Matching Patterns of Line Segments Using Affine Invariant Features

This dissertation, "Matching Patterns of Line Segments Using Affine Invariant Features" by Chi-ho, Chan, 陳子濠, was obtained from The University of Hong Kong (Pokfulam, Hong Kong) and is being sold pursuant to Creative Commons: Attribution 3.0 Hong Kong License. The content of this dissertation has not been altered in any way. We have altered the formatting in order to facilitate the ease of printing and reading of the dissertation. All rights not granted by the above license are retained by the author. Abstract: Abstract of thesis entitled "Matching Patterns of Line Segments using Affine Invariant Features" Submitted by Chan Chi Ho for the degree of Master of Philosophy at The University of Hong Kong in December 2005 Image matching is a basic process of computer vision that consists of finding in two images the features which represent the original feature in the observed scene. It is the first step towards computational computer vision and is crucial to the performance and accuracy of subsequent processes such as depth estimation or 3-D model reconstruction. In this thesis, a new feature matching method is proposed. Patterns of line segments are extracted from two images. The method matches these patterns of line segments using affine invariant features. In the development of the method, we have considered the derivation of affine invariant features, representation of line segments, disambiguation and feature matching techniques. In the domain of affine transformation, it is shown that for each object formed by two line segments, affine moment invariant and relative area features together can characterize the relationship between two line segments adequately up to an affine transformation. The affine moment invariant features represent each object in a global way, while the relative area features represent it in a local way. Hence, the two invariant features complement each other to give a better description of the object. Since line segments cannot be extracted accurately, the invariant value of a corresponding line segment pair cannot be expected to be exactly the same. A noise model for the invariant is developed for matching line segments perturbed by noise. This study proposes a voting method to find line segment matches, whereby each match is supported by a number of votes. The matches with highest number of votes are regarded as matches. A number of verification tests (e.g. projection test) are then applied to the matches to determine if they are real matches. The matching process is iterative, and a disambiguation method using connectivity information is adopted to increase the number and accuracy of matches. A complete feature matching method, based on the studies described above, is then developed. Line segments are first extracted from the two images. Affine invariant features are created for each pair of line segments. Putative matches are determined by comparing these features. Voting and disambiguation methods are employed to obtain reliable matching results. After a set of line segment matches is found, a post process called RANSAC is run to find the fundamental matrix and remove mismatches. With this matrix, a fine- tuning process using graph matching is then run to adjust the matching results. Reliable matches of line segments and their end point correspondences were obtained in experiments on both synthetic and real images Finally, a novel method to estimate the fundamental matrix using line segment correspondences is discussed. It uses the intersection points of matched line segments to estimate the fundamental matrix. Further, a new Hilbert invariant feature which is invariant up to an affine transformation is derived for matching in future development. DOI: 10.5353/th_b3462725 Subjects: Transformations (Mathematics) Image processing Computer vision Geometry, Affine