Reconstructing 3D Buildings from LIDAR Using Level Set Methods

We present a novel approach to reconstructing cities and buildings from LIDAR data using level set methods. Traditional approaches to building extraction from LIDAR data use image segmentation algorithms to determine the outlines of rooftops, estimation of height/depth maps, polygonal mesh generation and extrusion to generate 3D models resulting in buildings with high quality rooftops but flat sides with little or no detail shown on vertical surfaces (e.g. overhangs and windows on walls). Texturing these flat side polygons with aerial and geo-registered ground imagery create acceptable photo-realistic models although the resulting buildings are generally not geometrically accurate causing stretching and waviness in texture-mapping. Our approach uses the LIDAR data directly as constraints in a variational framework and can estimate the geometry more accurately and demonstrate its effectiveness with simulated data.