Accident scenario generation method and system based on autonomous driving virtual simulation

Abstract

The present invention provides an accident scene generation method and system based on virtual simulation of autonomous driving, which randomly generates autonomous driving seed scenes; selects the exploration strategy stage, and uses the exploration queue to extract the autonomous driving seed scenes as the scenes to be tested; selects the utilization strategy stage, and uses the queue to extract the scenes with the highest probability of accident danger in historical tests, and generates variant scenes as the scenes to be tested; uses the autonomous driving simulator and the autonomous driving system software to conduct joint simulation tests, and archives the accident scenes if a traffic accident occurs; if no accident occurs, selects and extracts the scene features during the execution of the autonomous driving scene; and performs model reasoning on the extracted scene features to predict the risk probability of an accident that may occur in this scene. It can complete efficient accident scene generation with a shorter number of test rounds, and still maintain the diversity of test scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and in particular relates to a method and system for generating accident scenarios based on autonomous driving virtual simulation. Background Art

[0002] When operating in the real world, autonomous driving systems need to maintain safety and reliability in a variety of complex and unpredictable environments. Before deploying these vehicles, comprehensive testing must be performed to ensure the safety of their intended functions. Systems for simulating autonomous driving scenarios are mentioned in patents CN115757125A, CN115859788A, CN116737540A, CN112784485A, and CN115455718A, but these solutions focus on generating key scenarios for autonomous driving. These key scenarios may not necessarily trigger problems in the autonomous driving system. Using reinforcement learning models, the early training of the model requires a large data set. However, since key scenarios and abnormal scenarios are very scarce in the overall natural driving scenario, reinforcement learning cannot efficiently learn the description method of key scenario generation, thereby triggering the occurrence of autonomous driving accident scenarios. Moreover, the reinforcement learning model relies on natural driving data for training during the early training process, resulting in the types of scenarios that the model can learn being directly limited by the diversity of natural driving data. If the training data lacks sufficient diversity, the generated scenarios will also lack diversity and cannot fully cover all possible complex driving scenarios. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for generating accident scenes based on virtual simulation of autonomous driving, aiming to solve the problem that the prior art is inefficient in generating accident scenes for autonomous driving and has poor diversity in generating accident scenes.

[0004] The present invention is implemented as follows: a method for generating an accident scene based on automatic driving virtual simulation, the method comprising the following steps:

[0005] Step S1: Generate an autonomous driving scenario description script according to the autonomous driving scenario description specification, and randomly generate an autonomous driving seed scenario;

[0006] Step S2: Select the autonomous driving test scenario to be executed. If the exploration strategy stage is selected, the exploration queue takes out the autonomous driving seed scenario as the scenario to be tested and executes step S3; if the utilization strategy stage is selected, the utilization queue takes out the scenario with the highest accident hazard probability in the historical test, adjusts the scenario features, generates a variant scenario as the scenario to be tested and executes step S3;

[0007] Step S3: Use the autonomous driving simulator and the autonomous driving system software to conduct a joint simulation test. If a traffic accident occurs, the accident scene is archived; if no accident occurs, step S4 is executed;

[0008] Step S4: Select and extract scene features during the execution of the autonomous driving scene;

[0009] Step S5: Perform model inference on the scene features extracted in step S4 to predict the risk probability of an accident that may occur in this scene. The obtained risk probability will be added to the utilization queue together with the scene.

[0010] A further technical solution of the present invention is: in step S1, the parameters of the autonomous driving scene description script include the starting coordinates of the autonomous driving main vehicle position in the map, the target coordinates, the map, the weather and the behavior of traffic participants.

[0011] A further technical solution of the present invention is: in step S2, the exploration strategy phase and the utilization strategy phase are executed alternately according to the custom configured weights.

[0012] A further technical solution of the present invention is that the scene characteristics include the physical environment, traffic environment, traffic participants and the running status of the vehicle.

[0013] A further technical solution of the present invention is: in step S5, a risk prediction model is established; in the training phase, scene samples are selected from the scene data of historical tests, feature data of each scene is extracted, a feature data set of the scene is obtained, and the risk prediction model is trained.

