[Abstract] In order to generate test cases covering the complex and changeable real traffic operation process in the corresponding scene in the scenario-based test of autonomous vehicles, multiple actual samples of the vehicle entering the scene are extracted from the highD data set, and the motion parameters and participation are analyzed. The positional relationship between vehicles is used to establish a description model of the vehicle entry scene, and the risk level of the solution is evaluated based on the collision time of the entry point. Combined with the distribution of parameters in the description model, a Monte Carlo method is used to generate test cases. The results show that the generated vehicle entry test cases can cover all risk levels and can better support autonomous driving testing.
Keywords: Autonomous driving test highway vehicle entry test scenario test case Monte Carlo method
1 Introduction
Autonomous vehicles provide new ways to solve problems such as “traffic accidents, traffic congestion, environmental pollution, and energy shortages” [1]. A scientific and complete testing, verification and evaluation system is crucial to improving the research and development efficiency of autonomous vehicles, improving relevant laws and regulations, and promoting the development of intelligent transportation [2]. Since 2009, Google’s self-driving cars have conducted more than 5.6 million kilometers of public road testing and billions of kilometers of virtual testing [3]. UBER, Volvo, Tesla and other companies have also conducted a large number of actual road autonomous driving tests [4]. However, multiple accidents represented by the UBER self-driving test vehicle accident show that before self-driving technology matures, actual road testing carries high safety and social risks. In addition, unlike the human-vehicle binary independent test of traditional cars, autonomous vehicles are a strongly coupled system of human-vehicle-road-environment. Traditional road site testing can no longer meet the testing needs of autonomous vehicles.
Scenario-based virtual testing technology has flexible test scenario configuration, high efficiency, strong repeatability, safe process, and low cost. It can realize automatic testing and accelerated testing. At the same time, the virtual testing system can simulate dangerous or difficult extreme scenarios in actual vehicle testing. [4-5], which greatly reduces the difficulty and risk of testing
and reduces the workload. Therefore, scenario-based virtual testing has become an indispensable and important link in the testing and evaluation of autonomous vehicles. Researchers have made a lot of efforts in constructing test scenarios [6-7]. Xia et al. used the analytic hierarchy process to select important factors in the test scenario and constructed virtual test scenarios through the combined testing method [8-9]. Zofka et al. used multiple A variety of sensors collect data, and a method is proposed to construct high-risk test scenarios using real traffic data in a virtual environment [10].
In the current field of autonomous driving, there is no clear and unified definition of "scenario". At the same time, existing research focuses on the analysis and construction of autonomous driving test scenarios, and lacks research on the design and generation of test cases. Therefore, this article proposes a definition of test scenarios, and also proposes a test case design and description method based on actual traffic scenarios, and verifies it by analyzing highway vehicle entry scenarios, using highD data sets to implement test case design, and using Monte Carlo Method (Monte Carol Method) generates test cases.
2 Autonomous driving test scenarios and use cases
2.1 Test scenario concept
There are different scenario definitions in the existing research on autonomous vehicle test scenarios. According to the classification of scenes, Elias Rocklage et al. proposed that "a scene is a combination of the observed vehicle itself and other static or dynamic objects on a section of road" [11]. Considering the continuity of time, Gelder defined the scene as the combination of observing the vehicle's own movement and the static and dynamic environment within a period of time, highlighting the continuity of the scene [12]. Hala Elrofai et al. proposed that "a scenario is the continuous change of the dynamic environment around the test vehicle within a specific time range, including the behavior of the test vehicle in that environment. Based on summarizing the existing scenario definitions, this paper proposes a method suitable for
testing Scenario definition of scene construction. For autonomous vehicle testing, the test scenario is the process in which the test object interacts with other relevant traffic participants in the traffic environment to achieve driving intention. Among them, driving intention refers to the test object completing the change of its own motion state, For example, complete a free lane change or turn at a traffic intersection. The traffic environment is the static environment such as roads, weather, lighting, etc. where the test object operates. Traffic participants refer to the scene elements in the scene that have an impact on the motion state of the test object, such as with the test vehicle. Vehicles, pedestrians, etc. with conflicting driving routes can be ignored in the observation and analysis of the scene. Elements that have no impact on the movement of the test object can be ignored. 2.2 Definition of test cases In self-driving car testing, the scene is a type of
real
traffic environment. An abstract description of the running process. A test case is an executable instance of the corresponding test scenario. A complete test case description includes the test scenario description to which the test case belongs and the test case element information. Among them, the test scenario description provides an abstraction of the test environment and test process. Description: The test case element information provides the values of each element in the test environment and the starting status, status change process and expected operating results of the traffic participants in the test.
This article takes vehicle entry in a highway environment as an example to describe test cases. Since the traffic scene in the highway environment is relatively simple, the test scenario is the vehicle cutting in from adjacent lanes on a straight road. The vehicle under test is driving along the line in lane 2, and the target vehicle is driving smoothly and straightly in lane 1 adjacent to the vehicle under test. At the cutting start time T0, the target vehicle is located in front of the vehicle under test and starts cutting into lane 2. At time T 1, the target vehicles cut into lane 2 and drive to time T 2. During the cut-in process, there is a risk of collision between the two vehicles. The vehicle entry process is shown in Figure 1, and the test case elements are shown in Table 1.
2.3 Test case construction
Real traffic data contains a variety of traffic scenarios. The method of generating test cases in this scenario based on real traffic scenarios is shown in Figure 2. Real traffic scenes can be decomposed into static traffic environments and movement processes of traffic participants. The static traffic environment covers road conditions, roadside facilities, weather and lighting conditions and other environmental conditions that affect vehicle operation, from which the static environmental elements in the test scene can be extracted. The movement process of traffic participants is the process in which the observed object interacts with other traffic participants in a static environment and realizes its own movement intention. The movement process of traffic participants is the core of the test scenario. In order to generate test cases, it is necessary to model the movement process and analyze a large number of examples of real scenarios, and estimate the description parameter distribution of the model, thereby realizing the transformation from real scenarios to standard test scenarios. On this basis, for different test objects, test methods and test content, different values of description parameters can be flexibly assigned, and corresponding test cases can be generated by combining different static elements to adapt to different testing needs.
