Chaining Planar Homographies for Fast and Reliable 3D Plane Tracking

This paper addresses the problem of tracking a 3D plane over a sequence of images acquired by a free moving camera, a task that is of central importance to a wide variety of vision tasks. A feature-based method is proposed which given a triplet of consecutive images and a plane homography between the first two of them, estimates the homography induced by the same plane between the second and third images, without requiring the plane to be segmented from the rest of the scene. Thus, the proposed method operates by "chaining" (i.e. propagating) across frames the image-to-image homographies due to some 3D plane. The chaining operation represents projective space using a "plane + parallax" decomposition, which permits the combination of constraints arising from all available point matches, regardless of whether they actually lie on the tracked 3D plane or not. Experimental results are also provided