Fast pose estimation for visual navigation using homographies

In this paper we propose a new algorithm for relative pose estimation between two images based on a new decomposition for an homography matrix faster than the classical solutions. We introduce in our method approximate information about the planes in the reference images but this additional information allows the decomposition to avoid the multiplicity of solutions. An exhaustive analysis of the error propagation through the method is provided. The new decomposition can be used for visual navigation of robots moving on a planar surface. The approach is based in the well known teaching-by-doing scheme with the route defined by a set of images. The parameters of the planes can be computed automatically from the reference images before the navigation. The contributions are the easy method to extract the parameters of all the planes and the use of this information in navigation tasks using a fast homography decomposition. The experimental results show the performance of the proposal.