We describe a parallel-distributed approach to simultaneously solving the correspondence problem and the pose estimation problem. When the correspondence is known our method reduces to a generalized Hopfield network with a node for each image point that has an activation that converges to the ray-length through the image point to the object point. Inexact knowledge of the correspondence between object points and image points is represented as an assignment array that encodes the degree of confidence in each of the possible correspondences, similar to the traditional `assignment problem' of optimization theory. Two neural networks, one for pose estimation and one for solving the assignment problem, interact via the introduction of `pseudo image points,' created from the inexact correspondences. These `pseudo image points' reduce to the actual image points when the correspondence is known.

Date Published: 14 February 1992
PDF: 7 pages
Proc. SPIE 1613, Mobile Robots VI, (14 February 1992); doi: 10.1117/12.135176

Published in SPIE Proceedings Vol. 1613:
Mobile Robots VI
William J. Wolfe; Wendell H. Chun, Editor(s)