Realistic Landscape Modelling with High Level of Detail

An approach for generating realistic 3D models of natural landscapes from overlapping aerial images is presented. Based on stereoscopic correspondence analysis a 3D terrain model is computed. The 3D reconstruction of a scene from 2D projections is underconstrained and details are lost due to occlusion and image resolution. The ambiguity and lack of detail is overcome by integrating a priori knowledge about the scene in the modelling process using image interpretation. Segmentation of linear features employs edge detection, and segmentation of areal features uses region growing based on a statistical texture model. Knowledge about object geometry is used to model continuous surfaces separately and to force height steps at discontinuities as forest edges. For visualization the landscape model is approximated by a polygon mesh with overlaid photo texture.