On reconstruction of 3-D volumetric models of reefs and benthic structures from image sequences of a stereo rig

Acquiring 3-D computer models of objects from images has been an active computer vision research topic. A common approach uses overlapping images from different views around the object, and attempts to fuse all the information to construct the model by volume intersection. Most current methods have in common that the camera poses for these views are known. To achieve this, a monocular camera circles the target object on a robot arm, or equivalently, views a target on a rotating platform (e.g., a turntable), while the images are acquired at known intervals. We are interested in the deployment of a submersible platform equipped with stereo vision to construct 3-D volumetric models of benthic habitat objects, e.g., coral reefs and other seafloor structures. In this application, we would automatically acquire continuous 360-degrees stereo views of object(s) of interest, by processing the images online to estimate and control camera trajectory and views. In particular, we show that the redundancy from binocular cues improves the robustness and accuracy. Upon completion of the image data acquisition, a bundle adjustment formulation is proposed to recompute the trajectory and the stereo rig poses, as well as to generate object model(s).