Implementation of the Barnard and Thompson algorithm on Transputers for SPOT stereo

The Transputer provides a powerful new capability in parallel computing which can readily be optimised for specific applications. The authors compare two alternative organisations of a large problem on a network of 21 Transputers, from the standpoints of speed, hardware efficiency and ease of programming. The authors programmed two highly parallel implementations of an algorithm which constitutes part of a `real-time´ system to generate terrain relief maps from satellite stereo image pairs. The particular algorithm used is the stereo matching strategy due to Barnard & Thompson (1980). The algorithm attempts to identify corresponding points in two image frames, in this case raw SPOT images, in order to refine initial estimates for the relative orientation parameters of the sensor in the two cases. The initial estimates are extracted from satellite header information, and the refined estimates are used to provide approximate epipolar image data for use with subsequent edge-based matching techniques. The authors describe one stage of the sequence implemented on the Transputers: apply the Barnard and Thompson algorithm to rough epipolar SPOT stereo pair, to obtain a spare disparity map