Self-Calibration of Stereo Cameras Using Correspondences of Edges

Abstract

This paper describes a method for self-calibration of camera parameters of stereo vision systems.In stereo vision, for a point in one image its corresponding point lies on the so called epipolar line in the other image. The epipolar line is calculated from camera parameters. If the parameters include some errors, the errors cause small displacements between corresponding pairs and, thus, epipolar cons traint can not be used effectively. The errors of camera parameters can be calculated if a enough number of displacements of corresponding pairs are obtained. To detect enough displacements, an image is divided into small regions and at every region, a two-dimensional histogram of the disparity and the vertical displacement of features is generated, and the displacement is detected as the position of the prominent peak, if found. The detection of displacements is repeated until enough displacements are obtained everywhere in an image, where more reliable features are given a higher priority to detect reliable displacements earlier than less reliable ones.
Since edges are used as features, structural descriptions are not necessary. Thus, this method is applicable to scenes with natural objects such as trees, bushes and so on. The calibration of displa cements provides establishing correct correspondences.