, optimization sampling search methods would result in high search costs. A research team modeled scene elements, transforming the optimization sampling problem in infinite space into an evolutionary search problem in finite space (see: Guanpeng Li, Yiran Li, Saurabh Jha, Timothy Tsai, Michael Sullivan, Siva Kumar Sastry Hari, Zbigniew Kalbarczyk, and Ravishankar Iyer. 2020; Av-fuzzer: Finding safety violations in autonomous driving systems. In Proceedings of IEEE International Symposium on Software Reliability Engineering (ISSRE). 25–36; Abdessalem RB, Panichella A, Nejati S, et al. Testing autonomous cars for feature interaction failures using many-objective search[C] / / 2018 33rd IEEE / ACM International Conference on Automated Software Engineering (ASE). IEEE, 2018: 143-154; Ben Abdessalem R, Nejati S, Briand L C, et al. Testing advanced driver assistance systems using multi-objectivesearch and neural networks[C] / / Proceedings of the 31st IEEE / ACM international conference on automated software engineering.2016:63-74; or Onieva E, Hernández-Jayo U, Osaba E, et al.A multi-objective evolutionary algorithm for tuning fuzzy rule bases for uncoordinated intersections in autonomous driving[J]. Information Sciences, 2015, 321:14-30); For example, modeling the behavior of dynamic participants in a scene (such as vehicles and pedestrians), generating initial scenes through random combinations of behaviors; formulating a multi-objective fitness function to measure the generated scenes, filtering out scenes with high criticality and large differences; and generating more critical and diverse scenes through evolutionary operations such as crossover and mutation, allowing for more comprehensive testing of the autonomous driving system. However, this approach often gets stuck in local optima, repeatedly discovering similar critical scenes. For autonomous driving technology providers like Baidu Apollo that release software updates weekly, this scene generation method is difficult to meet testing requirements. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a simulation testing method and system for autonomous driving systems based on key scenarios generated from influential behavioral patterns. The definition of influential behavioral patterns is as follows:

[0007] Influencing behavioral patterns refer to behaviors that contribute to a collision in a traffic accident, including the behavior of those involved in the collision and the behavior of other participants who have a potential impact on the occurrence of the accident. For example, in Figure 1 In a traffic accident, car A was traveling in front of car B. Although no collision occurred between them, car A's actions led to a collision between cars B and C. Therefore, for car B, the actions that caused the collision included a participant vehicle cutting in (car A's actions before the accident) and a participant vehicle crossing the road (car C's actions before the collision). For car C, the actions that caused the collision included a participant vehicle cutting in (car A's actions before the accident) and a participant vehicle turning (car B's actions before the collision). The influential behavioral patterns in this accident include two types: participant vehicle cutting in + participant vehicle turning (for car C) and participant vehicle cutting in + participant vehicle crossing the road (for car B). Based on real traffic accidents and following the above approach, we explored the influential behavioral patterns of traffic participants from the actions of those involved in the collision and those potentially influencing the accident.

[0008] By utilizing impactful behavioral patterns in traffic accidents, the behaviors of dynamic participants in a scenario are defined, and abstract scenarios are constructed. For each abstract scenario, different test scenarios are generated by solving the trajectory specifications of the behaviors. The criticality of the generated scenarios is evaluated using the Minimum Estimated Collision Time (METTC), and the criticality of the scenarios is improved by incrementally adding participant behavior trajectories.

[0009] Technical solution of the present invention:

[0010] A method for generating key scenarios for simulation testing of autonomous driving systems, comprising the following steps:

[0011] 1) Construct a corresponding abstract scenario based on each influential behavior pattern, and set the expected driving route of the master vehicle and the behavior of each participant in the abstract scenario; the influential behavior pattern refers to the behavior of the participants that directly or indirectly cause the master vehicle to have a traffic accident; the master vehicle is a vehicle connected to the autonomous driving system; the participants are vehicles or pedestrians other than the master vehicle in the abstract scenario.

[0012] 2) For each abstract scenario, generate multiple different trajectories for each participant in the abstract scenario according to the set behavioral norm constraints, combine the trajectories of each participant to obtain multiple abstract scenario instances; and instantiate each abstract scenario into multiple specific test scenarios.

