3D Point Cloud Generation, Consolidation And Reconstruction For Architecture

With the rapid development of laser scanning, internet images, remote sensing, vir-tual reality technology and proposal of”Virtual Earth”, big cities of the world are rushingto reconstruct their own virtual3D cities. Large-scale3D virtual scene is fundamental tolots of applications, such as urban planning, city surveillance, intelligent transportation,3D maps, protection of city’s cultural heritage, realistic video games, virtual battlefield-s, emergency management of natural disasters and terrorist attacks, pollutant dispersionmodeling and simulation etc. Moreover, Large-scale3D virtual scene will be broadlyapplied to new applications as the development of information technology.Architectures are main components of modern cities, and research on reconstructionof architectures is a hot topic in city scenes reconstruction. To overcome the deficienciesof existing methods on3D points generation, processing and reconstruction, this thesisfocuses on the following issues: unorganized architecture photos sorting for urban re-construction, multi-view stereo correspondence, consolidation and denoising for LiDARpoints by vehicle-based laser scanner, and grammar-based3D facade segmentation andreconstruction. The major contributions are described as follow:1. In order to avoid the registration failure caused by self-similarity of architec-ture photos and accelerate the structure-from-motion, we proposed an efficient methodfor sorting unorganized architecture photo collections. This method can quickly create apiecewise-planar facade model, an image connected graph, and an initial pose for eachcamera. This method first estimates single-view, generates piecewise planar geometryfrom each photo, then merges these single-view models together in an analysis phase thatreasons about the global scene geometry. The result is not only useful in itself as an ap-proximate scene model, but also represents a good initialization for structure from motionand multi-view stereo techniques from which refined models can be derived, at greatlyreduced computational cost compared to prior techniques.2. In order to improve the result of stereo correspondence induced by simple neigh-bourhood system, we proposed a new stereo correspondence method which integratesimage segmentation and high-order markov random field. We first introduced a mod-el which can map high-order clique to graph cut model, then apply high-order MRF tostereo correspondence. We segmented image into several segments based on colors of pixels, and assumed that the pixel in the same segment has the similar disparity or depthvalue. In order to ensure that the pixels in the same segment had the similar disparityand depth value, we introduced high-order clique in the energy function of ordinary MRFframework.3. Vehicle-based laser scanning devices extract depth through active sensing, allowfast scanning of urban scenes. Such rapid acquisition incurs imperfections: large region-s remain missing, significant variation in sampling density is common, and the data isoften corrupted with noise and outliers. We proposed a method for points consolidationinspired by non-local filtering. This method utilizes the self-similarity and repetitionsof architecture, represent repetitions as base-geometry, apply points consolidation, filter-ing, and completion by base-geometry. We tested our method on a variety of buildingsscans obtained by vehicle-based scanning devices, the consolidated scans are complete,uniformly distributed, clean, and can be used as input of reconstruction.4. Traditional surface reconstruction methods are widely used for irregular objec-t modeling, but they perform poorly on some man-made objects, such as architecture,especially when3D point cloud is incomplete. We proposed a method for reconstruc-tion of3D polygonal models from3D point cloud. This method first segments3D pointcloud into several depth layers, applies a grammar system to segment each layer into2Drectangles according to3D points distribution. Moreover, rectangles between layers aregeneratedbyextrusionaccordingto3Dpointsdistribution. Thefinalmeshofthearchitec-ture is composed of those3D rectangles. This method mitigates limitations of traditionalsurface reconstruction methods on architectures, introducing less polygons.