Geometric Feature Representation in Active Sonar Signal Processing

Proper feature representation of target signatures have been an active challenge to autonomous target recognition for many decades. The fundamental bottleneck is uncertainty of target-specific features in the face of unknown geometry, ping angles, as well as structured environmental interference. In particular, non-linear overlap between target features as well as with features generated due to spectral interference from the environment make it difficult to create robust feature dictionaries that are persistent against environmental uncertainties and ping directions. In this work, we introduce a novel feature representation based on braid manifolds that provides a mathematical basis to render and quantify non-linear target features that may overlap in the acoustic color and other multi-dimensional representations of active sonar returns. The ideas are introduced as a representational framework, and as such this work should be taken in the spirit of a position paper rather than a comprehensive report. Preliminary results based on sonar target returns from the Malta Plateau field experiment are presented.