[0013] 3) For each specific test scenario, calculate the minimum expected collision time (METTC) of the main vehicle based on the expected driving route of the main vehicle and the trajectories of each participant;

[0014] 4) If the METTC in the specific test scenario is greater than the set threshold, new participants and their behavioral trajectories are added incrementally to the specific test scenario to generate a new specific test scenario until the minimum expected collision time (METTC) of the main vehicle in the newly generated specific test scenario is less than or equal to the set threshold.

[0015] 5) The specific test scenarios in which the minimum estimated collision time (METTC) of the main vehicle is less than or equal to a set threshold are used as key scenarios for simulation testing of the autonomous driving system.

[0016] Furthermore, in step 3), based on the expected driving route of the master vehicle and the trajectory of each participant, the expected collision time between the master vehicle and each participant is calculated. (x1, y1) are the current coordinates of the main vehicle, and (x2, y2) are the current coordinates of the participants; (x + ,y +) is the expected collision location, v1 is the current speed of the main vehicle, v2 is the current speed of the participant; θ1 is the angle between the current speed direction of the main vehicle and the road, θ2 is the angle between the current speed direction of the participant and the road; the minimum value of the expected collision time ETTC is taken as the minimum expected collision time METTC.

[0017] Furthermore, based on the type rt of the road used in the test scenario, a behavior pattern that meets the feasibility constraint of type rt is selected from the set of influencing behavior patterns to construct an abstract scenario.

[0018] Furthermore, for multiple specific test scenarios of the same abstract scenario, the difference between the trajectories of the same participant in any two specific test scenarios is at least D meters; the trajectory tr of the same participant in the first specific test scenario i The trajectory tr in the second specific test scenario j Each trajectory is divided into n waypoints, with adjacent waypoints having the same time step; the difference between the trajectories is Distance(tr i ,tr j ) = min r～[0,n] {Dist(tr i r ,tr j r )}; For trajectory tr i The r-th waypoint on the road, For trajectory tr j The r-th waypoint on the path, Dist(tr i r ,tr j r ) is a waypoint and The distance between them.

[0019] Furthermore, the Z3 solver is used to generate multiple trajectories for each participant.

[0020] Furthermore, the simulator performs simulation tests based on the key scenarios, records the scenarios in which the main vehicle commits safety violations, and provides a scenario replay method to reproduce the scenarios of the safety violations.

[0021] A key scenario generation system for simulation testing of autonomous driving systems is characterized by including an abstract scenario construction module, a specific test scenario generation module, and a test scenario criticality evaluation module.

[0022] The abstract scene construction module is used to construct a corresponding abstract scene based on each influential behavior pattern, and to set the expected driving route of the master vehicle and the behavior of each participant in the abstract scene; the influential behavior pattern refers to the behavior of the participants that directly or indirectly cause the master vehicle to have a traffic accident; the master vehicle is a vehicle connected to the autonomous driving system; the participants are vehicles or pedestrians other than the master vehicle in the abstract scene.

[0023] The specific test scenario generation module is used to generate multiple different trajectories for each participant in each abstract scenario according to the set behavioral norm constraints, combine the trajectories of each participant to obtain multiple abstract scenario instances, and instantiate each abstract scenario into multiple specific test scenarios.

[0024] The test scenario criticality assessment module is used to calculate the minimum estimated time to collision (METTC) of the master vehicle for each specific test scenario based on the master vehicle's expected driving route and the trajectories of each participant. If the METTC in the specific test scenario is greater than a set threshold, new participants and their behavioral trajectories are incrementally added to the specific test scenario to generate a new specific test scenario, until the minimum estimated time to collision (METTC) of the master vehicle in the newly generated specific test scenario is less than or equal to the set threshold. Then, the specific test scenario where the minimum estimated time to collision (METTC) of the master vehicle is less than or equal to the set threshold is used as the critical scenario for the simulation test of the autonomous driving system.

[0025] A server is characterized by comprising a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the steps of the methods described above.

[0026] A computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the above-described method.

