As the level of autonomous driving increases, automotive systems become more complex. Changeable weather, complex traffic environments, diverse driving tasks and dynamic driving conditions all pose new challenges for the test evaluation of autonomous vehicles. In particular, the test and evaluation objects of autonomous vehicles have changed from the dual independent system of people and vehicles in traditional cars to a strong coupling system of people-vehicle-environment-task.
Scenario-based virtual test technology test scenario configuration is flexible, test efficiency is high, test repeatability is strong, the test process is safe, the test cost is low, automatic testing and accelerated testing can be realized, saving a lot of manpower and material resources. Therefore, scenario-based virtual testing has become an indispensable and important part of the test evaluation of autonomous vehicles. Many scientific research institutions and researchers at home and abroad have carried out extensive research on this. This paper summarizes and summarizes the scenario-based virtual test research from three aspects: automatic driving test scenario, automatic driving virtual test platform, and automatic driving acceleration test method, and analyzes and looks forward to the problems faced by the current research, as well as the future development trend of autonomous vehicle virtual test research.
Autonomous driving test scenarios
The definition of the scene
The word "Scenerio" is derived from the Latin word Olinda, meaning stage play, and now refers to specific situations in life. With the development of science and technology, the concept of scenes has gradually been applied to the development and testing process of industrial production.
The application area of the "scene"
Scenario-based testing is first applied to the development of software systems, and "scenarios" are used to describe how the system is used, the requirements for use, the environment in which it is used, and to envision more viable systems. At this stage in the field of automatic driving, there is no clear and unified definition of "scenario". However, according to the definition of RAND, PEGASUS and other different institutions, its core elements are the same: they all contain road environment elements, contain other traffic participants, and contain vehicle driving tasks, and at the same time, these elements will last for a certain period of time and have the characteristics of dynamic change.
The understanding of the autonomous driving test scenario in this paper is that the scene is the overall dynamic description of the autonomous vehicle and its driving environment components over a period of time, and the composition of these elements is determined by the functions of the autonomous driving car to be tested. In short, the scene can be considered as an organic combination of the driving situation and the driving situation of the self-driving car.
The features of the scene
Determining scenario features is the first step in conducting virtual testing of scenario-based autonomous vehicles. This paper synthesizes the research of different scene elements and proposes the specific situation of the scene elements shown in Figure 2. The test scene elements mainly include two categories: test vehicles and traffic environment elements, of which the test vehicle elements include the basic elements of the test vehicle, target information and driving behavior, and the traffic environment elements include weather and lighting, static road information, dynamic road information and traffic participant information.
Scene features
The data source for the scene
The data sources of the autonomous driving test scenario mainly include three parts: real data, simulated data and expert experience, and the specific content is shown in Figure 3.
Test the scene data source
The real data sources mainly include natural driving data, accident data, roadside unit monitoring data, and typical test data such as driver tests, intelligent car closed test site tests, and open road tests. A typical natural driving scenario data acquisition vehicle configuration is shown in Figure 4.
Typical natural driving scenario data collection vehicle
Simulation data: Simulation data sources mainly include driving simulator data and simulation data. Driving simulator data is the scene element information obtained by testing with the driving simulator. Compared to road testing, driving simulators test safe, efficient, and reproducible, and can perform a wide range of driver-in-the-loop tests for hazardous and extreme conditions.
Expert experience data refers to the scene element information obtained through the summary of empirical knowledge of previous tests, and the standard regulatory test scenario is a typical data source of expert experience scenarios. At present, there are more than 80 types of autonomous driving test laws and regulations in various countries in the world. The latest "Test Procedures for Autonomous Driving Functions of Intelligent Connected Vehicles (Trial)" released by Mainland China proposes 34 test scenarios, including the identification and response of traffic signs and markings.
How the scene is handled
There are differences in the format and type of scene data between different data sources, and there are a large number of invalid data and error data in the original data, which requires appropriate processing of scene data to form a truly usable autonomous vehicle test scenario. According to the existing typical scene data processing methods, this paper summarizes and proposes the scene data processing process shown in Figure 5.
