The Lightweight Reconstruction Of Urban Scene Under Poor Quality Data

In recent years,with the vigorous development of virtual reality,video games,geographic information systems,autonomous driving and other fields,and the recent proposal of the concept of "metaverse",the demand for 3D models of urban scenes in all walks of life is increasing.Due to the complex environment of urban scenes,high density of buildings,serious occlusion and self-occlusion,the actual collected images and point cloud data often have a lot of noise and data missing,which brings great difficulties to the reconstruction task.On the other hand,the scale of urban scenes is huge,and how to process massive urban scene data quickly and efficiently while obtaining models with ideal quality is also a huge challenge.Based on the above background,this paper studies the reconstruction of urban scenes.Although the urban scene environment is complex and in large-scale,we observed that most of objects in the urban scene,especially the dominant buildings,mostly have simple and regular plane structures.The details are also not important in applications like live maps and autonomous driving.Due to the low quality of the data,it is quite unrealistic to reconstruct the detailed model.Therefore,this paper focuses on the light-weight reconstruction of urban scenes from low-quality data.The light-weight is not only reflected in the simple geometric structure of the reconstructed model,which is very friendly to rendering and storage,but also reflected in the fast and efficient reconstruction process.In order to improve the accuracy of reconstruction results and deal with noise,uneven and incomplete data,this paper combines data from multiple modalities(point cloud,line cloud and image)for reconstruction to improve the robustness of the algorithm.Specifically,this paper has done the following researches:1.We proposed two algorithms to extract planes from 3D line clouds for different scenarios.The graph-clustering-based plane extraction algorithm is proposed for the first time to convert the line cloud plane extraction problem to a clustering problem.It ensures a more robust plane extraction result by fully considering the global relationship of line segments.The Manhattan-world-assumption-based plane extraction algorithm improves the random selection strategy for seed plane of RANSAC,and reduces the generation of virtual planes by adding Manhattan-world constraints,which also has high operating efficiency.2.We proposed the concept of cuboid representation to express the buildings with Manhattan-world structure in a lightweight way.Starting from the corner-a significant building feature,we fit the cuboid representation from sparse point clouds to reconstruct Manhattan-world buildings.The method directly reconstructs from sparse point clouds without the time-consuming dense reconstruction process,and the reconstruction speed is fast.The method can deal with the building data with missing backsides by fitting the point cloud from the corner s.The outputted lightweight model has prominent advantages in rendering and storage.3.We proposed an image-based algorithm for registering cuboid representations.The algorithm further improves reconstruction accuracy under low-quality data by backprojecting the cuboid representation onto the image,finding corresponding edges for the projected edges on the image,and adjusting the size of the cuboid for alignment.This article also realizes the texture mapping of the cuboid representation,and finds regular and uniform texture sources for the lightweight model,which further improves the realism and practicability of the model.Experiments on different data sets have verified that the method in this paper can achieve rapid reconstruction of buildings in the case of sparse point cloud data and face missed data of buildings,and obtain a model with a high degree of restoration.Compare to other methods,our method has advantages in both reconstruction speed and results.