| In recent years,thanks to the development of deep learning,the research and de-velopment of autonomous driving has also made great breakthroughs.Many domestic and overseas companies have developed their own autopilot system frameworks,such as Google’s Waymo,Baidu’s Apollo and Tesla’s Autopilot.The most well-known Autopilot has been successfully deployed in real-life commercial Tesla sedan.in.Although all companies have advertised that the safety of today’s autopilot systems is high enough,several autopilot accidents in the past year or two have caused concern about the safety performance of autopilot systems.such as traffic accidents on Tesla cars and Uber autopilot cars in 18 and 17 years.Accident analysis show that all the reasons were caused by different degrees of misjudgement of the automatic driving system.In order to improve the safety level of automatic driving system,many researches on automatic driving test have been done in the industrial and academic community.The most important work is the DeepXplore[1]and DeepTest[2]method,that is,a test framework that can automatically generate test cases for the automatic driving system.The above method reduces the cost of collection and generation of test cases for automatic driving system,and improves the stability of the system.However,the works above also have serious defects,that is,the generated test cases are much different from the real scene data,and the synthetic road picture is not true.In this paper,we have done a lot of empirical research on the existing in-depth learning technology,mainly including confrontation generation network and image style conversion technology.In order to explore which in-depth learning techniques are most effective for automatic driving test case generation,that is,the synthesis of driving scene pictures.In Chapter 2,we introduce a metamorphic testing framework based on DeepTest testing framework,which synthesizes test cases using antagonistic generation network technology.Chapter 3 and Chapter 4,aiming at improving the quality of test cases,respectively,implement antagonistic generation network technology and neural style migration technology for image style transformation.Empirical research,the fifth chapter makes statistics and analysis of the existing experimental results,and draws a corresponding summary.The main contributions of this work are:1.The first empirical study of applying large-scale deep learning technology,mainly generative adversarial network and neural style transer technology,to auto-generate test case in autonomous drving system.2.A large number of experimental data in empirical research were statistically analyzed,and relevant conclusions were drawn.It is pointed out which deep learning techniques are suitable as test case generation techniques for autopilot systems,and the performance comparison of each technology under different evaluation indicators is given.3.Expanded the DeepRoad test framework. |
