Research On Multi-mode Photorealistic Reconstruction Technology For Urban Smart Car

Scene efficient modeling technology is of great significance to the development of urban smart cars.Algorithm development and verification on the pre-constructed rich city simulation scenarios will greatly improve the development efficiency of autonomous driving algorithms.Traditional scene modeling methods rely entirely on manual implementation for data collection and modeling,which is inefficient,and it is difficult to construct details,resulting in scene patch redundancy affecting rendering efficiency.This paper relies on urban smart car data,and design algorithms to efficiently construct a simplified model of urban scenes.The main work of this article is as follows:A scheme for extracting the coordinate sequence of building facades from a point cloud continuously collected by urban smart cars is proposed.The vehicle-mounted point clouds are massive and disorderly.In view of the large number of ground objects,this paper first cuts point clouds of ground objects other than buildings into fragments,and then the spatial scale information is used to extract multiple building point cloud clusters,to avoid the problem that the classification threshold is difficult to set accurately when the collection angle is limited.In this paper,the building model is equivalent to a continuous rectangular fa?ade to reduce the weight of the model,and the main body of the building is retained by over-segmenting the plane.A search and merge rule is set to restore the continuous structure of the building,and the coordinate sequence of the rectangular patch is finally obtained.An idea to automatically search for the texture of buildings from images continuously collected by urban smart cars is proposed.The coordinate points extracted from the building point cloud are projected onto images collected at different positions,and the texture sequence corresponding to each patch is obtained by cropping.Aiming at the phenomenon of building texture occlusion,a texture restoration scheme based on prediction is designed.Combining the development results of deep learning,first use semantic segmentation to classify pixels,and then binarize the building category with the rest of the categories to obtain an image occlusion mask to form a one-to-one mapping to indicate the area to be repaired,and then train the gating-based convolution The image repair network selects the best optimal texture for repair according to the sizeof the occlusion area and the clarity of the image.In this paper,a system for batch generation of buildings and roads is designed to automatically generate white molds and texture maps of buildings,and automatically generate roads in the driving direction of intelligent vehicle owners.Relying on the longitudinal project,the six-free driving simulator was used as the experimental platform to design and develop a visual simulation system.The scene was imported into the driving simulator,and the frame rate and collision detection of the scene were verified,which proved the availability of the proposed scheme.