A new stratified self-calibration algorithm under small camera rotations

In this paper, we consider the problem of self-calibration from image sequence taken by a camera with constant intrinsic parameters. Stratified approach starts from projective calibration, refines it into affine calibration and finally upgrades it to metric calibration. Both linear and nonlinear algorithms were proposed for projective and metric steps. The affine step was reported to be the most difficult one in the whole process. The current technique mostly depends on nonlinear optimization followed by random search or dense search. This paper solves this problem by assuming that the camera rotation between subsequential images is small, which can be easily satisfied in real applications. We found that the relationships among three main diagonal elements of the infinity homography of such two views can be approximated as linear in the affine stage. The idea was developed into a simple, linear algorithm to compute the affine calibration from projective calibration. The result can be used as the start point of nonlinear optimization, such as modulus constraint. Simulations and experiments are presented in the paper.