Research On Test Method Of Autonomous Driving System Based On Generative Adversarial Network

In recent years,traffic accidents involving autonomous driving systems(ADS)have occurred frequently,indicating that there are serious reliability and safety problems in the software of autonomous driving systems,and it is of great research significance to fully and effectively test them.The investigation of the existing testing methods of autonomous driving systems found that many testing methods still have problems such as low quality of synthetic test cases,which do not fully cover real driving scenarios and test predictions.Therefore,this paper uses the image scene conversion model based on Generative Adversarial Network(GAN)to synthesize real and high-quality driving images,and applies a new automatic driving system test method to detect the potential defects of autonomous driving models,so as to effectively ensure the quality of autonomous driving models.The main research contents of this paper are as follows:(1)This paper proposes a new image scene transformation model based on generative adversarial network,which can flexibly output driving pictures in different scenarios(such as sunny days,rainy days,foggy days,etc.)in multiple domains,and solve the problem of image scene transformation with rich categories and highly complex semantic structure.The combination of generative adversarial network,attention module,and scene segmentation module enables the model to correctly identify and transform the Region of Interest(ROI)and keep other regions unchanged.In order to further improve the diversity of outputs,a new type of regularization loss is proposed to suppress potential noise.In addition,to avoid mode collapse due to lack of noise constraints,a noise separation block is embedded in the discriminator.Experiments show that compared with six models such as Cycle GAN,UNIT,MUNIT,etc.,the average FID score of the generated image is increased by about 7.25%,and the average score of KID is increased by about 19%,so the model can generate images with better visual effects in any scene.(2)In order to maximize the "boundary" of the autonomous driving system and cause the autonomous driving model to produce more error outputs,this paper proposes a new test method for the autonomous driving system,which combines the simple image transformation and the GAN-based image scene transformation model to maximize the coverage of neurons to generate test cases in different scenarios to cover more extreme driving scenarios of the automatic driving system.Construct a loose metamorphosis relationship to automatically determine the wrong behavior of the autonomous driving system to solve the test prediction problem.Experiments show that this method can not only detect a large number of false steering behaviors of the autonomous driving system,but also retrain the autonomous driving system to increase its robustness by using the synthetic test case set.Based on the above research and software engineering ideas,the autonomous driving system test platform is designed and implemented,and users can realize image scene conversion and generate conversion result report on the platform,realize automatic driving system testing and generate visual test results.After testing,the functional modules of the platform operate normally,simplifying the testing process of the autonomous driving system and facilitating the research of relevant researchers.