Feature Point Matching of Stereo Images by using a Projective Invariant

In Image-Based Modeling, which reconstructs 3-dimensional structures and extracts a texture from 2-dimensional images, one of the greatest problems is to match feature points of two images. The well-known template matching is used for automatically relating individual points in different images. However, since it matches each image in which a viewpoint position differs, incorrect feature points are often related. In this study, we propose the automatic matching technique of feature points that is not sensitive to parallax. We aim to improve the accuracy of feature point matching by using a projective invariant on a plane surface. In order to calculate the projective invariant, we use collinear four points or coplanar five points. Firstly, in each image corners of objects are picked out as feature points. Next, an edge image is created from the original image. The connective relation of edge points between feature points is measured. All of the collinear four feature points and the five feature points are estimated from the connective relation, and a projective invariant is calculated for each set of points. Finally, all projective invariants of one image are compared to those of another image and feature points with an equal projective invariant are associated. Then we confirm the effectiveness of the proposed technique by the experiments.