[0027] A simulation testing method for an autonomous driving system that utilizes influential behavioral patterns to generate key scenarios, the method comprising:

[0028] Test Scenario Generation and Evaluation: This includes abstract scenario construction, concrete scenario generation, and scenario evaluation. Based on influential behavior patterns, a simple abstract scenario is first constructed, where participants (vehicles and pedestrians) are configured according to the selected behavior pattern, defining the behavior of each participant in the abstract scenario. Then, for each participant's behavior, multiple different trajectories are generated by constraining the behavior rules. These trajectories are combined to obtain multiple abstract scenario instances; each abstract scenario is then instantiated into multiple concrete test scenarios. For each generated concrete test scenario, before execution, the minimum predicted collision time (METTC) of the master vehicle (the vehicle connected to the autonomous driving system) is calculated based on its expected travel route and the participants' trajectories. This minimum METTC is calculated at any given moment along the master vehicle's expected travel route. For concrete test scenarios with a METTC greater than the required threshold, a new influential behavior pattern is selected. Based on the behavior within this pattern, new participants and their behavioral trajectories are incrementally added to the concrete test scenario, generating a new test scenario. For test scenarios with a METTC less than or equal to the required threshold, execution is performed in the simulator to test the autonomous driving system; for test scenarios with a METTC greater than the required threshold, this process is iterated. The ETTC calculation method is as follows: based on the planned route of the master vehicle and the trajectories of the participants in the scenario, at each moment, the time when the master vehicle and all participants continue to move according to their current states and collide is calculated, and the minimum value is taken; for the estimated collision time between the master vehicle and a participant at a certain moment, the calculation formula is as follows:

[0029]

[0030]

[0031]

[0032] (x1, y1) are the current coordinates of the main vehicle, and (x2, y2) are the current coordinates of the participants; (x + ,y + θ1 represents the predicted collision location, v1 represents the current speed of the driver vehicle, and v2 represents the current speed of the participant. θ1 is the angle between the driver vehicle's current speed direction and the road, and θ2 is the angle between the participant's current speed direction and the road. The minimum predicted collision time for the driver vehicle at each time point in the scenario is the METTC of the scenario. The smaller the METTC, the higher the probability of the autonomous vehicle committing a safety violation in the scenario.

[0033] Simulation testing includes test scenario execution, monitoring, recording, and replaying of safety violation scenarios. For test scenarios that meet ETTC requirements, the autonomous driving system is connected for simulation testing. The performance of the autonomous vehicle during scenario execution is monitored, safety violations are recorded, and a scenario replay method is provided to automatically reproduce the recorded autonomous vehicle safety violation scenarios.

[0034] The advantages of this invention are as follows:

[0035] 1) Constructing test scenarios using influential behavioral patterns can improve the effectiveness of test scenarios in addressing the challenges of autonomous driving systems;

[0036] 2) For participants in a specific test scenario, automatically solving multiple trajectories with significant differences can improve the efficiency and diversity of test scenario generation;

[0037] 3) By evaluating test scenarios and increasing the criticality of scenarios, the efficiency of discovering critical scenarios can be improved. Attached Figure Description

[0038] Figure 1 Examples of influence on sexual behavior.

[0039] Figure 2 This is the overall system flowchart.

[0040] Figure 3 This serves as an example for assessing and improving the criticality of a scenario. Detailed Implementation

[0041] The present invention will now be described in further detail with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0042] The system of this invention comprises four modules: an abstract scene construction module, a test scene generation module, a scene criticality evaluation module, and a simulation testing module. The overall system flow is as follows: Figure 2 As shown, the input is a set of influential behavioral patterns; the following describes the four modules of the system.

[0043] 1. Abstract Scene Construction Module

[0044] Based on the road type (rt) used in the test scenario, select a behavior pattern from the set of influencing behavior patterns that meets the feasibility constraint of rt, and construct an abstract scenario. rt can be a highway, a city street (without intersections), or an intersection road; not all behaviors are allowed on a specific road type. For example, crossing an intersection is not feasible on a city street without intersections. The following infeasible behavior patterns are defined for each rt: rt = highway, infeasible behavior patterns include behaviors involving "pedestrians," "crossing," or "turning"; rt = city street, infeasible behavior patterns include behaviors involving "crossing" or "turning"; rt = intersection, no infeasible behavior patterns.

