Knowing Every Hen: Multi-Camera Poultry Monitoring
2025–present · Prairie View A&M University
NSF-funded research at PVAMU: two synchronized cameras, homography fusion and color re-identification track individual laying hens through their day — feeding, drinking, laying — to catch problems like bumblefoot before a human would.
The problem
The industry’s shift to cage-free housing is good for hens, but it broke the old way of monitoring them: when 60 birds share a pen, nobody knows which hen stopped eating yesterday. Bumblefoot — a painful bacterial foot infection — is a perfect example. It’s common, it’s treatable, and it usually gets noticed only when a bird is already limping.
Two cameras, one hen
Our setup pairs a top-down camera with a side-view camera over a 2 m × 3.5 m research pen. A homographic projection maps both views onto a single coordinate system of the pen floor, and color-based re-identification keeps each hen’s identity stable over long periods — so the system can say not just "a hen is at the feeder" but "hen #14 has been at the feeder half as long as usual".
The YOLOv12 segmentation model behind it reaches 0.976 box / 0.980 mask mAP@50. From the tracks we extract per-bird behavioral metrics — feeding frequency, drinking duration, laying activity — and a time-series model turns those into a bumblefoot probability for each bird.
Running at the edge
Everything runs in the barn, not the cloud: a Raspberry Pi 5 with a 26-TOPS Hailo AI accelerator handles real-time inference on-device. The tooling — pre-trained detectors and dataset-creation utilities for multi-animal pose estimation — is open source on GitHub.
The Spirulina experiment
The same rig powers a second study: an AI-IoT framework that pairs the vision system with environmental gas sensors to measure how Spirulina-supplemented diets change bird activity and emissions. Both papers are currently under review — this one was presented as a poster at AI in Agriculture 2026.