Perception-based automatic classification of impulsive-source active sonar echoes

Abstract

Impulsive-source active sonar systems are often plagued by false alarm echoes resulting from the presence of naturally occurring clutter objects in the environment. Sonar performance could be improved by a technique for discriminating between echoes from true targets and echoes from clutter. Motivated by anecdotal evidence that target echoes sound very different than clutter echoes when auditioned by a human operator, this paper describes the implementation of an automatic classifier for impulsive-source active sonar echoes that is based on perceptual signal features that have been previously identified in the musical acoustics literature as underlying timbre. Perceptual signal features found in this paper to be particularly useful to the problem of active sonar classification include: the centroid and peak value of the perceptual loudness function, as well as several features based on subband attack and decay times. This paper uses subsets of these perceptual signal features to train and test an automatic classifier capable of discriminating between target and clutter echoes with an equal error rate of roughly 10%; the area under the receiver operating characteristic curve corresponding to this classifier is found to be 0.975.

Performances of human listeners and an automatic aural classifier in discriminating between sonar target echoes and clutter

“…An approach that has been investigated recently to address this challenge was to exploit the sensitivity of human hearing for distinguishing between these two classes of echoes. A prototype automatic aural classifier was implemented that used features associated with the notion of timbre in musical acoustics as potential discriminating cues (Young and Hines, 2007). The results of tests that used a sample of echoes gathered from experimental data looked promising.…”

“…As mentioned in the preceding text, a sample of echoes taken from an actual experiment at sea served as the basic data for the the development of an automatic aural classifier (Young and Hines, 2007). These same data were used as stimuli for a human-performance study, the results of which were used to produce a baseline.…”

Feature-aided initiation and tracking via tree search

“…While perceptual features (aural features) of sonar signals have been recently applied for target classification [1], [2], the combined tracker-classifier technique, called feature-aided tracker hereafter, incorporates feature information to improve tracking performance. The proposed tracking architecture uses treesearch track initiation [6] for initializing the target tracks, the tree-search tracker with likelihood surface discretization [4], [3] for tracking initialized targets, and embedded track management [5] for identifying/removing false tracks.…”

Automatic target classification for low-frequency anti-submarine warfare sonars

“…Some are based on analysis of lowlevel sonar data, such as matched filtered and normalized data [1], [2], [3], [4], [5], [6], while others focus on higher level data such as track information [7], [8], [9]. Also the nature of the classification schemes vary, e. g. pattern recognition [10], inversion [3], track classification [8], spectral characterization [1], [5], principal component analysis [5], [9], and machine learning approaches like decision tree learning [11], various kinds of neural networks [9] and evolutionary approaches [12].…”