CabriTrack: Accelerometer data for automated behavioural monitoring of grazing Creole goats

Abstract

The availability of sensors and AI-based methods offers new perspectives for monitoring animal behaviour. In particular, accelerometers can record individual acceleration data for weeks, which can then be used to identify the activity of the animal. Several research articles have demonstrated the capacity of this technology, particularly using machine learning or deep learning, for behaviour estimation. These techniques need high-quality datasets to train and validate the models, particularly with a great diversity of examples for each considered behaviour and recorded animals. The diversity of the data is an important prerequisite for deploying these solutions at a large scale. In this context, the dataset presented here contains more than 144 hours of tri-axial accelerometer data, collected from 59 different animals. The data were collected from March 2023 until March 2024. Two to five adult Creole goats were equipped with an accelerometer on one horn and allowed to graze in a small experimental pasture. While grazing, the behaviour of the animals was recorded with a CCTV camera. The videos were then manually annotated using the software Boris to identify the behaviour of each animal when it was possible to do so. Five behaviours were considered: ruminating/chewing, grazing, resting, displacement, and other, which includes behaviours such as scratching or fighting with a congener. Finally, the behaviour sequences were associated with the corresponding acceleration sequences based on a time synchronization procedure, so that each acceleration sequence is associated with one behaviour. This dataset can be used to train and evaluate any prediction methods for behaviour prediction from acceleration data using tri-axial accelerometers mounted on the horn of grazing Creole goats.