Feature-point based matching: a sequential approach based on relaxation labeling and relative orientation

Abstract

This paper presents a solution for the problem of correspondence and relative orientation (RO) estimation for a pair of images. The solution is obtained by relaxation labeling using multiple metrics applied to image primitives. Besides the use of metrics based on radiometric elements (intensities, gradient, and coefficient of correlation), and on geometry (distance ratios), two additional metrics are considered. One is based on angular relations between primitives and another based on the volume of Matching Parallelepiped (MP) that allows the inclusion of the epipolar geometry constraint directly in the similarity and compatibility computation. The proposed solution was applied to synthetic and real images. The results showed that the use of multiple metrics contribute to the automation of the correspondence process and RO determination, even considering pairs of images subject to convergence, rotation, differences in scale, and presence of repetitive patterns.