STITCHING LARGE MAPS FROM VIDEOS TAKEN BY A CAMERA MOVING CLOSE OVER A PLANE USING HOMOGRAPHY DECOMPOSITION

Abstract. For applications such as underwater monitoring a platform with a camera will be moving close to a large roughly planar scene. The idea to map the scene by stitching a panorama using planar homographies is nearby. However, serious problems occur with drift caused by uncertainty in the estimation of the matrices and un-modelled lens distortions. Sooner or later image points will be mapped to infinity. Instead this contribution recommends using the homographies only for the composition of local patches. Then the homography obtained between the first and the last frame in such patch can be decomposed giving an estimate of the surface normal. Thus the patches can be rectified and finally stitched into a global panorama using only shift in x and y. The paper reports about experiments carried out preliminarily with a video taken on dry ground but a first under water video has also been processed.