Significant Feature Detection and Matching in Image Pairs

Abstract

We present a technique for identifying and matching significant features in image pairs . Significant features are found by first detecting contours on each image, and then choosing distinctive or significant edge elements along these contours. Each contour is treated independently. The feature detection algorithm is iterative and is designed to work on complex scenes . Significant features are first matched between images by using only local criteria such as curvature and intensities about the features, and we do not initially assume that we know the approximate position of matches in the images. After an initial set of possible matches is found, crosscorrelation is then used only to confirm or reject these matches. Feature detection and matching are extended to scale space . Features are also used to segment contours, and can be used to match contour segments between images. The algorithm is appropriate for binocular stereo, motion stereo and object motion.