Automated stereophotogrammetry

In this paper we outline a structural pattern recognition approach to the stereo matching problem of automated stereophotogrammetry. Oriented-edge graphs are obtained with the edge vectors and filtered to obtained feature points for matching purposes. A resolution pyramid is based to aid the accurate matching of the feature points. At each stage of the pyramid, pseudo-hexagonal gray scale arrays are used to bypass the four-eight connectivity paradox in the implementation of association of oblique vectors in Bowker's association filter. The Fisher-$\text{z}$ transform is used to determine the correlation function threshold in matching. To fill the gaps between the matched nodes, a local interpolation method that linearly weights the disparity values in a window was developed. The interpolated points are not used for matching, but give the initial approximation at the next lower stage of the resolution pyramid. Finally, a modified normalized gray-scale correlation is used to refine the parallax found at the lowest level of the pseudo-hex resolution pyramid using the original 2D raster. The correlation scheme used here works in both directions: from left to right and from right to left, to obtain reliable matching. We have tried to use as simple operations as possible in each stage to allow the algorithm to run in real-time, possibly leading to an economical hardware implementation.