[0014] A further technical solution of the present invention is: in the inference stage, the trained risk prediction model is used to process new scene data through the same scene feature extraction module, and new scene features are generated as model input to obtain a risk probability prediction result, and the risk probability value range is 0-1.

[0015] A further technical solution of the present invention is: in step S5, the scenarios in the queue are sorted from high to low according to the risk probability prediction results.

[0016] Another object of the present invention is to provide an accident scene generation system based on autonomous driving virtual simulation, which is used to execute an accident scene generation method based on autonomous driving virtual simulation, comprising:

[0017] Seed scenario generation module: used to generate autonomous driving scenario description scripts according to autonomous driving scenario description specifications and randomly generate autonomous driving seed scenarios;

[0018] Scenario selection and scheduling module: used to select the autonomous driving test scenario to be executed. If the exploration strategy stage is selected, the exploration queue will take out the autonomous driving seed scenario as the scenario to be tested; if the utilization strategy stage is selected, the utilization queue will take out the scenario with the highest probability of accident danger in the historical test, and adjust the scenario characteristics to generate a variant scenario as the scenario to be tested;

[0019] Scenario execution test module: used to conduct joint simulation tests using the autonomous driving simulator and the autonomous driving system software. If a traffic accident occurs, the accident scene is archived; if no accident occurs, the module goes to the scenario feature extraction module.

[0020] Scene feature extraction module: used to select and extract features during the execution of autonomous driving scenes;

[0021] Scenario risk prediction module: used to perform model inference on the extracted scenario features and predict the risk probability of an accident that may occur in this scenario. The obtained risk probability will be added to the utilization queue together with the scenario.

[0022] A further technical solution of the present invention is: the scenario risk prediction module is also used to establish a risk prediction model. In the training stage, scenario samples are selected from the scenario data of historical tests, and the feature data of each scenario is extracted to obtain a feature data set of the scene, and the risk prediction model is trained; in the reasoning stage, the trained risk prediction model is used to process new scenario data through the same scenario feature extraction module, generate new scenario features as model input, and obtain a risk probability prediction result, and the risk probability value range is 0-1.

[0023] The beneficial effects of the present invention are as follows: the present solution focuses on efficiently identifying and generating abnormal scenarios that may trigger autonomous driving accidents, thereby improving the pertinence and effectiveness of the test. When designing a scenario generation mechanism, a method of introducing a scenario generation diversity mechanism is introduced, which can complete efficient accident scenario generation in a shorter number of test rounds while still maintaining the diversity of test scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0026] Embodiment 1:

[0027] Figure 1The present invention shows a method for generating an accident scene based on autonomous driving virtual simulation, the method comprising the following steps:

[0028] Step S1: Generate an autonomous driving scenario description script according to the autonomous driving scenario description specification, and randomly generate an autonomous driving seed scenario;

[0029] Step S2: Select the autonomous driving test scenario to be executed. If the exploration strategy stage is selected, the exploration queue takes out the autonomous driving seed scenario as the scenario to be tested and executes step S3; if the utilization strategy stage is selected, the utilization queue takes out the scenario with the highest accident hazard probability in the historical test, adjusts the scenario features, generates a variant scenario as the scenario to be tested and executes step S3;

[0030] Step S3: Use the autonomous driving simulator and the autonomous driving system software to conduct a joint simulation test. If a traffic accident occurs, the accident scene is archived; if no accident occurs, step S4 is executed;

[0031] Step S4: Select and extract scene features during the execution of the autonomous driving scene;

[0032] Step S5: Perform model inference on the scene features extracted in step S4 to predict the risk probability of an accident that may occur in this scene. The obtained risk probability will be added to the utilization queue together with the scene.

[0033] Preferably, in step S1, the parameters of the autonomous driving scenario description script include the starting coordinates of the autonomous driving main vehicle position in the map, the target coordinates, the map, the weather, and the behavior of traffic participants.

[0034] Preferably, in step S2, the exploration strategy phase and the exploitation strategy phase are executed alternately according to the custom configured weights.

