Realistic 3D city models are used in important applications such as flight simulators, war game simulations, security & surveillance, entertainment, etc. Many of these applications require that the model be as close as possible to the real world that it simulates, both in terms of geometry and texture. Recently, aerial LIDAR (Light Detection and Ranging) has gained popularity as a way to quickly collect 3D information about a site. LIDAR scanning of a site produces dense, unorganized points that require further processing to identify buildings, trees, and bare ground.

We present a fully automatic approach to creating 3D geometric models of the buildings and terrain from LIDAR data. Our objective is to create compact, watertight geometric models of buildings that fit as close as possible to the original LIDAR points using the minimum number of triangles, as opposed to a dense mesh. We propose to model the class of buildings that can be constructed by combining several simpler prismatic models. Such primitives are also amenable to automatic semantic interpretation as well as intuitive interactive editing.

We first segment the building roofs, trees, and terrain from the LIDAR points. We next fit local planar patches to the segmented roof points that are then grouped together to identify individual faces of the roof as polygons. By analyzing their proximity to each other, we construct a planar graph that represents the topology of the roof structure, including adjacency, symmetry, and orthogonality constraints between roof faces. Instances of simple prismatic models in the scene, such as hip roof configurations, can be identified by subgraph matching. The geometric parameters of these models are then refined based on the original point cloud. Using the above approach we can automatically model complex roofs by combining simpler prismatic objects.