Dirichlet process mixture models for autonomous habitat classification
Daniel Steinberg, Oscar R. Pizarro, Stefan B. Williams, Michael V. Jakuba · OCEANS'10 IEEE SYDNEY · 2010
There is a need for truly unsupervised approaches to understanding acquired data in autonomous exploratory missions with minimal, or zero, bandwidth communication. This paper presents results of using a Bayesian non-parametric Dirichlet Process mixture model - the Infinite Gaussian Mixture Model (IGMM) - for the classification of benthic habitats. The IGMM is trained completely autonomously, without being given labelled data, or knowing the number of habitats present. It is able to infer the number of habitats present in the training data, and is also able to infer the presence of habitats in the test data that were not present in the training data. This is a powerful model for entirely autonomous labelling of benthic datasets, and will be used as the basis of completely autonomous approaches to understanding data in the future.