ABSTRACT
Correlation is widely used as an effective similarity measurein matching tasks. However, traditional correlation basedmatching methods are limited to the short baseline case. Inthis paper we propose a new correlation based method formatching two images with large camera motion. Ourmethod is based on the rotation and scale invariantnormalized cross-correlation. Both the size and theorientation of the correlation windows are determinedaccording to the characteristic scale and the dominantdirection of the interest points. Experimental results on realimages demonstrate that the new method is effective formatching image pairs with significant rotation and scalechanges as well as other common imaging conditions.
1. INTRODUCTION
Image matching plays an important role in many applications.A lot of matching algorithms have been proposed inthe literature [1,2]. Matching two uncalibrated images withlarge camera motion such as significant rotation and scalechanges still remains a difficult problem. One effectivestrategy is using feature matching approach, which extractssalient features such as corners in the two images and thenestablishes reliable feature correspondences [3,4].Normalized cross-correlation has found application in abroad range of computer vision tasks such as stereo vision,motion tracking, image mosaicing, etc. Normalized crosscorrelationis the simplest but effective method as asimilarity measure, which is invariant to linear brightnessand contrast variations. Its easy hardware implementationmakes it useful for real-time applications.There have been some image matching methods basedon normalized cross-correlation [5,6,7]. However, thesemethods cannot perform well when there are significantrotation and scale changes between the two images. This isdue to the limitation that normalized cross-correlation issensitive to rotation and scale changes. Therefore,traditional correlation based matching methods are notrobust against rotation and scale changes. There are alsogeneralized versions of cross-correlation, which calculatethe cross-correlation for each assumed geometrictransformation of the correlation windows [8,9]. Althoughthey are able to handle more complicated imagingconditions, the computational load grows very fast in themean time.This paper presents a new image matching methodbased on normalized cross-correlation, which can efficientlyhandle image pairs with significant rotation and scalechanges. First, interest points are detected in the two imagesseparately. Each interest point is assigned one characteristicscale and one dominant direction. Then the new methoduses rotation and scale invariant normalized crosscorrelationas the similarity measures between two interestpoints to establish the interest point matches. In order to beinvariant to rotation and scale changes, both the size and theorientation of the correlation windows are determinedaccording to the characteristic scale and dominant directionof the interest points. Finally, the epipolar geometryconstraint is imposed to reject the false matches. Experimentalresults demonstrate that the new method performswell on real images with different imaging conditions suchas large angle rotation and significant scale changes.The remainder of the paper is organized as follows.Section 2 describes extracting interest points withcharacteristic scale and dominant direction. Section 3introduces the matching algorithm based on rotation andscale invariant normalized cross-correlation and presents indetail the calculation of similarity measures between twointerest points. Section 4 describes rejecting the falsematches by imposing epipolar geometry constraint. Section5 presents some experimental results on real images andSection 6 concludes the paper.


Download full report
http://jdl.ac.cn/doc/2006/Image%20Matchi...lation.pdf