Abstract
Modeling a 3-D object from 2-D images is useful in a variety of applications. Optical systems can be ineffective in certain underwater environments with turbidity. However, acoustic signals can penetrate through suspended particles in turbid waters. Thus, sonar imaging systems are preferred in this case.
Multi-beam forward scan (FS) sonar systems form images by transmitting multiple acoustic beams and measuring the time and strength of the acoustic returns from the scene surfaces. The elevation angle of the scene surfaces is lost during the 2-D image formation process, leading to the rather common many-to-one projection in FS sonar images.
Among a few methods utilizing multiple sonar images taken at known sonar poses to overcome the many-to-one projection ambiguity is a method based on the space carving (SC) paradigm. Here, a successive carving process is performed to remove the "non-target volume", leaving the remaining volume as the estimation of the volume occupied by the target. No image intensity information is used in this method. However, once a 3-D model has been derived, we can synthesize the 2-D FS sonar images utilizing an image formation model based on object shape and reflectance properties. The discrepancy between the synthesized and real sonar images provides guidance on how the model may be adjusted to achieve better consistency.
In this research, we utilize the 2-D FS sonar images captured by a DIDSON deployed near the sea surface, where multi-path reflections from the water surface offer additional visual cues. This scenario is common where we image targets, floating at or near the sea surface, e.g., ship hull, buoys, mines, and the likes. Our goal is to improve the reconstruction accuracy of an initial 3-D model derived from any other methods, here based on the SC solution.
The proposed iterative optimization process minimizes the discrepancy between the real sonar images and the synthesized views of the 3-D model. We demonstrate its effectiveness by performing experiments on simulated and real data.