| Résumé : |
(Auteur) Various applications demand realistic 3D city models. For urban planning, analyzing in a 3D virtual reality world is much more efficient than imaging the 2D information on maps. For public security, accurate 3D building models are indispensable to make strategies during emergency situations. Navigation systems and virtual tourism also benefit from realistic city models. Manual creation of city models is undoubtedly a rather time consuming and expensive procedure. On one hand, images are for long the only data source for geometric modelling, while recovering of 3D geometries is not straightforward from 2D images. On the other hand, there are enormous amounts of objects (for example buildings) to be reconstructed, and their structures and shapes show a great variety. There is a lack of automated approaches to understand the building structures captured by data. The rapid development of cities even adds to the cost of manual city model updating. In recent years, laser scanning has been proven a successful technology for reverse engineering. The terrestrial laser point clouds are especially useful for documenting building facades. With the considerable high point density and the explicit 3D coordinates of terrestrial laser point clouds, it is possible to recover both large structures and fine details on building facades. The latest developments of mobile laser scanning technology also make it more cost-effective to take large-scale laser scanning over urban areas.
This PhD research aims at reconstructing photorealistic building facade models from terrestrial laser point clouds and close range images, with a largely automatic process. A knowledge base about building facade structures is established first, where several important building features (wall, door, protrusion, etc.) are defined and described with their geometric properties and spatial relationships. Then constraints for feature extraction are derived from the knowledge base. After a laser point cloud is segmented into planar segments by surface a growing segmentation algorithm, each segment is compared with the feature constraints to determine the most likely feature type for each segment. The feature extraction method works fine for all facade features except for windows, because there are usually insufficient laser points reflected from window glass. Instead, windows are reconstructed from the holes on the wall features. Then outline polygons or B-spline surfaces are fit to all feature segments, and the parts without laser points are hypothesized according to knowledge. A complete polyhedron model is combined from both fitted and hypothesized outlines.
Since laser data contains no colour information, the building models reconstructed from only laser data contain only geometric information such as vertices and edges. To obtain photorealistic results, textures must be mapped from images to the geometric models. The fusing of laser points and image requires accurate alignment between laser space and image space, which is accomplished after a semi-automated process. Because of the limitations of modelling methods, the geometry model reconstructed from laser points may contain many errors which would cause poor texturing effect. Therefore, significant line features extracted from images are compared with the initial model's edges, and necessary refinements are made to correct the model errors, or at least make the model edges consistent with the image lines. Finally, in the texturing stage, the texture of each model face is selected automatically from multiple images to ensure the optimal visibility. Texture errors caused by occlusions in front of a wall are also removed by analyzing the locations of the wall, the occlusions and the camera position.
Experiments with three data sets show that building reconstruction are considerably accelerated by the presented methods. Our approach is more than 10 times faster than the traditional approach when reconstructing the same buildings, and the models by our approach contain more fine details such as doors and windows. The reconstruction of wall facades and roofs are fully automatic, while some manual interactions (48 percent of the total reconstruction time) are still required for editing the fine details. It should also be faster to make global statistics (number of floors, number of entrances, etc.) and modifications (deriving models with a lower level of detail, applying pre-defined textures, etc.) later on to our models, since different model parts have been associated with the semantic labels. While the reconstruction efficiency is improved by our approach, the visualization effects of our models are also comparable to the models by the traditional approach. The future work will focus on improving the knowledge base and developing a fully automated camera parameter estimation procedure. The completeness and adaptability of the knowledge base will be especially important for the further automation of our reconstruction approach. |
Knowledge based building facade reconstruction from laser point clouds and images (2010)