[0045] 2. Test Scenario Generation Module

[0046] For each abstract scenario, by solving the behavioral specifications, multiple different trajectories conforming to the specified behaviors of each participant in the scenario are generated, resulting in multiple specific test scenarios. To ensure that the constructed test scenarios can challenge the host vehicle (the vehicle connected to the autonomous driving system under test), participants need to be close to the host vehicle when performing behaviors. To achieve this, the concept of a convergence zone is introduced, which is a designated area of ​​size L*L meters in the road. When calculating the trajectory for the participants, the participants and the host vehicle are required to enter the convergence zone on the road simultaneously to increase the probability of a collision with the host vehicle.

[0047] The convergence area can be determined by the testers. To increase the probability of primary vehicle collisions, a survey of traffic accident trajectory datasets is conducted to cluster collision locations, identifying areas where collisions are concentrated. The clustering results are then used to guide the setting of the convergence area. Specifically, for roads with intersections, it is recommended to set convergence areas at corners and connecting roads; for highways and urban streets without intersections, they can be set in the middle of the road.

[0048] After determining the convergence area, the target path of the master vehicle and the trajectories of the participants are further defined. One or more connecting lanes are selected that traverse the convergence area; then, along the selected lanes, a position before the convergence area is chosen as the master vehicle's initial position, and a position after the area is chosen as the master vehicle's destination. In the simulation scenario, the movement of the participants is defined by trajectories, which consist of a series of waypoints. The trajectories of the participants are generated, enabling them to perform corresponding actions and enter the convergence area with the master vehicle, thus generating impactful behaviors that can influence the autonomous driving system in appropriate locations. We define a set of trajectory specifications for each behavior and use a constraint solver to calculate the trajectory of each participant.

[0049] Participant behaviors include vehicle behaviors and pedestrian behaviors. Vehicle behaviors are categorized as: driving along a lane, following another vehicle, changing lanes, cutting in, parking, driving against traffic, crossing the road, and turning. Pedestrian behaviors are categorized as: walking on the crosswalk and crossing the road. The trajectory specification for each behavior includes three parts: constraints on the participant's initial position and destination, constraints on the time and position of the participant starting and completing the behavior, and constraints on intermediate waypoints. The meanings of the symbols used in the constraints are shown in Table 1 (E represents the primary vehicle, and B represents the participant performing the behavior).

[0050] Table 1. Explanation of Symbols for Behavioral Norms

[0051]

[0052]

[0053] We will use the entry behavior as an example to explain the trajectory specifications. The trajectory specifications for other behaviors are similar in concept to those described below. The first set of constraints is shown below. This set of constraints requires: 1) The endpoint is in the convergence area, ensuring that the participant can approach the master vehicle during its journey to influence it; 2) The start and end points are located in different lanes of the same road; 3) The direction from the start point to the end point does not contradict the direction of the lane where the start point is located.

[0054]

[0055]

[0056]

[0057] The second set of constraints is as follows: 1) The participant's starting position and the starting point are in the same lane, and the position at the end of the cut-in is in the same lane as the destination. Furthermore, the direction from the starting position to the end of the cut-in does not contradict the direction of the lane where the starting point is located; 2) The direction from the starting point to the starting position does not contradict the direction of the lane where the starting point is located, and the direction from the end of the cut-in to the destination does not contradict the direction of the lane where the destination is located; 3) At the start of the cut-in, the host vehicle is located in the lane where the participant's destination is located, and within a distance of H meters behind the participant. This second set of constraints ensures that the participant can drive normally on the road and perform the cut-in action, and that the action can affect the host vehicle.

[0058]

[0059]

[0060]

[0061]

[0062] The third set of constraints is shown below. This set of constraints requires that the participant be able to complete the journey between the critical waypoints determined by the first two sets of constraints, within the range of maximum speed and acceleration, and according to the waypoint location and timing requirements. This set of constraints can generate intermediate waypoints and ensure that the participant's journey complies with the road and vehicle dynamics requirements.

