CNN-Based End-to-End Unmanned Driving Simulation in Curved Scene

Abstract

Nowadays, the research on autonomous driving technology is more and more in-depth, and the purpose of the research is that autonomous vehicles can achieve safe driving in various driving scenarios. In order to improve the existing autonomous driving scene data set, this paper proposes an end-to-end scene reality simulation implementation method based on CNN. This method is used to train an end-to-end model based on vision sensors for cornering scenes, establishing a direct mapping relationship between vision sensors and driving behavior. In this experiment, the Udacity simulator autonomously collects the data set of the curve scene, and after cutting and normalizing the original data, an end-to-end model with a low loss value is trained, and then the image data is divided into boxes. , Gaussian, median and bilateral filtering operations, the results show that filtering the data can quickly reduce the network loss value. Finally, comprehensive analysis can prove that this method is a high-performance and efficient scene implementation method.