Generic Estimation Procedures for Orientation with Minimum and Redundant Information

1 Motivation Orientation of cameras with minimum and redundant information is the rst step in 3D-scene analysis. Compared to image interpretation it looks simple, it seems to be solved in photogrammetry and is expected to be implemented within a few weeks. All experience shows that camera calibration and orientation needs much eeort and the solutions provided in photogrammetric textbooks cannot be directly transferred to automatic systems for scene analysis. The reasons for this situation lie in the hidden complexity of the calibration and orientation tasks. { Camera modelling requires a thorough understanding of the physics of the image formation process and of the statistical tools for developing and re-ning mathematical models used in image analysis. High precision cameras used in aerial photogrammetry have put the burden of obtaining high precision on the manufacturer, leading to the-only partly correct-impression that calibration can be standardized, and thus is simple. The long scientiic struggle photogrammetry went through in the 70's, which is not mentioned in today's publications, must now be repeated under much more diicult boundary conditions: non-standardized video cameras, non-standardized applications , the requirement for full automation, therefore the integration of error-prone matching procedures, etc.. { The 3D-geometry of orientation reveals high algebraic complexity. This is overseen when assuming the calibration and orientation to be known or at least approximately known, as the iterative techniques used in photogramme-try and the spatial intersection (triangulation)-in general-lead to satisfying results. Again, the eeorts of photogrammetric research in the 70's and early 80's for generating guidelines for a good design of so-called "photogrammet-ric blocks", where hundreds and thousands of images are analysed simultaneously for object reconstruction, speciically mapping, has to be invested for the diierent tasks of 3D-scene reconstruction in computer vision, especially in the area of structure from motion. It is interesting and no accident