Abstract
Building 3-D object models from 2-D images is a key capability for target classification and identification, reacquisition, and environmental mapping, among many applications in underwater with poor visibility. We present a novel approach in utilizing multiple 2-D forward-look sonar images from known sonar poses to localize an acoustically opaque target and reconstruct its 3-D shape. Based on projections onto various images, the 3-D space not occupied by an imaged target within the sonar field of view is sequentially carved out, leaving the remaining volume as the estimate of the 3-D object region. The estimation generally improves with information from new distinct views, and moreover with images acquired through sonar roll motions, rather than circumnavigating the target. Computer simulations allow assessing the convergence properties and performance of the approach for convex and concave polygons. Additionally, results from experiments with real images of amorphous coral rocks and a miniature wood table demonstrate performance in the 3-D modeling of small objects with varying reflectance properties.