Scene data processing process
Virtual testing of autonomous driving
A typical autonomous vehicle test verification architecture is shown in Figure 6, and the test methods mainly include virtual tests in a virtual environment such as model-in-the-loop testing, driving simulator testing, hardware-in-the-loop testing and vehicle-in-the-loop testing, as well as real-vehicle testing in closed test grounds and public roads. Among them, the virtual test method mainly includes model-in-the-loop testing, hardware-in-the-loop testing and vehicle-in-the-loop testing.
Autonomous driving verification framework
Model in the loop test using simulation scenarios, vehicle dynamics models, sensor models, decision planning algorithms for automatic driving testing in the virtual environment, which is mainly used in the initial stage of system development, without hardware participation in system testing, mainly used to verify the correctness of the algorithm.
Hardware-in-the-loop testing mainly includes environment-aware system-in-the-loop testing, decision planning system in-the-loop testing and control execution system-in-the-loop testing, etc., and its test requirements include: continuous testing (automatic testing can be carried out according to the purpose at this time), combined testing (different standards are evaluated in the same scenario, such as safety, comfort, etc.), scalability (test results of simple functions have scalability, for example, for lane keeping test results can be applied to advanced autonomous driving functions).
Typical hardware-in-the-loop test scenario
Vehicle in-the-loop test is to embed the whole vehicle into the virtual test environment for testing, through the simulation of the scene to test the performance of the vehicle, mainly including the closed site vehicle in the ring and the turntable platform vehicle in the ring, the key lies in the vehicle information to the simulation environment and the sensor information generated in the simulation environment to the vehicle controller.
Vehicle-in-the-loop test protocol
Scenario-based accelerated testing of autonomous driving
At present, there are two main ways to accelerate the test of scene-based autonomous driving: one is to build the rapidity and repeatability of the test scene based on the virtual environment, randomly generate the test scene according to the test requirements, and generate a large number of test scenes in a short period of time; the other way is to refer to the dangerous scene enhancement generation method proposed by the vehicle strengthening corrosion test method.
The test scenario is randomly generated
The technical route of random generation of test scenes mainly includes the generation method based on random sampling represented by Monte Carlo simulation method and fast search random tree, the generation method based on the hierarchical analysis of the importance of scene features, and the method based on machine learning.
Compared with building real-world test scenarios in the real world, generating test cases in a virtual environment can greatly reduce time and resource consumption. However, due to the low probability of accidents under natural circumstances, the use of random scene generation may still face the problem of a large number of calculations, and the method of enhanced generation of dangerous scenes can solve this problem very well.
Dangerous scene enhancement spawn
If a self-driving car performs well in dangerous situations, its system safety can also be well guaranteed. Therefore, testing the performance of self-driving cars in dangerous scenarios has attracted more and more attention from scholars.
According to the definition of dangerous scenarios, Zhao D et al. proposed a method of accelerating the generation of dangerous scenes with importance sampling, the core idea of which is to introduce a new probability density function f*(x) instead of the original f(x), increase the probability of generating dangerous scenes, thereby reducing the number of tests. When using the randomly sampled scene generation method, the probability density function of the dangerous scene is f(x), and the minimum number of tests is:
where γ is the probability of the occurrence of a dangerous scenario, β is a given constant, and z is related to the inverse cumulative distribution function of N(0,1).
When using importance sampling for hazard scenario generation, the probability density function of the hazard scenario is f*(x) and the minimum number of tests is:
where I(x) is the indicator function of the dangerous event, L(x) is the likelihood ratio of the importance sampling, and the probability of the dangerous scene occurring after changing the probability density function
By verifying the method of enhancing the generation of dangerous scenes in typical scenarios such as front car cutting in and front brake, it is proved that the test speed can reach 7000 times that of Monte Carlo test simulation.
