Vision-Based Detection of Estrus-Related Interaction Behavior in Hanwoo Cattle Using Lightweight YOLO Models

Timely estrus detection determines reproductive efficiency in Hanwoo cattle, yet visual observation is increasingly difficult as herd sizes grow and standing behavior becomes less consistently expressed. Fixed farm cameras offer a non-invasive alternative, but mounting, chin-resting and sniffing involve two overlapping animals and occur rarely, complicating automated detection. This study evaluated how annotation strategy, augmentation policy and detector choice affect detection of these cues. A surveillance dataset of 846 red–green–blue (RGB) frames, containing 200 estrus-interaction instances against 2180 posture instances, was recorded at a commercial farm. Three annotation strategies, partial bounding-box label (PBL), full single bounding-box label (FSBL) and full dual bounding-box label (FDBL), were combined with three augmentation stages and three lightweight detectors (YOLOv5n, YOLOv8n, YOLO11n), giving 27 configurations. Rare-class performance varied approximately twice as widely as overall performance and depended more on annotation and augmentation than on detector family. PBL maximized sensitivity (recall 0.718), FSBL maximized localization quality (average precision, AP50 0.798) and overall accuracy (mean average precision, mAP50 0.864), and FDBL preserved relational context (recall 0.700, AP50 0.775). Annotation and augmentation should be selected according to whether the purpose is screening candidate interactions or supporting expert confirmation. Detections represent a behavioral proxy, not physiological confirmation of estrus.