Many stereovision algorithms are based on the assumption that the images are registered in epipolar geometry. This requires that the two cameras of a vision system are aligned such that the camera axes are parallel and also perpendicular to the camera base. Initially, the cameras of a moving robot may be perfectly aligned, but this condition is likely to change in time. It is argued that the camera positions should be periodically surveyed and, if necessary, readjusted. The camera positions can be determined by an automatic orientation. In this paper, a method to compute the epipolar geometry from known camera positions is described. After defining the problem, the authors derive algorithms to resample the original images such that they satisfy the condition of epipolar geometry.
Toni F. Schenk,
Charles K. Toth,
"Recovering epipolar geometry in 3-D vision systems", Proc. SPIE 1457, Stereoscopic Displays and Applications II, (1 August 1991); doi: 10.1117/12.46295; https://doi.org/10.1117/12.46295