Problems in area-based image registration for stereo

While registration of stereo image pairs is an old problem, it is one which is of great importance, not only to cartographers, but also in machine vision applications. Many of the existing, successful cartographic systems are commercial, and the exact techniques used are not well publicised. We have been using area-based correlation methods in our work on wide-area imaging through the atmosphere, in which we have successfully registered, point-by-point, atmospherically-distorted images from a time sequence recorded by an astronomical telescope. The resulting registration provides both a measure of the turbulence (allowing its visualisation) and a reference for subsequent geometric restoration of the images. Following our success in astronomical imaging, we decided to apply our technique to the registration of image stereo pairs. This was done for several reasons-to see how our method compared with others in this traditional area, to determine how robust is is over different problem domains, and to further refine the method. Although the concept of area-based correlation is simple and well known, we have found that, in application, it requires a great deal of experimentation and "hand-crafting" to achieve useful and successful results. In the stereo application, we found that further fine-tuning has been required, but the end result is a very robust registration technique which can be applied to a wide variety of problems.