Modeling and Information Processing
ANAÏS 3D - Analysis in 3D imaging
Extracting 3D models at iso-altitude in the Pleiades context
This work is the continuation of the studies dedicated to the extraction of buildings of any shape from very high resolution (between 20 and 50 cm ) airborne or satellite images. The objective is to adapt the methodology developed previously to the Pleiades high resolution optical context. This adaptation mainly revolves around the resolution of two different problems of varying difficulty: taking a triplet of images and a vegetation mask into account. In order to prepare for this development, the images we are working on come from either the IGN's digital camera (Pelican) for the triplets or Quickbird for the vegetation mask.
Integrating a triplet of images
The great agility of the Pleiades satellites allows triplets of images to be acquired by aiming the instrument. The computing code developed up to now for our 3D applications was handed down from stereoscopy. In order to take a third image into account to improve the robustness of the algorithm, major developments are needed in the method of calculating correlations and in the use of the scene's geometry.
We start from the geometry of the central image with a variable altitude. The correlations are calculated in sub-pixels, with the search area varying according to altitude. The search criterion is the minimum of the three correlation coefficients (left-center, left-right, and center-right). The extraction of the buildings is based, as before, on the deformable models (polygonal model) and is done interactively. It is initialized by segmenting the image into 3 zones: building, masking, ground (Figure 1) calculated by increasing areas of disparities in the image. The foreground minimization function takes into account the correlation inside and outside building, the cost of masking, the gradient on the contours and a geometric adjustment term.
The first results on a triplet of images of Amiens show the precision of the resetting obtained (Figure 1 on the right) after the convergence of the deformable model.
Figure 1 - Pelican triplet:
left, central with initialization of the snake; right, final cropping
l [vidéo]
Conclusions and prospects
The next stage in this work consists of adapting the algorithm for the extraction of buildings of variable heights. In practice, up to now, the extraction of buildings has been constrained by a fixed height combined with a low tolerance level. Finally, work is proceeding in order to find complementary data in order to improve the energy of the deformable models.
Publications
[1] H. Oriot, Statistical snakes for building extraction from stereoscopic aerial images, in ISPRS Conference on Photographic Image Analyses (Institute of Photogrammetry and Geoinformation, U. Hannover, Hannover, Germany. 2003), pp. 218-226.
[2] H. Oriot, A. Michel, Building Extraction from Stereoscopic Aerial Images, Applied Optics, Volume 43, Issue 2, pp 218-226, janvier 2004.
