Machine Learning in Commercial Poultry Farming: Computer Vision-Assisted Broiler Count Verification During Manual Poultry Harvest Weighing

Abstract

Commercial poultry harvesting in many Philippine farm settings still depends on manual bird handling, weighing, and record verification procedures, particularly where conveyor- based systems remain too costly for routine use. These methods place substantial demands on workers and may prolong the time broilers spend in a stressful handling environment. This study evaluates oriented bounding box (OBB) object detection models for harvest-stage monitoring in a practical weighing-in and count verification scenario. A camera positioned near the weighing station captures broiler images while a computer vision pipeline performs automated detection of broiler heads. OBB detection is used to better represent the irregular orientations and overlapping positions of birds during manual handling. The study compares YOLO-based OBB detectors in terms of detection accuracy, localization quality, and count-verification relevance under operational farm conditions. Held-out test results showed that the medium model achieved the strongest pure localization performance, while the small and nano models provided stronger count-verification behavior at the frame and event levels, respectively. Unlike prior work focused on feeding, drinking, camera placement, and other poultry monitoring scenarios, this study examines computer vision performance during harvest operations, where count verification, workflow visibility, and documentation are directly relevant to farm management. The proposed approach is intended to support workers during verification while giving farm owners and veterinarians better visibility into harvest activity. By improving operational efficiency, the system may also help shorten manual weighing and count verification workflows, with potential welfare benefits for broilers during harvest.