BUILDING ROOF BOUNDARY EXTRACTION FROM LiDAR AND IMAGE DATA BASED ON MARKOV RANDOM FIELD

Abstract

In this paper a method for automatic extraction of building roof boundaries is proposed, which combines LiDAR data and high-resolution aerial images. The proposed method is based on three steps. In the first step aboveground objects are extracted from LiDAR data. Initially a filtering algorithm is used to process the original LiDAR data for getting ground and non-ground points. Then, a region-growing procedure and the convex hull algorithm are sequentially used to extract polylines that represent aboveground objects from the non-ground point cloud. The second step consists in extracting corresponding LiDAR-derived aboveground objects from a high-resolution aerial image. In order to avoid searching for the interest objects over the whole image, the LiDAR-derived aboveground objects' polylines are photogrammetrically projected onto the image space and rectangular bounding boxes (sub-images) that enclose projected polylines are generated. Each sub-image is processed for extracting the polyline that represents the interest aboveground object within the selected sub-image. Last step consists in identifying polylines that represent building roof boundaries. We use the Markov Random Field (MRF) model for modelling building roof characteristics and spatial configurations. Polylines that represent building roof boundaries are found by optimizing the resulting MRF energy function using the Genetic Algorithm. Experimental results are presented and discussed in this paper.