High-precision relative orientation using feature-based matching techniques

Abstract

This paper presents a method for relative orientation using feature-based matching techniques. Both the interest operator and feature-based matching are used to select distinctive points and measure their parallaxes for relative orientation automatically. After the feature-based matching a fine least-squares matching is performed in order to improve accuracy. The matched points are then used for the relative orientation. The precision of relative orientation depends on the final least-squares matching. Some aspects of the least-squares matching have also been studied. An off-line experiment was carried out. Results of 2.08 μm for the standard error of unit weight (σ 0) of the γ-parallax for a simulated stereo model and 3.83 μm for a real stereo model are obtained. The estimated precision of feature-based matching together with least-squares matching for stereo parallax measurement is 0.06 pixels. The experiment also shows the high potential for automating the relative orientation procedure.