DynROI-Hybrid: Dynamic modeling of functional regions for feeding and drinking behavior detection in laying hens

The automatic identification of feeding and drinking behaviors in commercial laying hen houses remains challenging due to of high stocking densities, frequent occlusions, and the pendular movement of feeding and drinking equipment, which compromise approaches that rely solely on fixed regions or bounding boxes. This study proposes the DynROI-Hybrid framework, which integrates object detection with YOLOv11, selective segmentation with Segment Anything Model 2 (SAM 2), and dynamic modeling of functional interaction regions associated with feeders and drinkers. The methodology combines spatial criteria (overlap, proximity, and intersection) with temporal stabilization via hysteresis, enabling distinction between true interaction events and mere spatial proximity. Results demonstrate robust performance, achieving 94% overall accuracy and a Cohen’s Kappa coefficient of 0.91 compared to gold-standard validation. The proposed approach reduces false positives caused by equipment movement and structural resource geometry while maintaining computational feasibility for Edge AI applications. The method contributes to automated monitoring of feeding-related behaviors in commercial poultry production systems, with potential applications in welfare assessment and decision support within Precision Livestock Farming.