An algorithm for image matching is proposed, which uses both epipolar and homography constraints. At the first step, the Forstner algorithm is employed for features extraction. The features description and matching method of SIFT are used to find a group of original correspondences, then the correspondences are refined by LSM(Least Square Matching). With the refined correspondences the RANSAC algorithm estimates the fundamental matrix robustly and the more accurate correspondences with less outliers are gotten, which are called as correspondences candidates. As the features extracted from the image are all the edge inflexions, texture nodes with maximal intensity or corners of the objects in the 3D world. The features which are adjacent can form a local plane or quasi-plane. So the homography constraint is proposed for image matching. At the second step, the corresponding features seeds around the feature to be matched are recognized from the correspondences, which associate with a real 3-D scene plane or quasi-plane. Then with the seed correspondences the local homography matrix is computed. At last, under the guide of the local homography matrix, the coarse position of the target feature on the opponent image is found, then with the constraints of the epipolar line and the coarse position, the normal correlation and LSM matching methods are employed to match the features accurately. The algorithm searches for the corresponding feature only in a very small region and works quickly. Experimental results show that the algorithm is efficient and it improves the robustness and accuracy of the automatic image matching.

Date Published: 22 January 2008
PDF: 7 pages
Proc. SPIE 6833, Electronic Imaging and Multimedia Technology V, 68330Z (22 January 2008); doi: 10.1117/12.757770

Published in SPIE Proceedings Vol. 6833:
Electronic Imaging and Multimedia Technology V
Liwei Zhou; Chung-Sheng Li; Minerva M. Yeung, Editor(s)