Time-Series Processing and Classication for the Automatic Assessment of Wild Animals Activities

Ethology is the science that aims to understand how and why a certain behavior occurs. In order to be able to answer these questions, experts must observe the species of interest for a long time in order to have a sufficiently large database to represent its typical behavior. Once these data have been collected, it is possible to move on to the next phase, the purpose of which is to be able to infer the causes and dynamics that bind different behaviors. If in the second phase the expert's knowledge and skills appear fundamental, in the first phase the human presence causes several problems. For this reason, the proposed work aims to provide a tool to support the expert during the phase of annotation of the activities of the species considered. In detail, it is a methodology to produce machine-learning frameworks capable of recognizing the behavior assumed by the animal, analyzing the time-series measured by sensors such as tri-axial accelerometer and GPS. The absolute novelty that characterizes the proposed methodology is the fact of being focused on the pre-processing phases, in particular, composed of filtering and inconsistencies detection and features extraction, where the role of the data-analyst is central. The classifier is left only with the task of drawing the final conclusion. This methodology has been applied in a real case study, demonstrating the gain in accuracy that a data-driven approach can bring compared to the works already present in the state of the art.