Camera Self-Calibration

Camera calibration is the problem of estimating a camera’s intrinsic parameters, which among others include the focal length and image center of the camera. The estimation of the camera’s intrinsic parameters is important as it is a prerequisite to a wide variety of vision tasks related to motion and stereo analysis. Historically, camera calibration was performed by knowledge of 3D world points and their corresponding image locations. For example, first a flat plate with a regular pattern, whose 3D coordinates are well known, is placed in the field of view of the camera. Then the image coordinates of a number of well recognizable and measurable features of the pattern are used to estimate the camera’s parameters [1]. The camera calibration problem can be solved in this way with the knowledge of a minimum of 6 point correspondences [2]. The above method estimates the camera parameters very accurately [1], but it has a couple drawbacks. The first drawback is that a calibration grid may not be available, and secondly it is not possible to calibrate the camera when the camera is on the field and it’s parameters may be changing [1]. The change in the camera parameters may be due to mechanical or thermal variations or may be due to focusing and zooming of the camera [1]. It has become apparent, however, that camera calibration can also be performed without the use of a calibration grid, but instead by only using a number of image correspondences [3]. Many methods have been proposed that perform this self-calibration. Most of the self-calibration methods relate the image correspondences to the absolute conic which stays fixed under all Euclidean transformations [4]. The methods, furthermore, seem to fall into one of three categories: