Research On Test Scenario And Accelerated Evaluation Method Of Automated Vehicles

In recent years,automobile intelligence has attracted more and more attention from consumers and has gradually entered People’s Daily life.Perfect testing and verification is an important part of the application of autonomous vehicles.Various safety problems existing in autonomous driving functions can be found through test data,so as to improve and upgrade the autonomous driving algorithm.The traditional test method requires a large number of road tests,so it is faced with such problems as long cycle,high cost,and difficult to reproduce scenarios,and it is difficult to meet the safety test requirements of high-level automatic driving system.Because of the above drawbacks of traditional testing methods,simulation testing has attracted great attention in today’s society.However,at present,there are still some problems in autonomous vehicle simulation test technology,such as rare key scenarios and low efficiency of test and evaluation.Based on the analysis and summary of the existing autonomous vehicle simulation evaluation methods,this paper focuses on the generation and evaluation of autonomous vehicle test scenarios.First of all,aiming at the problems of lack of track data and low scenario coverage of current automated driving road test,a scenario feature description method based on stochastic model was proposed,which could effectively describe the statistical characteristics of natural driving scenarios and increase the coverage of scenarios.On this basis,the least square method was used to fit the parameters in the stochastic model of the leading vehicle behavior for the following scenario.In front of the vehicle insertion scenario,Gaussian check scenario feature variable distribution convolution smoothing was adopted to avoid missing key test scenarios.The scenario feature description method ensures that the scenario complies with the natural driving statistical characteristics and the coverage of the test scenario is guaranteed.Secondly,aiming at the problem that the test scenarios generated by Monte Carlo method are too safe and lead to low simulation test efficiency,an automatic driving vehicle risk test scenario extraction method is proposed based on the combination of scenario exposure rate and scenario risk degree.This method tests the probability distribution of behavioral characteristics of traffic participants in the scenario through targeted migration.More dangerous scenarios can be generated that can affect the driving behavior of the tested autonomous vehicle to a greater extent.Simulation tests show that the proposed method can ensure the accuracy of the vehicle safety level test results in the accelerated test,and improve the efficiency of autonomous driving simulation test to more than 100 times.On this basis,aiming at the low efficiency of controller parameter optimization based on Monte Carlo method,a controller performance optimization method is proposed.This method integrates particle swarm optimization algorithm and importance sampling method,which can optimize controller parameters quickly.The simulation results show that the optimized controller can reduce the accident rate,conflict rate and discomfort rate.Finally,aiming at the problem of poor universality of test scenario library developed by specific automatic driving algorithm,an optimization method of automatic driving test scenario library based on Bayesian optimization was proposed.This method adopted the agent model with specific automatic driving function to build the test scenario library.According to the performance of the black box test object,the test scenario library was updated and iterated by Bayesian optimization.Search for test scenarios that have greater impact on black box test objects,and optimize and upgrade the test scenario library.Simulation results show that the proposed method can further improve the efficiency of black box object testing on the basis of importance sampling.