Area-based automatic image matching combined with a region-growing technique are able to provide a dense and accurate set of corresponding points. However, the region-growing process may stop at image patches where the horizontal x-parallax has an abrupt change. In such cases new pairs of corresponding points (seeds) must be provided, usually by a human operator. The region growing procedure restarts then from the new seed points. Depending upon the type of image and the 3D-structure of the mapped area, the human intervention may be considerable. A fully automatic alternative for finding conjugate points in stereo pairs was proposed by the authors in a prior work. The method combines the scale invariant feature transform, the Least-Squares matching and the region-growing technique. This work presents an extension of that technique. Basically, improvements in the matching step of the SIFT algorithm are proposed, which explores properties of stereo images produced by aerial and orbital sensors. Experiments conducted on stereo pairs from both airborne and satellite imagery show that the benefit of the proposed changes is twofold. Firstly, the number of true matches increases substantially with no significant increase in the proportion of false matches. Secondly, the computational load is dramatically reduced.
Photogrammetry; Least Square Correlation; Least Square Matching; Region Growth; SIFT; LSM; DEM; DSM
