Clustering Based Planar Roof Extraction from Lidar Data

An approach to generate 3-D models of buildings from lidar data collected from an urban setting is presented. The present research focuses on extracting roof structures from a point cloud of a building using a combination of datamining techniques. To extract the roof structure, an assumption of planarity has been made, i.e. it is assumed that the roof can be modeled by a set of planar segments. The task then is to map each of the building point to a planar segment. We present a method that first separates points lying on or near breaklines, by which we mean points that are near the intersection of two planar surfaces or points that are near step-edges. Treating these points to be ambiguous, (i.e. they belong to more than one plane), we separate them from points that are exclusively planar. We then use neighborhood functions to determine what we call planar patches, and their direction cosines using their eigenvalue and eigenvector characteristics. Then an unsupervised data clustering technique to cluster these planar patches into one single plane is described. Most clustering techniques require that the number of clusters be known. In our case these clusters represent roof planes. We present a way to determine the number of roof planes (clusters) by using an iterative combination of k-means and density based clustering methods.