Photogrammetry Papers

 
Photogrammetry Research

A recent research thrust within the MAPSLab has been the utilization of photogrammetric knowledge within computer vision algorithms. The standard view within the computer vision community has been that photogrammetry is necessary only at the end of the processing chain, to translate image-based results into some object space coordinate system. Our approach is to build our algorithms upon a rigorous photogrammetric basis and to integrate photogrammetric knowledge throughout. This integration has several aspects:

• The use of object-space and image-space geometry to aid in geometric reasoning. This has proven especially valuable in building extraction; for example, we use line direction labelings (horizontal, vertical) derived from geometric principles to form and evaluate building hypotheses.

• The ability to make meaningful object space measurements, instead of working in arbitrary image (pixel) dimensions. This allows us to reason about significant properties of objects, such as building heights and lengths.

• Use of simultaneous image orientation solutions, rigorously tied to world coordinate systems. Most computer vision algorithms produce results in arbitrary coordinate systems. This is acceptable for algorithms working in isolation, but when trying to use a variety of cartographic data sources as input, and to produce results for inclusion in cartographic products, the ability to rigorously position algorithm results within a well-defined coordinate frame is essential. A cornerstone of this technique is the ability to do simultaneous multiple image solutions, in order to produce the most accurate and consistent set of image orientations.

• Generation and utilization of precision and reliability information. As important as the actual parameters associated with a sensor model or with generated object space coordinates is the precision information available from rigorous least-squares resection solutions. The precision information on camera parameters can be propagated into object space coordinates or into derived quantities, such as epipolar lines, establishing rigorous search areas and error bounds instead of arbitrary thresholds.