3D city site model extraction through point cloud generated from stereo images

Abstract

It is a grand challenge to automatically extract 3D city site models from imagery. In the past three decades, researchers have used radiometric and spectral properties of 3D buildings and houses to extract them in digital imagery with limited success. This is because their radiometric and spectral properties vary considerably from image to image, from sensor to sensor, and from time to time. The locations and shapes of 3D buildings and houses are invariant and painfully obvious in a terrain-shaded relief image generated from a point cloud. Based on this observation, we have developed AFE (Automatic Feature Extraction) that can automatically extract 3D city site models from a point cloud which is automatically generated from stereo images. Point cloud generation from stereo imagery is a key technology which has been used in the geospatial industry for more than two decades. We have developed NGATE (Next Generation Automatic Terrain Extraction) that matches every pixel across all selected stereo image pairs. For each XY location, an array of Z coordinates are computed from a number of different stereo image pairs using a voxel 3D grid. The voxel 3D grid is statistically filtered for outliers and weighted averaging is used to generate a very dense and accurate point cloud. The AFE algorithms consists of the following components: identify and group 3D building and house points into regions; separate buildings and houses from trees; trace region boundaries; regularize and simplify boundary polygons; construct complex roofs. As shown in the following figures, 1505 buildings and houses have been extracted by AFE, from a point cloud generated by NGATE, using 12 stereo images (GSD 15cm) over downtown Oakland, California, USA. The background image is the terrain shaded-relief image generated from a point cloud. NGATE used all the 66 stereo image pairs and generated a point cloud of 50 million 3D points with a spacing of 30cm.