Pattern classification for enhanced detection using the underwater radar-encoded laser system

In this work we investigate the use of pattern classification algorithms to enhance detection performance of the underwater radar-encoded laser system. A challenge encountered with this system is the automatic detection of the return from an underwater object in highly-scattering and/or low signal-to-noise ratio (SNR) conditions. Previous efforts were largely based on threshold detection and result in detection errors in such challenging conditions. Other efforts attempt to use signal processing to remove scatter returns, but this does not address low SNR cases. We take a different approach here, investigating the use of machine learning to develop classifiers which combine various shape and statistical features to discriminate between object and non-object returns. Such pattern classifiers are commonly used in a variety of applications; the novelty in this work is applying such techniques to the problem of automatic object detection in a degraded visual environment, namely turbid water. We describe our framework and features, then demonstrate the performance of three pattern classification detectors using a series of test data collected in a variety of water conditions in a laboratory test tank. All three pattern classification detectors outperform a standard detection method. There are subtle performance differences between the classifiers that may result in application-specific tradeoff considerations.