Braid feature extraction and representation of active sonar data

Abstract

As an active sonar system pings an object of interest, the moving platform collects data descriptive of both the target and background clutter. We hypothesize that the relevant features of the target echo tend to persist and smoothly morph over time in some domain as the sonar platform continues its route. This work leverages the preceding and presents initial findings of a feature extraction algorithm resulting in representations of active sonar data that persistently appear and are physically and statistically informed. This is done by isolating the targets’ features from background clutter using a statistically meaningful threshold, extracting features onto a manifold representation, and mathematically describing and quantifying the extracted features. The background clutter is modeled using a Rayleigh probability density function, and the echo is isolated by keeping statistically significant responses. Feature extraction is performed on a ping-by-ping basis by minimizing the angle between the isolated echo samples and assigning a pseudo probability to sequential samples. These methods are used to extract features from ping sequences that are unique to targets and quantified using a correlation metric. Results reported are on simulated data and experimental field data. [Work supported by NREIP, NDSEG, and ONR, Grant No. N00014-21-1-2420.]