Automatic 3d Rendering of High Resolution Space Images

Virtual flights over a Digital Surface Model (DSM) textured with an image are now widely used for communication purposes towards the general public, decision-makers and non-specialists in general. Furthermore, recent advances in computer science and virtual reality make very easy to perform such flights, even on a standard laptop. Examples will be given with SPOT5 2.5m colour images applied to HRS-derived DSM on a 60 km by 60 km surface. Performing the same kind of virtual flight over metric images needs to solve a certain number of problems related to the fact that we have to move from a 2.5D description of the landscape to a real 3D description, essential for the recognition of all the objects that can be seen in such images. Therefore, the surface can no longer be described as a regular grid of altitudes, especially to be able to account for vertical walls. It is necessary to use a vectorized description of the landscape geometry, for example an irregular network of triangles (TIN) in a Virtual Reality Mark-up Language (VRML) format. The problem of estimating a 3D model from images will not be addressed here. For our purposes, we will consider that this 3D geometric description is available from other sources (city models...) and that we “only” want to be able to virtually generate the landscape by applying a metric image over a given 3D model. In this case, the first step is to check its quality and eventually simplify the 3D model -its quality and level of detail can vary considerably from one source to another, even for a same townand then to merge it with a traditional Digital Terrain Model, giving the altitude of the ground and of the object’s bases. The second step consists in registering the image with the geometric model with a sub-pixel accuracy. The third step deals with the texturing of the geometric model with the image, which means associating each triangle with the image chip representing its radiometry. This extraction process must take into account the viewing angle of the surface in the image and its orthogonal projection. At this level, we have produced a virtual mock-up modelling a real 3D landscape that we can use to generate several outputs: an image in any projection (an ortho-image, for instance), any perspective view of the landscape or a virtual flight. We will illustrate the different stages of the processing with Pléiades-HR simulated images texturing the 3D model of Toulouse.