A statistical approach to structure and motion from image features

Abstract

The estimation of structure and motion from image sequences using corresponding points, lines, conics and structured patches is treated. Recent research has provided good tools for obtaining good initial estimates of structure and motion using point, line, conic and curve correspondences. These estimates are, however, not so accurate. In this paper it is shown how to obtain statistically optimal estimates of structure and motion using a combination of such image feature correspondences. The question of using proper weighting is important when different types of features are combined. We show how weights can be chosen in a statistical optimal sense. Experiments with real data are used to evaluate every step of the algorithm.