Monitoring horse behaviour with deep learning models.

Abstract

Detailed analysis of stabled horse behaviour can reveal accurate information about its well-being. Advances in deep learning now allow these behaviours to be tracked without being invasive through the use of video data. This study evaluated a convolutional neural network for recognising standing, lying, and drinking behaviours in a horse housed in a wooden stall and recorded continuously over 29 consecutive days. Model predictions were compared with manually annotated ground truth data. Standing was detected with high precision (97.5%) and high recall (89.2%). Lying behaviour was classified with high precision (92.8%) but lower recall (63.1%). Activity patterns showed that standing dominated daily time budgets (>85%), lying accounted for 5-10%, and drinking occurred most often between 04:00 pm and 10:00 pm. These results demonstrate that deep learning can classify common equine behaviours from video, supporting its use in automated welfare monitoring. Future evaluations will explore the recognition of less frequent behaviours.