Vehicle Following Virtual Test Technology Based On Multi-dimensional Logical Scenario

With the rapid development of autonomous vehicles,autonomous driving technology proposes a new solution to problems such as traffic safety and traffic congestion.Traditional road testing cannot meet the testing needs of autonomous vehicles due to its test mileage and cost,so virtual simulation-based testing methods have emerged.However,how to build highfidelity test scenarios and how to discover critical scenarios in the scene parameter space to achieve accelerated testing and other issues need to be solved urgently.Car-following scenarios account for a high proportion of natural driving data and are prone to dangerous accidents,so this paper mainly studies the virtual simulation test problem based on vehicle following scenarios,and its research content is as follows:1.Multi-dimensional logic scene construction.In this paper,the characteristics of NGSIM dataset and car-following scene are combined to select appropriate critical scene parameters and deconstruct critical scene parameters in NGSIM.When refactoring scenarios,test cases are generated in a hierarchical structure of scene abstraction.Static scene elements are not considered in the NGSIM dataset,and this paper adds static scene elements when generating test scenes by combining the scene elements that affect the movement state of the car-following the scene.At the logical scene level,a multi-dimensional logical scene is constructed with three dynamic scene elements and static scene elements.2.Autonomous driving driver model.In this paper,an autonomous following algorithm is designed,the autonomous vehicle is controlled by the Intelligent Driver Model(IDM)and the Automatic Emergency Braking(AEB)algorithm,the time to collision(TTC)reaches the AEB set threshold as the switching logic,and the forward vehicle movement mode is a random front-vehicle model based on the Markov chain,and the car-following scene is built on the Prescan/Simulink joint simulation platform.3.Generate critical scenarios based on importance sampling.For dynamic scene construction,this paper uses a Gaussian mixture model to fit its multi-dimensional joint probability distribution,and generates test cases by Monte Carlo method based on Gibbs sampling.In order to improve the coverage of critical scenarios,TTC was used to generate more critical scenarios by using the importance sampling method,and eight typical test cases were generated for key scene clustering by K-means clustering algorithm.Finally,based on the autonomous driving following algorithm,the typical scenarios are evaluated on the Prescan/Simulink joint simulation platform,and these eight typical scenarios cover safety scenarios and dangerous scenarios,indicating that the importance sampling generation of critical scenarios is more critical in the scene space.4.Generate critical scenarios based on optimized search.In this paper,the simulated annealing algorithm optimizes the search for discrete scene parameter space,takes the minimum TTC in the scenario simulation process as the cost function,and searches for critical scenarios in the Prescan/Simulink joint simulation platform based on the autonomous driving following algorithm,and the results show that the number of critical scenarios explored by the optimized search algorithm is 2.15 times that of the random algorithm.