Registration and Stereo Papers

 
Image Registration and Stereo Research

The perception of depth from a stereo pair of images provides important qualitative and quantitative cues for scene interpretation which are not available from a monocular view. Many tasks in cartography can only be reliably accomplished using a stereo viewing model. For example, the computation of the underlying digital elevation model is a typical first step in the generation of an orthophoto from which planimetric features can be compiled. Our research in image registration and stereo is directed toward the recovery of cartographic features within the images.

Image registration is a fundamental requirement of a number of image analysis tasks such as stereo matching, multi-image matching for temporal changes, and image sequence or motion analysis. As a result, there exists a rich variety of techniques to perform image registration. Our research has been directed toward the automation of the relative orientation between two aerial images and the quantitative analysis of the accuracy of the resulting registration. While absolute error measures are important for algorithmic comparisons, the effect of errors on downstream feature extraction and stereo matching must be evaluated within the context of the overall end-to-end system.

Our research in stereo image analysis has ranged from traditional stereo pair analysis provided by near nadir (parallel axis) mapping photography, the generalization of these techniques to use oblique image pairs, and the exploration of simultaneous matching of multiple (three or more) images acquired under a variety of imaging geometries.

Using oblique imagery introduces several problems to the process of stereo matching that are not found with near-nadir imagery:

• As the image obliquity increases, approximate image warping techniques traditionally used to align the stereo pair prior to image matching break down severely. A rigorous photogrammetric approach to resect the images and generate the epipolar reprojection into a stereo model is required.

• Under obliquity the ground-plane is not perpendicular to the camera axis. Many stereo matching algorithms implicitly rely on this assumption, as well as the assumption that building roofs are perpendicular to the camera axis; unfortunately this does not generally hold even with vertical (near-nadir) imagery. Further, building walls are commonly observed rather than being exceptional features. These problems require significant generalizations to the image registration process and assumptions made by the stereo matching algorithm.