ESTIMATION OF CAMERA PARAMETERS FROM STEREO PAIRS WITH NO EXTERNAL CONTROL INFORMATION
I. Kalisperakis, George C. Karras, Eleni Petsa · 2005
Leaving planar rectification aside, in most practical cases of architectural photogrammetry – where mostly non-metric cameras are now used – the camera interior orientation must be known. Of course, it is infeasible to anticipate every possible case and calibrate in advance all available (including auto-focus and zoom) lenses. Therefore, tools are required for the on-the-job estimation of interior orientation parameters. This paper focuses on the simplest instances, namely when no ground control is available. To this end, there exist both single-image methods (for images with clearly defined vanishing points) as well as a multi-image approaches (self-calibrating bundle adjustment which, in principle, also allows estimating the camera interior parameters without exterior information). Most applications, however, still rely on the stereo pair. Thus, the question of determining camera parameters solely from point correspondences on two images is addressed here. Indeed, it is known that simple point homologies in uncalibrated stereo pairs can impose the epipolar constraint, expressed by the ‘fundamental matrix ’ (well documented in computer vision literature), and even permit partial self-calibration. For instance, the camera constant of normal cameras – even if different for the two images – may be recovered if the principal point is assumed to be known. Several non-iterative algorithms have also been presented for computing one or two camera constants (the principal point can, too, be localized for given camera constants). The pertinent critical configurations have also been studied. In this contribution, the specific question of estimating a common camera constant is treated. After using a linear algorithm for obtaining initial values, bundle adjustment of the stereo pair is performed for the estimation of relative orientation together with