An Efficient and Flexible Solution for Camera Autocalibration from N≥3 Views

Abstract

This paper presents an efficient and flexible solution for camera autocalibration from N≥3 views, given image correspondences and zero (or known) skew only. The knowledge is not required on camera motion, 3D information, scene, or internal constraints. Our method is essentially based only on the fundamental matrices and its main virtues are threefold. Firstly, it is shown that, in the center-oriented metric coordinates, the focal length and aspect ratio can be estimated independent of considerable principle point shift (PPs). Thus, our method includes recursive steps: estimating focal length and aspect ratio and then calculating the PPs. Secondly, the optimal geometric constraints are selected for calibration by using error propagation analyses. Thirdly, the Levenberg–Marquardt algorithm is adopted for the fast final refinement of four internal parameters. Our method is fast and efficient to derive a unique calibration. Besides, this method can be applied to calibrate the focal length from two views, without requiring the prior knowledge of PPs. Good performance of our method is evaluated and confirmed in both the simulation experiments and the practical tests.