Lessons learned from the DSTO sonar ATR challenge / by John A. Fawcett and Warren A. Connors. : D68-6/184-2013E-PDF

In May 2013 DRDC Atlantic participated in a Defence Science and Technology Organization, Australia (DSTO)-sponsored sonar automatic target recognition (ATR) challenge. As part of this challenge, DRDC Atlantic and the other participants received approximately 11000 Remus/Marine Sonic sonar images in bitmap format as well as corresponding navigational information. A set of 654 sonar files and ground truth data for minelike objects and objects of interest were supplied for training the ATR methods. A larger set of files containing minelike and objects of interest were then used for testing the methods. The locations of the objects in the testing set were not known to the participants. A file with the locations of the DRDC Atlantic ATR detections was emailed to DSTO and the results were evaluated on the basis of the ground truth locations for the testing set. In this report, we describe our initial processing of the data to obtain the results which were sent to Australia. The set of sonar files provided were very challenging, including significant clutter regions and sand ripples. The results from the test evaluation indicated a low detection rate. After the test evaluation, Australia kindly supplied a file with the ground truth locations for the testing set. This report describes the subsequent analysis of our ATR approach. It was found that by dropping one of the constraints originally used in the ATR algorithm that the results were significantly improved.

Lien permanent pour cette publication :
publications.gc.ca/pub?id=9.821357&sl=1

• Format MARC XML
• Format MARC HTML

Demander des formats alternatifs