Large-Scale Accurate Reconstruction of Buildings Employing Point Clouds Generated from UAV Imagery

摘要

High-density point clouds are valuable and detailed sources of data for different processes related to photogrammetry. We explore the knowledge-based generation of accurate large-scale three-dimensional (3D) models of buildings employing point clouds derived from UAV-based photogrammetry. A new two-level segmentation approach based on efficient RANdom SAmple Consensus (RANSAC) shape detection is developed to segment potential facades and roofs of the buildings and extract their footprints. In the first level, the cylinder primitive is implemented to trim point clouds and split buildings, and the second level of the segmentation produces planar segments. The efficient RANSAC algorithm is enhanced in sizing up the segments via point-based analyses for both levels of segmentation. Then, planar modelling is carried out employing contextual knowledge through a new constrained least squares method. New evaluation criteria are proposed based on conceptual knowledge. They can examine the abilities of the approach in reconstruction of footprints, 3D models, and planar segments in addition to detection of over/under segmentation. Evaluation of the 3D models proves that the geometrical accuracy of LoD3 is achieved, since the average horizontal and vertical accuracy of the reconstructed vertices of roofs and footprints are better than (0.24, 0.23) m, (0.19, 0.17) m for the first dataset, and (0.35, 0.37) m, (0.28, 0.24) m for the second dataset.