Abstract
We present a novel method for determining the 3-D shape of a target from multiple 2-D forward-look sonar images captured at known sonar poses. It enables both the 3-D localization and modeling of scene objects made from any material visible to the sonar. The method is based on the sequential carving of the non-target space based on projections onto various images, leaving the volume deemed to be occupied by the object of interest. While the estimation generally improves with information from new distinct views, the images acquired through sonar roll motions at only a few views, rather than circumnavigating the target, are critical for more precise reconstruction. Experimental results include computer simulations to assess the convergence properties of our approach for convex and concave polygons. Additionally, we demonstrate performance in experiments with real images of two amorphous coral rocks. Potential applications of our technique include target classification, recognition, re-acquisition and avoidance, as well as 3-D environmental mapping.
