An algorithm for computation of the scene geometry by the log-polar area matching around salient points

Abstract

Computation of a scene geometry belongs to the one of the fundamental problems of computer vision. It can be computed from point correspondences found in a pair of stereo images or a video sequence. This is achieved by an image matching algorithm. However, found correspondences usually are burdened with large errors due to noise and outliers. In this paper we propose an improvement to the point matching algorithm which is twofold. At first the salient points are found which are corners detected by the structural tensor. Then the log-polar representations are computed around found salient points and the matching is done in the extended log-polar space. Such representation has very desirable properties which allow detection of a local change of scale and rotation of the matched areas. This feature is employed in the matching algorithm to eliminate outliers. The proposed method can be used in variety of computer vision tasks, such as stereovision, recovering shape and motion from video, or camera calibration and autocalibration. © Springer-Verlag Berlin Heidelberg 2008.