Toward video to geospatial reference image indexing

In this report, we are concerned with registration of data in geospatial databases, especially with registering images taken by different sensors and from different viewpoints of the same scene. Extant approaches break down as viewpoint and/or sensor vary beyond relatively small changes. Of particular interest in the current work is the development of techniques that allow an aerial video to index corresponding spatial location within a larger reference orthoimage, without detailed a priori knowledge of the relative acquisition scenarios (e.g., lacking telemetry). Such an approach can extend significantly the operational range of video to reference registration, as extant techniques make strong assumptions about the availability of good initialization. We present a uniform approach to representing video and reference imagery and for quantifying the goodness of match between two image samples, one captured from each type of source imagery, that have been brought under our representation. The approach combines image appearance, characterized in terms of texture defined regions, and image geometry, characterized in terms of relationships between textured regions. By construction, the matching methods are robust to a range of photometric and geometric distortions between image sources, including changes in greylevel contrast and affine geometric transformations. In application, the developed approach can serve to structure a reference image database that can be indexed directly via similarly represented video. Empirical investigations with real and synthetic data suggest the promise of the approach.