Monitoring cattle behavior plays an important role in improving farm productivity, maintaining animal welfare, and supporting efficient management practices. This study presents a multi-view behavior recognition system that uses synchronized top-view and front-view CCTV footage, combined with deep learning techniques. The system includes four main components: cow identification, behavior classification, identity-behavior association using Intersection-over-Union (IoU), and a decision-level ensemble to combine information from both views. YOLOv8 models are applied separately to each camera angle to detect individual cows and classify six key behaviors: drinking, eating, standing, lying, riding, and chin resting, with the latter two being relevant for estrus detection. The system matches cow identities to their behaviors within each view and then integrates the results to produce a final activity label for each cow.
Hanpinitsak et al. (Fri,) studied this question.