default Risk maps for performance evaluation of autonomous mine search

Risk maps for performance evaluation of autonomous mine searchs. Gips, Bart ; Strode, Christopher ; Dugelay, Samantha. CMRE-FR-2018-009. February 2019.

Autonomous underwater vehicles (AUVs) are used with increasing frequency in mine-hunting operation to reduce risk to personnel conducting mine clearance operations. When equipped with sonar imagery sensors such as side-scan sonar (SSS) or synthetic aperture sonar (SAS), AUVs can generate highresolution images of large patches of sea floor, which could then be passed to an operator. When combined with automatic target recognition (ATR) algorithms, AUVs can autonomously scan a survey area for the presence of mines. If we want to rely on such an autonomous survey, we need methods to quantify the AUV's performance. In the current work we propose a framework with which we can construct a map of the survey area that expresses the residual risk. Such a residual risk map (RRM) expresses the estimated probability of a false negative, i.e. the probability of the ATR algorithm not detecting a hypothetical target present at each location in the map. RRMs can aid in estimating the percentage of the targets that have been detected. Secondly, an RRM can be useful for communicating which parts of the survey area should be revisited for closer inspection, or if time and resources do not permit this, which areas should be avoided. The algorithm through which we can construct the RRM is based on through-the-sensor features such as target range, ping-to-ping correlation, and metrics that capture the texture of the image. The algorithm is flexible in the sense that it is agnostic to the specific sensor hardware and ATR type used, as long as we have reference data to constrain model parameters. This means that RRMs can be generated by multiple AUVs, if we want to deploy multiple assets in parallel to accelerate the survey. Since we make use of a probabilistic framework, it is possible to easily combine these individual RRMs into a single RRM of the area in a principled way. This study includes initial results based on reference SAS data gathered by the CMRE MUSCLE AUV during SHOEX17 off the coast of Elba Island, Italy, as well during ESPMINEX18 off the coast of Mallorca, Spain.