[0035] Preferably, the scene characteristics include the physical environment, traffic environment, traffic participants and the operating status of the vehicle.

[0036] Preferably, a risk prediction model is established in step S5. In the training phase, scenario samples are selected from the scenario data of historical tests, feature data of each scenario is extracted, and a feature data set of the scenario is obtained to train the risk prediction model.

[0037] Preferably, in the inference stage, the trained risk prediction model is used to process new scenario data through the same scenario feature extraction module to generate new scenario features as model input to obtain a risk probability prediction result with a risk probability value range of 0-1.

[0038] Preferably, in step S5, the scenarios in the utilization queue are sorted from high to low according to the risk probability prediction results.

[0039] Necessity of accident scenario generation for autonomous driving. Autonomous driving systems need to maintain safety and reliability in a variety of complex and unpredictable environments when operating in the real world. Before these vehicles are deployed, comprehensive testing is necessary to ensure the safety of their intended functions. However, traditional mileage-based testing methods have certain limitations due to their inability to fully cover extreme cases of long-tail distributions. These extreme cases represent rare but critical traffic error scenarios that autonomous vehicles must be able to handle safely. Therefore, it is very necessary to systematically expose vehicles to a variety of challenging driving conditions and enhance the robustness of autonomous driving systems by generating accident scenarios.

[0040] Achieving efficient accident scenario generation is an urgent problem. The generation of autonomous driving accident scenarios faces the problem of test efficiency. Autonomous driving test scenarios consist of complex structures of high-dimensional variables, and it is not feasible to traverse each element of all possible values ​​in a short time. In addition, each autonomous driving test usually consumes more time than traditional software testing, exacerbating the inefficiency problem. Therefore, it is necessary to optimize the mechanism design of the existing autonomous driving accident scenario generation method.

[0041] This method achieves efficient and comprehensive testing through the following key mechanisms: security risk prediction, scenario selection scheduling, and a balance strategy between unknown scenario exploration and known scenario utilization.

[0042] Definition of terms:

[0043] Accident scenario: During the virtual simulation test, the autonomous driving system may experience abnormal behaviors such as collision, traffic violation, and failure to complete the driving task within the specified time, which may affect personal and property safety and traffic efficiency;

[0044] Seed scenario: an initial autonomous driving scenario that has not been tested;

[0045] Scenario variation: refers to changing the autonomous driving test scenario by making minor changes to scenario elements such as weather and traffic participants without changing the main body of the test scenario such as the autonomous driving task, map, and road conditions.

[0046] This approach achieves efficient and comprehensive testing through the following key mechanisms:

[0047] Scenario risk prediction: This solution designs a method that can more accurately predict the risk of traffic accidents in autonomous driving test scenarios. By extracting key features during the autonomous driving test process and using a machine learning model trained with a small number of test scenarios, the risk probability of traffic accidents in autonomous driving can be more accurately predicted;

[0048] Scenario feature selection: This solution proposes to extract comprehensive features from the autonomous driving test process by combining the autonomous driving scenario description file and the monitoring data during the autonomous driving test, thereby helping the machine learning model to make more accurate risk predictions;

[0049] Scenario selection and scheduling mechanism: This selection and scheduling mechanism can guide the test process to prioritize and schedule scenarios that are more likely to cause safety-critical traffic accident behaviors. This mechanism significantly improves the test efficiency of accident scenario mining;

[0050] Exploration and Exploitation Balance: The exploration-exploitation queue mechanism is introduced to maintain a balance between exploring diverse scenarios and focusing on known high-risk conditions. This mechanism effectively avoids the testing process from falling into local optimality, thereby ensuring the extensive coverage of test scenarios and the efficiency of accident scenario mining.

[0051] Seed scenario generation: Generate autonomous driving scenario description scripts and autonomous driving test scenarios based on the OpenScenario standard or other autonomous driving scenario description specifications.

[0052] Scenario selection scheduling: Design a scenario selection scheduling mechanism to maintain untested and tested autonomous driving test scenarios by using an exploration-exploitation queue. In this stage, the autonomous driving test scenarios to be executed are selected.

