Abstract
Planar geometric transformations approximately model differences between images from a moving camera. Our registration technique consists of finding and matching featured primitives, invariant in the observed scene. The invariant shape to be maintained constrains the kind of approximation. Primitives stem from level-lines, inheriting their robustness towards contrast changes. The registration still improves it through efficient cumulative matching based on a multi-stage primitive election procedure. This paper is a continuation of a preliminary work on the simplest geometric transform, similarity, constructing two-segment-primitives. Our contribution is to validate further geometric transforms, completing the path “similarity, affine, projective”. Transformations are stable when they are computed on planar objects or from scenes which contain many coplanar facets and elements. Our approach works with cluttered images, and even if the estimation is done globally while the apparent displacement is not small and there are several different unknown motions in the scene. Results obtained with selected shapes of primitives are shown and compared in the corresponding sections. A gauge of transform goodness is elaborated, based on an assumption of spatial ergodicity incentive to estimate a conditional probability of finding similar pixel values in a neighborhood of the original and transformed images respectively.