Robust algorithm for point matching in uncalibrated stereo vision systems

Abstract

This article introduces a new point matching algorithm for stereo images. The cameras used for capturing the image do not need to be calibrated. The only requirement is the existence of a set of segmented corners in each image. In order to execute the point matching, the algorithm starts by applying non-parametric techniques to the pair of images and a set of candidate matches is selected. After that, the reliability of each point is calculated based on a proposed equation. Finally, the fundamental matrix of the system is estimated and the epipolar restriction is used to eliminate outliers. Tests made on real images demonstrate the viability of the proposed method.