Research Perspectives
Although scholars from various countries have conducted extensive research on scenario-based virtual testing of autonomous vehicles and achieved certain results, the current level of research cannot meet the urgent needs of autonomous vehicle testing, and a perfect virtual test evaluation system for autonomous vehicles has not yet been established worldwide. In the future, further in-depth research is still needed in the following aspects:
Scene deconstruction and automatic refactoring technology
The real traffic scene is complex and changeable, and the amount of data is huge, so the scene feature features should be extracted according to the scene feature analysis to achieve scene deconstruction. At the same time, the scene elements are complex and varied, and when testing different autonomous driving functions, the types of scene elements required are not the same. How to automatically reconstruct the test scene according to the test requirements is a key problem that needs to be solved urgently.
High-confidence modeling of human-vehicle-environment system integration
At present, the people, vehicles, and environment models are mostly constructed separately, and the coupling relationship between them has not yet been clarified, and the coupling mechanism of the sensor model and the environmental model should be described through the path loss, shadow attenuation and noise modeling of the sensor signal, and then the influencing factors of the driver, vehicle and environment are comprehensively analyzed, and a high-confidence model of the integration of people-vehicle-environment is constructed.
Build a standard toolchain for virtual testing of autonomous vehicles
At present, the virtual test of autonomous vehicles refers to the "V" type process, and in the future, the test advantages between different virtual test platforms should be clarified, and the integrated testing technology of the autonomous driving system in the loop under close service conditions and the multi-configuration executor are adopted to establish a unified and standardized virtual test standard tool chain for autonomous vehicles.
Hybrid traffic simulation and testing under different penetration rates of autonomous vehicles
Establishing a hybrid traffic model under the penetration rate of different autonomous vehicles, analyzing the traffic situation and vehicle behavior of different autonomous vehicles, and conducting hybrid traffic tests are a new research field for virtual testing of autonomous driving in the future.
Establish a dynamic adaptive random generation mechanism for test cases
According to the combination criteria and constraint relationship of scene elements, the construction of multi-hazard level test scenarios, the establishment of a dynamic adaptive random generation mechanism for test cases, and the realization of high-speed concurrency of massive data are the research priorities of virtual testing of automatic driving in the future
Establish a virtual test standard system for autonomous vehicles
Environmental complexity, task complexity, manual intervention, driving intelligence, etc. can be used as the evaluation content of virtual tests. In the future, we should develop a virtual test evaluation system architecture that adapts to the development trend of technology, and establish a test standard system.
Due to the extremely complex, infinitely rich and unpredictable characteristics of the driving scenarios of autonomous vehicles, the traditional road test methods can no longer meet the needs of automatic driving testing, and scenario-based virtual testing has become an indispensable part of autonomous vehicle test verification.
In this paper, the research progress of scenario-based virtual test technology for autonomous vehicles is systematically sorted out and summarized, the connotation of the test scene is clarified, the elements of the test scene are summarized, the data sources of the test scene are summarized, and the processing method of the scene data is systematically summarized. On this basis, the model-in-the-loop, hardware-in-the-loop and vehicle-in-the-loop test schemes and their key technologies are systematically sorted out, and the typical random generation methods of test scenarios and the enhanced generation methods of dangerous scenarios are summarized.
Scenario-based virtual testing of autonomous vehicles is still in the initial stage of technological development, which needs to be jointly promoted by scientific researchers in multiple fields, and future research should focus on breaking through core common technologies such as test scenario database based on deconstruction and automatic reconstruction, high-confidence modeling of human-vehicle-environment system integration, standard tool chain for virtual testing of autonomous vehicles, hybrid traffic simulation and testing under the penetration rate of different autonomous vehicles, and dynamic adaptive random generation mechanism of test cases. Establish a virtual test standard system for autonomous vehicles to provide strong support for the development of autonomous driving technology and industrial landing.
Reproduced from china highway journal, the views in the article are only for sharing and exchange, do not represent the position of this public account, such as copyright and other issues, please inform, we will deal with it in time.
-- END --