[0063]

[0064]

[0065]

[0066]

[0067] Based on the behavioral norms corresponding to each participant's actions, the Z3 solver is used to generate multiple trajectories for each participant in the initial abstract scenario. These trajectories are then combined to form multiple specific test scenarios. To improve the differentiation between test scenarios, a new constraint is introduced. To ensure that in multiple concrete test scenarios within the same abstract scenario, the difference between multiple trajectories for each participant is at least D meters. We define the difference between trajectories as follows: tr i r For trajectory tr i Given the r-th waypoint, Dist(a,b) calculates the distance between waypoints a and b. The trajectory is divided into n waypoints, with each waypoint having the same time step. The trajectory difference is Distance(tr i ,tr j ) = min r～[0,n] {Dist(tr i r ,tr j r By limiting the distance between the newly generated trajectory and the previously calculated trajectory based on the behavior of each participant, it can be ensured that the multiple specific test scenarios constructed are different from each other.

[0068] 3. Scene Criticality Assessment Module

[0069] For each generated scenario, the minimum estimated time to collision (METTC) of the master vehicle is used to assess its criticality. The master vehicle's planned trajectory is divided into segments with the same time steps as the waypoints of the participants' trajectories. For each waypoint, the ETTC of the master vehicle and each participant at that time is calculated, and the minimum value is taken as the master vehicle's ETTC at that time. The minimum ETTC of the master vehicle at each time step is the METTC of the scenario. The smaller the METTC, the higher the probability of a safety violation.

[0070] For test scenarios where METTC exceeds the threshold, select other influential behavioral patterns and add their behaviors one by one to the current test scenario. For each newly added behavior, generate the corresponding participants and solve for their N trajectories, generating N new scenarios. Evaluate the newly generated scenarios, iterating through this process for those where METTC does not meet the requirements.

[0071] Examples of scenario criticality assessment and improvement, such as Figure 3 As shown. For a scenario containing two trajectories: crossing and turning; if its METTC is greater than a threshold, the "pedestrian crossing + parking" behavior mode is selected. First, the participant performing the pedestrian crossing is added to the scenario, generating N trajectories for that participant, thus generating N new test scenarios; for the scenarios where the METTC is greater than the threshold, the participant performing the parking is added to the scenario, generating N trajectories for that participant, thus generating N new test scenarios.

[0072] 4. Simulation Test Module

[0073] Input the METTC-compliant test scenarios into the simulator for simulation testing. During scenario execution, the simulator reloads the roads, initializes the environment, and removes all participants from the previous scenario. Then, a new scenario is executed in the simulator. During the execution of each scenario, after all participants have generated at their initial positions, the master vehicle begins its journey towards its destination. During scenario execution, the simulator checks for any safety violations by the master vehicle; if any occur, the scenario is recorded and saved as a critical scenario, which can be automatically reproduced in the simulator.

[0074] Although specific embodiments of the invention have been disclosed for illustrative purposes to aid in understanding and implementing the invention, those skilled in the art will understand that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the invention and the appended claims. Therefore, the invention should not be limited to the content disclosed in the preferred embodiments, and the scope of protection claimed by the invention is defined by the claims.

Claims

1. A method for generating key scenarios for simulation testing of autonomous driving systems, comprising the following steps: 1) Construct a corresponding abstract scenario based on each influential behavior pattern, and set the expected driving route of the master vehicle and the behavior of each participant in the abstract scenario; the influential behavior pattern refers to the behavior of the participants that directly or indirectly cause the master vehicle to have a traffic accident; the master vehicle is a vehicle connected to the autonomous driving system; the participants are vehicles or pedestrians other than the master vehicle in the abstract scenario. 2) For each of the abstract scenarios, generate multiple different trajectories for each participant in the abstract scenario according to the set behavioral norm constraints, combine the trajectories of each participant to obtain multiple abstract scenario instances; and instantiate each of the abstract scenarios into multiple specific test scenarios. 3) For each specific test scenario, calculate the minimum estimated time to collision (METTC) of the master vehicle based on the master vehicle's expected driving route and the trajectories of each participant; wherein, calculate the estimated collision time between the master vehicle and each participant based on the master vehicle's expected driving route and the trajectories of each participant. ; , (x1, y1) are the current coordinates of the main vehicle, and (x2, y2) are the current coordinates of the participants; (x + , y + ) is the expected collision location, v1 is the current speed of the main vehicle, v2 is the current speed of the participant; θ1 is the angle between the current speed direction of the main vehicle and the road, θ2 is the angle between the current speed direction of the participant and the road; the minimum value of the expected collision time ETTC is taken as the minimum expected collision time METTC. 4) If the METTC in the specific test scenario is greater than the set threshold, new participants and their behavioral trajectories are added incrementally to the specific test scenario to generate a new specific test scenario until the minimum expected collision time (METTC) of the main vehicle in the newly generated specific test scenario is less than or equal to the set threshold. 5) The specific test scenarios in which the minimum estimated collision time (METTC) of the main vehicle is less than or equal to a set threshold are used as the key scenarios for simulation testing of the autonomous driving system.