[0053] Scenario execution test: A joint simulation test is performed using an autonomous driving simulator (such as Carla) and the autonomous driving system software. If a traffic accident occurs, the accident scene is archived; if no accident occurs, the scene feature extraction is performed;

[0054] Scene feature extraction: Select and extract features during the execution of autonomous driving scenes.

[0055] Scenario risk prediction: The trained machine learning model uses the scenario features of the previous stage to perform model reasoning to predict the risk probability of an accident that may occur in this scenario. The risk probability obtained in this stage will be added to the exploration-exploitation queue together with the scenario.

[0056] The main task of scenario selection scheduling is to select the autonomous driving scenario tasks to be executed. This module selects scenarios based on the selection strategy of the current stage by maintaining the exploration-exploitation queue.

[0057] In the exploration strategy phase, the diversity of scenarios is the main consideration. At this point, randomly generated seed scenarios are selected for scenario testing. Due to the introduction of random elements, these randomly generated seed scenarios can ensure the breadth of test coverage and the diversity of scenarios. This method can discover more potential accident scenarios instead of being limited to known high-risk areas; in the utilization strategy phase, the efficiency of accident scenario mining is mainly considered. At this point, the scenarios with the highest probability of accident danger in historical tests are given priority, and are tested by making minor adjustments to these scenarios. These adjustments are based on existing high-risk scenarios, and subtle changes are made to further explore potential problems.

[0058] The scenario selection scheduling alternately executes the exploration and exploitation strategies according to the custom configured weights to select the scenario tasks to be executed. This mechanism ensures that the test process can both cover a wide range of scenarios and efficiently mine safety-critical accident scenarios. Random elements of seed scenario generation, introducing randomly generated seed scenarios in the exploration strategy, by randomly selecting and generating scenarios, the limitations of local optimal solutions can be broken and the diversity of testing can be ensured. The introduction of random elements allows the test to cover more different types of scenarios, including accident scenarios that may not be discovered in the exploitation strategy. The local optimal solution problem of the exploitation strategy, in the exploitation strategy, although the potential problems can be deeply explored by slightly adjusting the high-risk scenarios, this method is prone to fall into the local optimal solution and can only find problems under certain specific conditions. By alternating the exploration strategy, new and diverse seed scenarios can be introduced, avoiding the limitation of the test falling into a single type of problem, and further improving the comprehensiveness of the test.

[0059] In order to help the autonomous driving risk prediction model, the features of the selected autonomous driving scenarios should be as comprehensive as possible. Therefore, the four dimensions of physical environment, traffic environment, traffic participants and vehicle operation status are considered in the selection of categories that constitute the autonomous driving scenarios. The data sources of the features are the autonomous driving scenario description files and the monitoring data of the autonomous driving scenario testing process. The process monitoring data can better reflect the risks encountered by autonomous driving during the test process, so that risk prediction can be more accurate.

[0060] The options and selection ranges for scenario risk prediction feature selection are as follows:

[0061]

[0062] The scenario risk prediction task is to predict the probability of an accident occurring in a scenario without a traffic accident by using the machine learning model completed through previous training to perform inference.

[0063] Scenario risk prediction requires a preliminary model training phase. The machine learning model used in this solution is the random forest model. This model is particularly suitable for autonomous driving virtual simulation testing in terms of processing high-dimensional data, resisting noise, and being easy to interpret.

[0064] During the model training phase, we select small-scale scenario samples from the scenario data of historical tests. For example, we select 50 normal scenario samples and 50 accident scenario samples. Then, these samples are processed by the scenario feature extraction module to extract the feature data of each scenario. Finally, we get a feature data set of 100 scenarios for model training. The random forest model is trained with these feature data to build a model that can accurately predict scenario risks. Since the random forest model can automatically select and optimize important features and has strong noise resistance, it can perform well in such high-dimensional and complex scenario data.

[0065] In the inference phase, the trained random forest model is used to process new scene data through the same scene feature extraction module to generate new scene features as model input. After model inference, we can get a risk probability prediction result with a value range of [0,1]. This risk probability prediction result indicates the possibility of an accident occurring in the scene.

