Evaluation of detection based on statistics and template matching classification on SAS AUV data

Abstract

A trend in Mine Counter Measurement is towards Remotely Operated Vehicles or even Autonomous Underwater Vehicles (AUV) to increase the stand-off and reduce the risk for the personnel. The development of Synthetic Aperture Sonar (SAS) means that range independent high resolution images can be generated from an AUV. Hence the operator will be presented with a large amount of data to process. If the detection and classification process could be carried out automatically the workload would be reduced. Thus research into computer aided detection and computer aided classification (CAD/CAC) algorithms is an opportunity to improve the mine-hunting process. As indicated by the name CAD/CAC, the process is here considered as two consecutive steps. Beginning with an anomaly detection step where image pixels not belonging to the expected bottom reverberation pattern are identified. Anomaly pixels are then grouped into objects which are classified as a specific target. A detection method based on statistical properties and a classification method based on template matching are described. The methods are evaluated on SAS AUV data. The SAS AUV experiment was carried out in the Stockholm archipelago, partly in an area used by the Swedish navy for minehunting training. The seabed consists of soft postglacial clay and mud, a very suitable bottom for minehunting. After identifying six mine and mine-like objects the AUV was ordered to circulate around each target several times at different ranges. The purpose was to catch the angle dependence of the targets. The AUV was sailed at a constant depth and each object was illuminated from eight different horizontal angles repeated at three different ranges giving variation in elevation angle. Results are presented for the evaluation of the methods. Both detection based on statistics and template matching methods perform very well on this dataset. However, the used dataset has a high SNR and a high resemblance between training and testing data. Additional datasets are desirable including more difficult bottom characteristics. Furthermore, the used methods are applicable to side-scan sonar images.