3D Motion from structures of points, lines and planes

摘要

In this article we propose a method for estimating the camera motion from a video-sequence acquired in the presence of general 3D structures. Solutions to this problem are commonly based on the tracking of point-like features, as they usually back-project onto viewpoint-invariant 3D features. In order to improve the robustness, the accuracy and the generality of the approach, we are interested in tracking and using a wider class of structures. In addition to points, in fact, we also simultaneously consider lines and planes. In order to be able to work on all such structures with a compact and unified formalism, we use here the Conformal Model of Geometric Algebra, which proved very powerful and flexible.As an example of application of our approach, we propose a causal algorithm based on an Extended Kalman Filter, for the estimation of 3D structure and motion from 2D observations of points, lines and coplanar features, and we evaluate its performance on both synthetic and real sequences.