Underwater mosaicing and trajectory reconstruction using global alignment

This paper deals with the problem of constructing high quality mosaics of the sea bed. It focuses on the use of long image sequences with time-distant superpositions, such as the ones arising from loop trajectories or zig-zag scanning patterns. An algorithm is presented for the simultaneous creation of mosaics and the estimation of the camera trajectory. The method comprises three major stages. The first stage consists of the sequential estimation of the image motion, using a reduced image motion model. The set of resulting consecutive homographies is cascaded, in order to infer the approximate topology of the camera movement. The topology information is then used to predict the area where there is image overlap resulting from nonconsecutive images. Secondly, a motion refinement is performed, by iteratively executing the following two main steps. (1) Point correspondences are established between non-adjacent pairs of images that present enough overlap. (2) The topology is refined, by searching for the set of homographies that minimizes the overall sum of distances in the point matches. The final stage of the algorithm consists of estimating the set of homographies and a world plane description that best fit the observation data. As the main concern is attaining high registration accuracy, a general parameterization of the homographies with 6 DOF for the pose is used, which is capable of modelling the effects of wave-induced general rotation and translation. The overall method is fully automatic in the sense it does not require human intervention at any of the stages, apart from the beforehand specification of the most adequate motion model for the first stage. We present results obtained from shallow water image sequences acquired by a ROV. The images present some of the common difficulties of underwater mosaicing, such as non-planar sea-bottom, moving objects and severe illumination changes. This sequence also serves to illustrate the robustness and good performance of the presented algorithm