Abstract
Abstract. In this work, an automated approach for 3D building roof modelling is presented. The method consists of two main parts, namely roof detection and 3D geometric modelling. For the detection, a combined approach of four methods achieved the best results, using slope-based DSM filtering as well as classification of multispectral images, elevation data and vertical LiDAR point density. In the evaluation, the combination of the four methods yields 94% correct detection at an omission error of 12%. Roof modelling is done by plane detection with RANSAC, followed by geometric refinement and merging of neighbouring segments to clean up oversegmentation. Walls are then detected and excluded, and the roof shapes are vectorised with the alpha-shape method. The resulting polygons are refined using 3D straight edges reconstructed by automatic straight edge extraction and matching, as well as 3D corner points constructed by intersection of the 3D edges. The results are quantitatively assessed by comparing to ground truth manually extracted from high-quality images, using several metrics for both the correctness and completeness of the roof polygons and for their geometric accuracy. The median value of correctness of the roof polygons is calculated as 96%, while the median value of completeness is 88%.
Highlights
1In this work, we focus on the 3D building roof modelling
The fourth approach is fully based on the LiDAR data, so it does not include any drawbacks from image data, but has weaknesses, especially on building outlines, which are detected as trees
The spectral problems are eliminated using LiDAR data and height information, while the weakness of LiDAR data surface discontinuities is handled by using 3D edges from image matching
Summary
The test site is in Vaihingen, Germany. The dataset has been provided from the DGPF camera evaluation project Reference vector data has been collected by stereo measurement using the DMC images, with better than 15 cm accuracy. It consists of the planar roof polygons, which are larger than 25 m2 for 164 buildings. A slope-based progressive morphological filtering method (Zhang et al, 2003) has been used to reduce the DSM to DTM. Where is the height difference threshold, is the initial elevation difference threshold which approximates the error of DSM measurements (sigma of 0.2 m), is the maximum elevation difference threshold (3 m), s is the predefined slope (0.10), is the grid size and is the filtering window size (in number of cells) at the th iteration.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