[0066] The risk probability prediction results obtained will be used as the basis for scenario sorting and used in the scenario selection scheduling module. During the scenario selection scheduling process, this result helps determine which scenarios should be tested first. By using the exploration-exploitation queue mechanism, the module can maintain a balance between exploring diverse scenarios and focusing on testing high-risk scenarios, thereby improving the efficiency and coverage of the test.

[0067] Embodiment 2:

[0068] The present invention also provides an accident scene generation system based on autonomous driving virtual simulation, which is used to execute an accident scene generation method based on autonomous driving virtual simulation, including:

[0069] Seed scenario generation module: used to generate autonomous driving scenario description scripts according to autonomous driving scenario description specifications and randomly generate autonomous driving seed scenarios;

[0070] Scenario selection and scheduling module: used to select the autonomous driving test scenario to be executed. If the exploration strategy stage is selected, the exploration queue will take out the autonomous driving seed scenario as the scenario to be tested; if the utilization strategy stage is selected, the utilization queue will take out the scenario with the highest probability of accident danger in the historical test, and adjust the scenario characteristics to generate a variant scenario as the scenario to be tested;

[0071] Scenario execution test module: used to conduct joint simulation tests using the autonomous driving simulator and the autonomous driving system software. If a traffic accident occurs, the accident scene is archived; if no accident occurs, the module goes to the scenario feature extraction module.

[0072] Scene feature extraction module: used to select and extract features during the execution of autonomous driving scenes;

[0073] Scenario risk prediction module: used to perform model inference on the extracted scenario features and predict the risk probability of an accident that may occur in this scenario. The obtained risk probability will be added to the utilization queue together with the scenario.

[0074] Preferably, the scenario risk prediction module is also used to establish a risk prediction model. In the training phase, scenario samples are selected from the scenario data of historical tests, and the feature data of each scenario is extracted to obtain a feature data set of the scenario, and the risk prediction model is trained. In the inference phase, the trained risk prediction model is used to process new scenario data through the same scenario feature extraction module to generate new scenario features as model input to obtain a risk probability prediction result, and the risk probability value range is 0-1.

[0075] The solution is described below with reference to a specific embodiment.

[0076] (1) First, the seed scenario generates a scenario description configuration file. For example, the starting coordinates (x, y) of the autonomous driving vehicle in the map, the target coordinates (x, y), the map, the weather, the behavior of traffic participants, and other contents are provided in the form of configuration items;

[0077] (2) Assume that the weight ratio of the scenario selection scheduling configuration selection and utilization is 1:1, that is, the selection strategy and utilization strategy are executed alternately. The seed scenario enters the exploration queue. When in the exploration phase, the exploration queue takes out the scenario as the scenario to be tested;

[0078] (3) Scenario execution test process: Based on the scenario description file, the autonomous driving test object simulates the scenario through joint simulation with the autonomous driving simulator. In this process, the autonomous driving behavior status data will be monitored and recorded;

[0079] (4) If the scene is an accident scene, it will be directly archived; if no accident occurred in the scene, feature extraction will be performed on the scene;

[0080] (5) In the feature extraction stage, based on the process monitoring data of the scene and the scene description configuration file, the feature extraction module extracts the required feature data;

[0081] (6) The features extracted in the feature extraction phase are used as input to the machine learning model for reasoning, and the risk probability of a scenario prediction is obtained;

[0082] (7) The scenario and the corresponding predicted risk probability will be added to the utilization queue. In this queue, the scenarios will be sorted from high to low according to risk;

[0083] (8) In the next round, the scheduling module is selected. If it is the exploration phase, the process from (2) to (7) is continued. If it is the utilization phase, the scenario with the highest risk is taken out from the utilization queue and mutated (for example, the rainy day parameter is changed from 80 to 90, etc.), and the subsequent mutated scenario is obtained, and the process from (3) is continued.

[0084] Through the above process, this method can complete the efficient generation of accident scenarios in a shorter number of test rounds, while still maintaining the diversity of test scenarios.