2. The method according to claim 1, characterized in that, Based on the type rt of the road used in the test scenario, select a behavior pattern that meets the feasibility constraint of type rt from the set of influential behavior patterns, and construct an abstract scenario.

3. The method according to claim 1, characterized in that, For multiple specific test scenarios of the same abstract scenario, the difference between the trajectories of the same participant in any two specific test scenarios is at least D meters; the trajectory tr of the same participant in the first specific test scenario i The trajectory tr in the second specific test scenario j Each trajectory is divided into n waypoints, with adjacent waypoints having the same time step; the difference between the trajectories is Distance(tr i tr j ) =min r~[0,n] {Dist(tr i r tr j r )}; For trajectory tr i The r-th waypoint on the road, For trajectory tr j The r-th waypoint on the path, Dist(tr i r tr j r ) is the waypoint and The distance between them.

4. The method according to claim 1, characterized in that, The Z3 solver is used to generate multiple trajectories for each participant.

5. The method according to claim 1, characterized in that, The simulator performs simulation tests based on the key scenarios, records the scenarios in which the main vehicle commits safety violations, and provides a scenario replay method to reproduce the scenarios in which the safety violations occur.

6. A key scenario generation system for simulation testing of autonomous driving systems, characterized in that, It includes an abstract scenario construction module, a specific test scenario generation module, and a test scenario criticality evaluation module; The abstract scene construction module is used to construct a corresponding abstract scene based on each influential behavior pattern, and to set the expected driving route of the master vehicle and the behavior of each participant in the abstract scene; the influential behavior pattern refers to the behavior of the participants that directly or indirectly cause the master vehicle to have a traffic accident; the master vehicle is a vehicle connected to the autonomous driving system; the participants are vehicles or pedestrians other than the master vehicle in the abstract scene. The specific test scenario generation module is used to generate multiple different trajectories for each participant in each abstract scenario according to the set behavioral norm constraints, combine the trajectories of each participant to obtain multiple abstract scenario instances, and instantiate each abstract scenario into multiple specific test scenarios. The test scenario criticality assessment module is used to calculate the minimum estimated time to collision (METTC) of the master vehicle for each specific test scenario, based on the master vehicle's expected driving route and the trajectories of each participant. If the METTC in a specific test scenario is greater than a set threshold, new participants and their behavioral trajectories are incrementally added to that specific test scenario to generate a new specific test scenario, until the METTC of the master vehicle in the newly generated specific test scenario is less than or equal to the set threshold. Then, the specific test scenarios where the METTC of the master vehicle is less than or equal to the set threshold are used as critical scenarios for autonomous driving system simulation testing. Specifically, the estimated collision time between the master vehicle and each participant is calculated based on the master vehicle's expected driving route and the trajectory of each participant. ; , (x1, y1) are the current coordinates of the main vehicle, and (x2, y2) are the current coordinates of the participants; (x + , y + ) is the expected collision location, v1 is the current speed of the main vehicle, v2 is the current speed of the participant; θ1 is the angle between the current speed direction of the main vehicle and the road, θ2 is the angle between the current speed direction of the participant and the road; the minimum value of the expected collision time ETTC is taken as the minimum expected collision time METTC.

7. A server, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing each step of the method of any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images