APPLYING FOUNDATION SEGMENTATION MODELS TO ASSESS POULTRY PHENOTYPES AND BEHAVIORS

Advanced computer vision and deep learning methods have been developed to enhance precision poultry farming through automated health status identification and activity monitoring across multiple poultry species. A modified Segment Anything Model (SAM) pipeline, combined with pre- and post-processing techniques but without extensive model training, achieved 84.4% segmentation success for cage-free laying hens in thermal images, enabling non-invasive extraction of body temperature to support early health assessment. A hybrid YOLOv7 + SAM model using bounding box prompts achieved 98.0% segmentation accuracy, allowing precise individual identification. Additionally, an open-source, user-friendly Streamlit platform integrating SAM2 was developed to enable non-technical researchers to track animal activity across different species directly from video data without any model training or manual labeling. These tools minimize manual intervention, reduce animal stress, and improve decision-making by providing automated monitoring of phenotypic and behavioral indicators, with broad applicability in precision livestock farming and smart agricultural systems.