A family of algorithms for the automatic detection, isolation, and fusion of object responses in sonar data

A new, fully automated approach for obtaining an object’s acoustic-color template –uncontaminated by background seafloor – from low-frequency sonar data is proposed. The developed family of algorithms eliminates several time-consuming tasks that previously required execution by humans. First, a fast object detection algorithm generates a set of candidate objects of interest. For a given object, geometric and physics insights are then leveraged to perform an image-domain segmentation followed by a more precise time-domain segmentation. Finally, the masked time-domain data is converted to an acoustic-color representation displaying the object scattering amplitude as a function of frequency and aspect. When the scene has been interrogated at multiple aspects, the individual acoustic-color templates can be fused via another new algorithm to form a single unified template. As a result of the end-to-end automation, the creation of desired sonar data products is greatly accelerated. This fact in turn enables both the generation of sizable data sets for classifier training, and the realization of large-scale experiments to study the effects of various quantities. Example results of the algorithms on sonar data are provided. The environment’s impact on an object’s acoustic-color template is also shown quantitatively.