Research On Virtual And Real Scenes Fusion Technology For Autonomous Driving Test

Autonomous driving test is a key step to ensure the safety of autonomous driving technology.At present,the real test drives of autonomous driving are expensive,not reproducible and inflexible.While virtual simulation overcomes the above shortcomings of the real test drives,it also brings the degraded performance due to its low simulation accuracy of the vehicle dynamics models and sensor models,which causes the verified self-driving technology by simulation tests can not directly come to the next step.This article focuses on the fusion technology of virtual and real scenes for autonomous driving testing.Based on the scene construction technology and sensor simulation technology,a virtual autonomous driving scene is built.The image fusion and point cloud fusion of the virtual and real scenes are realized,and their effectiveness and superiority are verified by data sets comparison.Firstly a virtual scene of autonomous driving is built.After comparing the existing autonomous driving simulation platforms,Air Sim simulator is selected as the virtual scene construction platform considering the research needs.Based on the static environment and dynamic traffic scene construction technologies,the implementation methods of lighting and weather effects as well as the common sensor simulation technologies,the virtual scene for autonomous driving is built.Next the image fusion technology for the fusion of virtual and real scenes is studied.The imaging model formula of the monocular camera is derived and the camera simulation is carried out based on the camera principles.The fusion of virtual and real images based on the Poisson fusion algorithm is implemented and improved aiming at the problems of artifacts and bleeding.By comparing the authenticity of the image data generated by the copy method,Poisson fusion algorithm and the improved Poisson fusion algorithm proposed in this paper,the superiority of this algorithm is verified.Then the point cloud fusion technology for virtual and real scenes fusion is researched.According to LIDAR working mechanism,a LIDAR simulation model is built.The correspondence between point cloud and image is used to realize the mapping from point cloud to image pixel.The closest strategy is selected for the fusion of virtual and real point clouds and its fusion results show satisfactory results and good real-time performance.Finally,the image fusion and point cloud fusion algorithms are verified through the data set comparison.The simulation images,the real images and the images generated by virtual and real fusion proposed in the paper are made into three data sets.Their target detection evaluation indicators are compared through YOLO-v3,a target detection network model.The results show that the target detection indicators of the virtual and real fusion image are better than those of the simulation images.The semantic segmentation evaluation indicators of three data sets made by the point cloud generated by the simulator,the real point cloud and fused point cloud are compared after training by the point cloud segmentation model Squeeze Seg.The results show that the point cloud generated by the fusion technology in this paper is better than virtual point cloud.Pre Scan/Air Sim joint simulation illustrates the application scenarios of the algorithms proposed in this paper.