Image Point Feature Matching by Triangulation

Abstract

Matching point features is recognized as one of the most important processes in many computer vision areas, such as motion analysis, stereo image processing, and scene analysis. In this paper, we present a new method to find the correspondence of point features between two images. Our method first constructs a network of triangles using extracted point features through a triangulation algorithm. The corresponding triangles between two images are then found by comparing the intensity distribution, and the Mahalanobis distance of a feature vector which is composed of the area, the center of gravity, the sum of side lengths and orientation of each side, and the two moment invariants of the triangles. As each element of the feature vector is a function of the three vertices of any given triangle, the uncertainty modeling of this vector is based on the error propagation from these vertices. The correspondence of triangles between two images naturally provides the corresponding point features. The image deformation described by the first order differential invariants is also obtained. An efficient method to compute shape moment of a triangle is introduced in accordance with the design of a triangular network from the extracted point features. This method effectively computes any higher order shape moments in real time. Experimental results from real images demonstrate the proposed algorithm could be used as one of the key attributes for matching point features. © 1997 Taylor & Francis Group, LLC.