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for generating accident scenarios based on autonomous driving virtual simulation, characterized in that: The method comprises the following steps: Step S1: Generate an autonomous driving scenario description script according to the autonomous driving scenario description specification, and randomly generate an autonomous driving seed scenario; Step S2: Select the autonomous driving test scenario to be executed. If the exploration strategy stage is selected, the exploration queue takes out the autonomous driving seed scenario as the scenario to be tested and executes step S3; if the utilization strategy stage is selected, the utilization queue takes out the scenario with the highest accident hazard probability in the historical test, adjusts the scenario features, generates a variant scenario as the scenario to be tested and executes step S3; Step S3: Use the autonomous driving simulator and the autonomous driving system software to conduct a joint simulation test. If a traffic accident occurs, the accident scene is archived; if no accident occurs, step S4 is executed; Step S4: Select and extract scene features during the execution of the autonomous driving scene; Step S5: Perform model inference on the scene features extracted in step S4 to predict the risk probability of an accident that may occur in this scene. The obtained risk probability and the scene will be added to the utilization queue; In step S2, the exploration strategy phase and the exploitation strategy phase are executed alternately according to the custom configured weights.

2. The accident scenario generation method based on autonomous driving virtual simulation according to claim 1, characterized in that: In step S1, the parameters of the autonomous driving scenario description script include the starting coordinates of the autonomous driving main vehicle position in the map, the target coordinates, the map, the weather, and the behavior of traffic participants.

3. The accident scenario generation method based on autonomous driving virtual simulation according to claim 1, characterized in that: The scene characteristics include the physical environment, traffic environment, traffic participants and the operating status of the vehicle.

4. The accident scenario generation method based on autonomous driving virtual simulation according to claim 1, characterized in that: In step S5, a risk prediction model is established. In the training phase, scenario samples are selected from the scenario data of historical tests, and feature data of each scenario is extracted to obtain a feature data set of the scenario, and the risk prediction model is trained.

5. The accident scenario generation method based on autonomous driving virtual simulation according to claim 4 is characterized in that: In the inference stage, the trained risk prediction model is used to process new scenario data through the same scenario feature extraction module, and new scenario features are generated as model input to obtain a risk probability prediction result with a risk probability value range of 0-1.

6. The accident scenario generation method based on autonomous driving virtual simulation according to claim 1, characterized in that: In step S5, the scenarios in the utilization queue are sorted from high to low according to the risk probability prediction results.

7. An accident scene generation system based on autonomous driving virtual simulation, used to execute the accident scene generation method based on autonomous driving virtual simulation according to any one of claims 1 to 6, characterized in that: include: Seed scenario generation module: used to generate autonomous driving scenario description scripts according to autonomous driving scenario description specifications and randomly generate autonomous driving seed scenarios; Scenario selection and scheduling module: used to select the autonomous driving test scenario to be executed. If the exploration strategy stage is selected, the exploration queue will take out the autonomous driving seed scenario as the scenario to be tested; if the utilization strategy stage is selected, the utilization queue will take out the scenario with the highest probability of accident danger in the historical test, and adjust the scenario characteristics to generate a variant scenario as the scenario to be tested; Scenario execution test module: used to conduct joint simulation tests using the autonomous driving simulator and the autonomous driving system software. If a traffic accident occurs, the accident scenario is archived. If no accident occurs, go to the scene feature extraction module; Scene feature extraction module: used to select and extract features during the execution of autonomous driving scenes; Scenario risk prediction module: used to perform model inference on the extracted scenario features and predict the risk probability of an accident that may occur in this scenario. The obtained risk probability will be added to the utilization queue together with the scenario.

8. The accident scenario generation system based on autonomous driving virtual simulation according to claim 7, characterized in that: The scenario risk prediction module is also used to establish a risk prediction model. In the training phase, scenario samples are selected from the scenario data of historical tests, and the feature data of each scenario is extracted to obtain a feature data set of the scenario, and the risk prediction model is trained. In the inference phase, the trained risk prediction model is used to process new scenario data through the same scenario feature extraction module to generate new scenario features as model input to obtain a risk probability prediction result, and the risk probability value range is 0-1.

Images