Research On Data Generation Technology Of Urban Autonomous Vehicles Test Based On Augmented Reality

Simulation system has become an essential component in the development and validation of autonomous driving technologies.The existing simulation test systems can be divided into two categories.The one is the simulation software based on the traditional vehicle dynamics models.The other is to use game engines or high-fidelity computer graphics(CG)models to create driving scenarios.However,both of them need manual tasks that would be costly and time-consuming.In addition,the simulation data has an obvious style of computer synthesis,which is obviously different from the data collected by actual sensors.Using this kind of data for autonomous driving simulation testing will result in performance degradation.This paper focuses on improving the data authenticity and diversity of the autonomous driving simulation test system.This paper solved the problems of image inpainting,image enhancement and LIDAR simulation.Finally the paper realized autonomous driving test simulation based on augmented reality technology,built a software-in-the-loop platform,tested the authenticity of the test data generated by the data-driven method,and analyzed and verified the superiority of the autonomous driving simulation test system based on the augmented reality.This article established an image data enhancement system.Based on the real scene data collected by the autonomous vehicles,the data set was divided into static scene data and dynamic obstacle data by creaging semantic masks.A two-stage generation network was designed to inpaint the holes left by removing obstacles in the static scene data.Based on the graph convolutional neural network,a depth model corresponding to the threedimensional model was generated from the RGB image,and a large number of threedimensional models were generated using the separated obstacle data.With the help of Cycle,a renderer software,the images were rendered and enhanced,and the threedimensional models were enhanced and rendered to the restored static scene image to obtain enhanced image data.In addition,in order to increase the diversity of data,the conversion of image from day to night based on the generative confrontation network was realized.According to the principle of LIDAR,a simulation LIDAR model was designed and implemented.Using the deep learning algorithm,the obstacles in point cloud data collected from real world were moved so the pure scene point cloud data was obtained.By analyzing the distribution rule of all kinds of obstacles in different scenes.Based on the statistical analysis of the distribution of various obstacles in different scenes,the probability map of the distribution of obstacles in each scene is established to guide the placement of moving obstacles.The cube map model was designed,and the depth map,normal vector map and texture map were obtained by projecting the scene point cloud to generate the enhanced LIDAR point cloud data.Using the simulation test data generated by the method proposed on the paper,the practicability of this method in the simulation test process of automatic driving is verified.By comparing the enhanced simulation data generated by this method with the existing simulation data,the experimental results show that the simulation data obtained by this method is closer to the real scene data and has higher authenticity.