A hierarchical energy minimization method for building roof segmentation from airborne LiDAR data

Abstract

This paper presents a reliable and accurate method for building roof segmentation from airborne LiDAR data. In order to obtain the optimal results in both object level and pixel level, three energy minimization procedures are conducted consecutively in a hierarchical way. Firstly, an active multi-plane fitting method is conducted to obtain reliable initial segmentations. Then, the coarsest energy function composed of both the plane fitting errors in pixel level and the number of plane hypotheses in object level is minimized to obtain the optimal label space. Next, energy function composing of plane fitting errors and spatial smoothness between neighboring planes is minimized to obtain the optimal segmentation results. Finally, by taking prior knowledge of building roof structure into consideration, the optimal plane parameters for the segmented plane hypotheses are obtained by minimizing energy function of the structural adjusted plane fitting errors. Two real LiDAR data sets with different point densities and different building styles are used to evaluate the performance of the proposed method, and experimental results demonstrate that the proposed method is fast, stable, and reliable for accurate building roof segmentation from airborne LiDAR data.