This project explores the use of a unsupervised hierarchical Bayesian non-parametric topic modeling for automatic characterization and anomaly detection in data observed by a mobile exploration robot. The proposed model takes into account the spatio-temporal context of the observations, and encourages sparse representation of the data, in semantic space, through the use of Dirichlet process priors. We use the posterior topic distribution to generate visualizations of the observation data, which we hypothesize can speed up discovery of new processes and phenomena.

The following visualizations correspond to 16 dives by the SeaBed AUV at Hannibal seamount in Panama. These visualizations demonstrate that the learned model is automatically able to characterize the underlying terrain, and detect anomalous flora in image data collected by an underwater robot.

Authors: Yogesh Girdhar, Walter Cho, Jesus Pineda, Hanumant Singh