RECONSTRUCTION OF BUILDING MODELS WITH ROOF PATCH CLASSIFICATION

摘要

Building models are indispensable for the representation of the real world in 3D geographic information systems. The proposed paper combines the distinct edges from aerial imagery and accurate elevation from lidar data to reconstruct the building models. Most of the man-made buildings are constructed with planes and straight lines so this paper aims at the plane roofs. First, the lidar point clouds on the roof are extracted to form the roof patches. The connectivity among the roof patches is analyzed. We classify the roof planes to three types, namely, flat roof patches, mono-pitched roof patches and multi-pitched roof patches. The roof structure lines exist in the multi-pitched roof patches. Therefore, we detect the zone of roof structure lines and the boundaries of three different roof patches are tracked in the object space. The zone of roof structure lines and the initial boundaries are projected to the imagery for buffering the zones of structure lines. Next, we detect the edges with Canny detector in the work area and vectorization for each edge, the line segments are projected to object space to reconstruct 3D building models with SMS(Split-Merge-Shape). Finally, the height of building models are adjusted with lidar point clouds to determine the accurate height. The experimental cases include different kinds of roof planes. The proposed method is validated using manually measured 3D building models as reference data. The test data include (1) DMC aerial imagery with a spatial resolution of 17cm, and (2) lidar point clouds from Leica ALS 50 with a spatial resolution of 10 points/m~2. The experimental results indicate that the reconstruction may reach high accuracy and with high degree of automation.