A novel image matching method in camera-calibrated system

A novel approach for matching points from two views in a camera-calibrated system is presented in this paper. The interest points are selected using epipolar gradient features. Then three local region invariant descriptors, together with similarity measures are proposed. These descriptors are constructed from binary-threshold gray histogram, sample statistics of the edge pointspsila epipolar gradients, and average intensities of points on the epipolar line, respectively, all image patches based. Similarity measures relative to the descriptors work in cascade. Experimental results demonstrate that our matching scheme is tolerant to image deformations due to changes of viewpoint and effects of perspective, and can find more corresponding points.