Level-line primitives for image registration with figures of merit

摘要: 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 shape to be maintained constrains kind approximation. Primitives stem level-lines, inheriting their robustness towards contrast changes. still improves it through efficient cumulative based on multi-stage primitive election procedure. This paper is continuation preliminary work simplest transform, similarity, constructing two-segment-primitives. contribution validate further transforms, completing path "similarity, affine, projective". Transformations are stable when they computed planar objects or scenes which contain many coplanar facets elements. approach works with cluttered images, even if estimation done globally while apparent displacement not small there several different unknown motions Results obtained selected shapes primitives shown compared corresponding sections. A gauge transform goodness elaborated, an assumption spatial ergodicity incentive estimate conditional probability similar pixel values neighborhood original transformed respectively.