Automatic Building Boundary Extraction from Point Cloud Data

Abstract. This paper presents a methodology and solution for extracting building footprints from three-dimensional point cloud data. The extraction and normalisation of building boundaries represent a fundamental approach for three-dimensional modelling of buildings for urban mapping. The automatic building boundary extraction utilising dense point cloud data has emerged as a prominent and labour intensive subject in lidar and photogrammetric point clouds. The methodology comprises three principal stages: Roof Face Plane Detection, Boundary Detection, Outline Smoothing, and Footprint Generation. The building point clouds are segregated from surface point clouds by filtering the bare earth model and removing above-ground objects other than buildings. Rooftop structures are generated by roof face plane detection. An extracted rooftop point cloud is subsequently traced to produce a series of corner points. The regularisation phase is conceived to extract corners from the irregular boundary and yield a polygon that accurately represents a rectilinear building footprint. To assess the performance, centroids of manually obtained building footprints are compared with centroids of automatically generated footprints. Euclidean distance between centroids of up to 2m is achieved for 90% buildings derived from the airborne lidar point cloud. Most buildings have an Euclidean distance variation between 4m and 8m for high- resolution stereo satellite data. The footprints generated from airborne point cloud data exhibit a smooth texture and closely resemble the manually derived building footprints in comparison to those obtained from satellite point clouds. The results indicate that automatic building footprints predominantly rely on the density of the point clouds.