Precise Wide Baseline Stereo Image Matching for Compact Digital Cameras

Abstract

Numerous image matching methods for wide range of applications have been invented in the last decade. When high precision and reliability of the object space point coordinates is highly demanding, a stereo image matching method which can produce conjugate point of images and a standard deviation of the matched point is examined. In this approach, image gradients are used locally to seek a conjugate patch. The normalized cross correlation is first utilized to estimate an approximate location of the conjugate patch between two normalized images. Then the location of conjugate patch is further refined by using Gaussian-Newton least squares image matching. Both radiometric and geometric parameters of least squares models are used selectively in seeking the best possible accuracy. Iterative computation is conducted to incrementally refine the geometric location of the conjugate point. After a matched patch has been found, a variant-covariant matrix of the parameter is analyzed to inform the precision of the conjugate points both on images and object space. This method can compute high precision object space points and some examples demonstrate the insight of the approach.

A Relative Rotation between Two Overlapping UAV's Images

“…An array of 30 ground control points (GCPs) is established from a white concentric ring surrounded with dark background for point correspondences (Fig.2). To avoid false matches and to facilitate a possible highest accuracy of image coordinate measurements of GCPs on stereo images, least squares image matching are performed [3] to seek the best matches on stereo images as illustrated in Fig. 3.…”

“…In supporting a large scale urban city mapping [1] from UAV's images, mosaicking and compositing aerial images [2], as well as stereo matching [3], it is necessary to determine a three-dimensional motion of a rigid object (i.e. a flying UAV platform) from perspective images.…”