Hidden Markov Models for detecting anomalous fish trajectories in underwater footage

Abstract

In this paper we propose an automatic system for the identification of anomalous fish trajectories extracted by processing underwater footage. Our approach exploits Hidden Markov Models (HMMs) to represent and compare trajectories. Multi-Dimensional Scaling (MDS) is applied to project the trajectories onto a low-dimensional vector space, while preserving the similarity between the original data. Usual or normal events are then defined as set of trajectories clustered together, on which HMMs are trained and used to check whether a new trajectory matches one of the usual events, or can be labeled as anomalous. This approach was tested on 3700 trajectories, obtained by processing a set of underwater videos with state-of-art object detection and tracking algorithms, by assessing its capability to distinguish between correct trajectories and erroneous ones due, for instance, to object occlusions, tracker mis-associations and background movements.