An adaptively generated feature set for low-resolution multifrequency sonar images

Abstract

Many small Unmanned Underwater Vehicles (UUVs) currently utilize inexpensive, low resolution sonars that are either mechanically or electronically steered as their main sensors. These sonars do not provide high quality images and are quite dissimilar from the broad area search sonars that will most likely be the source of the localization data given to the UUV in a reacquisition scenario. Therefore, the acoustic data returned by the UUV in its attempt to reacquire the target will look quite different from the original wide area image. The problem then becomes how to determine that the UUV is looking at the same object. Our approach is to exploit the maneuverability of the UUV and currently unused information in the echoes returned from these Commercial-Off-The-Shelf (COTS) sonars in order to classify a presumptive target as an object of interest. The approach hinges on the ability of the UUV to maneuver around the target in order to insonify the target at different frequencies of insonification, ranges, and aspects. We show how this approach would allow the UUV to extract a feature set derived from the inversion of simple physics-based models. These models predict echo time-of-arrival and inversion of these models using the echo data allows effective classification based on estimated surface and bulk material properties. We have simulated UUV maneuvers by positioning targets at different ranges and aspects to the sonar and have then interrogated the target at different frequencies. The properties that have been extracted include longitudinal, and shear speeds of the bulk, as well as longitudinal speed, Rayleigh speed, and density of the surface. The material properties we have extracted using this approach match the tabulated material values within 8%. We also show that only a few material properties are required to effectively segregate many classes of materials.