Deep Ranking-Based DOA Tracking Algorithm

Renana Opochinsky, Gal Chechik, Sharon Gannot

Abstract: In this study, we present a weak-supervised deep neural network-based tracking algorithm for a moving source. A triplet-loss network is trained with instantaneous spatial features to estimate the time-varying DOA. The core idea is to minimize the use of labeled samples (i.e. samples which are accurately localized, and difficult to acquire) by using instead partial knowledge drawn from an unlabeled, and much larger, dataset in which only the relative spatial ordering between the samples is known. We use a deep learning architecture that stochastically combines a triplet-ranking loss for the unlabeled samples and a spatial loss for the labelled samples and learns a nonlinear deep embedding that maps acoustic features to an azimuth angle of the source. We show that it is unnecessary to train the network with a large number of random trajectories of a moving source, and that triplets of static sources from the same locus, which can be more easily acquired, are sufficient. A simulation study demonstrates the applicability of the proposed method to dynamic problems.