Abstract
Interpretation of acoustical images to understand the characteristics of a scene and identify under-water objects has become a problem of high importance in a variety of applications. In the meantime, the computer processing of forward-scan sonar video imagery has enabled significant capabilities in a wide variety of underwater operations within turbid environments. Further developments are inevitable with more accurate representation of the image formation principles of these sonars. Sonar images are formed by transmitting acoustical pulses and measuring the reflected sound power from the scene surfaces. The recorded signal encodes information about the shape and material properties of these surfaces. The inherent ambiguities in the interpretation of the 3-D world based on visual cues in a 2-D forward-scan sonar image arise as a result of both operating as a ranging device and loss of elevation angle information due to the projection geometry. One aspect of this work is modeling the image formed by a new class of high-resolution 2-D forward-scan sonar systems which supports our complementary aim of simulating FL sonar images. The other aspect is reconstructing 3-D objects by either recovering the unknown zenith angles from image brightness of a single sonar view or by applying a carving based technique on multiple sonar views. To elaborate, we first provide a detailed derivation for modeling the image formed by a forward-scan sonar. Our initial model derived for a small surface patch is generalized to handle multiple patches and account for the signal type and pulse-width. We exemplify and simulate our model for a circular target profile and validate it by comparing with the true intensities measured from two isolated cylindrical targets with circular cut sections. Then we demonstrate application of our model for multi-path reflections from bottom surfaces with hemi-cylindrical and hemi-spherical targets laid on the bottom. Next, we propose a method to recover the unknown zenith angles from image brightness and thus reconstruct 3-D objects. Our method applies to a single forward-scan sonar image, assuming that the scene objects have smooth surfaces that vary monotonically in terms of distance from the sonar, and cast visible shadows on a flat background. We present the results of experiments with real data to demonstrate the performance of our 3-D reconstruction technique. Finally, we develop a method for determining the 3-D shape of a target from multiple 2-D forward-scan sonar images at given sonar poses. Our strategy is based on sequentially carving the non-target space projecting onto the so-called dark pixels on various images. We demonstrate how the remaining space, deemed to be occupied by the target, approaches toward the object volume, given enough views. The images acquired through sonar roll motions, rather than circumnavigating the target, are critical for more precise reconstruction. Employing our image simulator, we perform extensive tests to assess the convergence properties of our 3-D approach for convex and concave polygons. Further experiments with real data, collected by DIDSON, ARIS, and Blueview sonars demonstrate the performance from image sequences of four amorphous coral rocks and a wooden table. Potential applications of our techniques include target localization and re-acquisition, classification and recognition, obstacle avoidance and precision navigation, scene mapping in support of studying underwater ecosystems, and investigation of deep-water archaeological sites within turbid environments.