79. Detection of Sow Feeding- and Drinking-related Behaviors in Farrowing Crates Using Oriented Bounding Boxes.

Abstract Sow feeding and drinking behavior during farrowing can provide useful management information, but continuous observation is difficult in commercial systems because of labor constraints and herd size. Camera-based monitoring offers a scalable alternative, although detection from overhead farrowing-crate images is complicated when sows are positioned diagonally relative to the camera frame. In those cases, conventional axis-aligned bounding boxes may include substantial background near the feeder or drinker, which can hinder discrimination among visually similar behaviors. This study evaluated the use of oriented bounding boxes for detecting sow feeding- and drinking-related behaviors from overhead video recorded in individual farrowing crates. The study was conducted at the Imported Swine Research Laboratory at the University of Illinois Urbana-Champaign across three farrowing trials involving 24 sows, with one overhead camera per crate. Video frames were annotated into three classes: Active_Feeding, At_Feeder_Not_Feeding, and Drinking. A total of 228 images from two trials were used for model development, and 54 images from the remaining trial, representing eight sows, were reserved for independent evaluation. Hyperparameters were selected using a structured grid search during model development. On the held-out trial, the model achieved 0.84 precision, 0.91 recall, and 0.91 mean average precision. Drinking showed the highest recall, whereas At_Feeder_Not_Feeding had lower precision, likely because of visual similarity to active feeding. These results support the feasibility of using oriented bounding boxes to detect feeding- and drinking-related behaviors in overhead farrowing-crate images. Additional testing with larger and more diverse datasets is needed to determine robustness and comparative performance against conventional annotation approaches.