Ambiguity reduction of underwater targets in framework of topic modeling

Abstract

An unsupervised track classification approach based on appropriate discriminative and aggregative features derived from beamformed and normalized matched-filtered data is applied to sonar multistatic tracking and extended to include discretised track velocity and heading rate. A clustering algorithm based on the Latent Dirichlet Allocation model is proposed. It is demonstrated how low-level, highly variable and non-stationary data components can be combined through an increased abstraction level with higher level kinematic tracking features. Improved discrimination of tracks associated with both stationary and moving scatterers is demonstrated.