FEATURE-AIDED TRACKING FOR MARINE MAMMAL DETECTION AND CLASSIFICATION

摘要

This paper presents a method to detect and classify odontocete echolocation clicks as well as to estimate the number of animals that are vocalizing. A transient detector using the Page test [1-3] is used to extract the clicks: the click time, the click duration, the click amplitude and the spectral information of the clicks are extracted,. A probability distribution over the species is assigned to each click, based on the spectral information of the click. The estimation of the number of animals is done using feature-aided multi-hypothesis tracking (MHT) algorithms. The association is based on the assumptions of slowly-varying click amplitude and intra-click timing [4-5]. This woik has been done on the dataset provided by the organizers of the 3rd International Workshop on the Detection and Classification of Marine Mammals using Passive Acoustics, Boston, July 2007. This dataset consists of training and test data; the training data includes vocalizations of three species: Blainville's beaked whale (Mesoplodon densirostris), Risso's dolphin {Grampus griseus) and short-finned pilot whale (Globicephala macrorhynchus).