Automatic generation of building primitives using multi-source data

Although many researchers have studied building reconstruction from remotely sensory data, most approaches are not yet satisfactory in terms of the degree of automation, the reconstructed details and the accuracy. With the innovation of sensory technology, more advanced sensors are now available and getting cheaper. This paper describes a framework for automatic generation of 3D building models from the data acquired from multi-sources, specifically, airborne LIDAR, digital camera, and a digital map. By combining these data, we derive building primitives in 3D space. The core primitives are step and intersection edges from images, step edges from building boundary of digital maps, patches and intersection edges from LIDAR data, and step edges from DSM generated from the LIDAR data. Then, we group these elements and refine the grouping results to generate polyhedral models of buildings. This framework was partially implemented and applied to real data. The experiment results show that the framework can open a possibility of automatic building reconstruction.