The use of correlation as a measure of similarity between two signals is a well known technique. Correlation is commonly used in stereo vision to solve the correspondence problem. In this context the aim is to find the position of a point in one of the two images of a weakly calibrated stereo pair which corresponds to a point in the other image. Template windows are chosen from the first image as 2D information samples to be matched in a region of interest of the second image. We propose a technique which exploits an adaptive template matching. It works as an iterative optimization of the involved parameters with a consequent iterative refinement of the correspondence. Matching is performed by transforming a template window using the current image-to-image affine transformation and exploiting the epipolar geometry to reduce the degrees of freedom of the template position in the region of interest. As a validation test of the presented technique we report a study on the accuracy of the obtained results with various template sizes. We show that subpixel accuracy can be reached with relatively small templates. A major concern of our work is the accuracy of the correspondence. Higher accuracy of the correspondences results in fact in a more realistic reconstruction of the viewed scene. In the case of a stereo pair having undergone an estimated distortion of 0.5 pixel (10 micrometers of pixel size) it shows an accuracy in the correspondence of 3 micrometers for template windows of 17 X 17 size selected in textured image patches. Experiments are under process to improve the mentioned results.

Date Published: 4 January 1995
PDF: 12 pages
Proc. SPIE 2356, Vision Geometry III, (4 January 1995); doi: 10.1117/12.198613

Published in SPIE Proceedings Vol. 2356:
Vision Geometry III
Robert A. Melter; Angela Y. Wu, Editor(s)