3D target shape from SAS images based on a deformable mesh

Abstract

The seafloor can nowadays be scanned with side-looking sonar that provides a very high resolution over a large swath, which has proved beneficial for underwater target detection and classification. For systems operated at hundreds of kilohertz, one may obtain centimeter resolution in the range and along-track direction. For these systems, the third dimension, height, is usually resolved much worse. Since the 3D shape is regarded a valuable clue in underwater target classification, it is important to extract height information as best

as possible. In this paper a new method for deriving 3D information from non-interferometric SAS images is described. The method is experimentally applied to multi-view reconstruction of calibrated target shapes imaged by the MUSCLE vehicle, and the results are compared to the actual dimensions of the observed objects in order to quantify the reconstruction accuracy. The reconstruction algorithm is a shape-from-shading approach that uses a

deformable mesh to preserve surface continuity while enforcing observational constraints. The technique is shown to have an important impact in object classification, both as a standalone method and combined with other 3D imaging techniques such as interferometry. Implications to target identification and vehicle autonomy are discussed.