An Automated Approach to Stereo Matching Seasat Imagery

This paper describes a procedure developed at University College London for the automatic stereo matching of SAR imagery from NASA's Seasat satellite. The method employed uses Gruen's least squares correlation technique to improve the match accuracy of randomly generated points as they are cascaded down an image pyramid, coupled with a sheet growing mechanism in order to produce a dense array of points. The overall result of the cascading procedure is the totally automatic production of a dense digital disparity model [DDM]. 1 . INTRODUCTION In order to apply the area based Otto-Chau [5] stereo matcher to any form of imagery, you must first accurately determine image conjugate points. These points then become seedpoint data for the matcher and act as starting points for its sheet growing mechanism. Without accurately determined image conjugate points the stereo matching process is unable to proceed. The Otto-Chau stereo matcher was developed under the aegis of the Alvey MMI-137 [Real-time 2.5D Vision System] project undertaken at University College London. Otto-Chau is an area based stereo matcher which incorporates both Gruen's [3] adaptive least squares technique and a sheet growing mechanism. To date, any seedpoint data needed for the Otto-Chau algorithm has been measured manually via an interactive display program on a personal workstation. However, this procedure is both time consuming and prone to error [1]. In an attempt to overcome these problems and to achieve a totally automated system, cascade [6] was developed. Cascading uses the idea that if stereo images are reduced in size by a large enough factor, then any randomly generated points can be transformed into accurate conjugate points in the original images by applying Gruen's least squares correlation technique at each tier of the image pyramid (fig. 1). Gruen's [3] least squares correlation technique is used to act as a "pull in" to improve the match accuracy at the new increased resolution which is associated with each consecutive tier of the image pyramid. And, by also applying the sheet growing mechanism of the Otto-Chau [5] stereo matcher near the top of the image pyramid it is possible to produce a dense stereo matched extent at the bottom of the image pyramid. It must be noted that if applying the technique of cascading to SAR imagery, before any stereo matching can take place the problem of speckle [1, 2] must be addressed. Speckle is perceived by the Otto-Chau matcher to be a series of micro disparity domains, the boundaries of which act to inhibit the growth of the texture dependent sheet growing mechanism. This means that if a conventional approach to stereo matching is made, the matched extent will be limited to the immediate area surrounding the seedpoints [1]. However, it should also be noted that the technique of cascading points down an image pyramid means that the prohibitive effect of speckle in the SAR imagery will be not encountered until, at the earliest, the bottom two tiers of the image pyramid. To overcome the problem of speckle the images must be convolved with an appropriate speckle reduction filter [4]. 2 . METHODOLOGY The first stage in the overall process is to convolve the left and right images of the stereo pair with a speckle reduction filter. For this series of experiments an adaptive BMVC 1991 doi:10.5244/C.5.44