README.md

PGH — Automated Chicken Feeding Count

Automated two-stage computer vision pipeline using YOLOv8 to count chicken heads at feeders from night-vision camera frames as a behavioural indicator of gut health in broiler chickens.

---

Overview

This project processes continuous IR camera footage from a broiler chicken farm to automatically count how many chickens are feeding at any given time. Feeding frequency is used as a non-invasive behavioural proxy for early gut health monitoring.

The pipeline runs in two stages:

1. Feeder segmentation — locate the feeder regions in each frame
2. Head detection — count chicken heads inside those feeder regions

---

Pipeline

Raw frame (IR camera)
        ↓
Stage 1 — Feeder Segmentation (YOLOv8n-seg)
        ↓ Top 2 feeders by bounding box area
Stage 2 — Feeder ROI Crop
        ↓
Stage 3 — Head Detection (YOLOv8s, conf = 0.50)
        ↓
Output — Head count per feeder per frame → CSV + time-series graph

---

Notebooks

Notebook	Description
01_train_feeder_seg.ipynb	Train YOLOv8n-seg on 40 labeled feeder images
02_extract_roi_crops.ipynb	Use feeder model to crop top 2 feeder ROIs from raw images
03_train_head_detection.ipynb	Train YOLOv8s on 59 labeled feeder ROI crops
04_run_count_pipeline.ipynb	Run full pipeline on all 223 video frames, export CSV + graph

---

Dataset

Model	Images	Split	Class
Feeder segmentation	40	28 train / 8 valid / 4 test	Feeder
Head detection	59 (after augmentation)	59 train / 6 valid / 3 test	Head

• Images sourced from a top-view IR night-vision camera (grayscale/BW)
• Labeled using Roboflow
• Frames extracted from continuous video at ~5-minute intervals (223 total frames)

---

Results

Feeder Segmentation (YOLOv8n-seg)

Metric	Score
mAP50 (box)	0.9950
mAP50 (mask)	0.9950
Precision	0.9955
Recall	1.0000

Head Detection (YOLOv8s)

Metric	Score
mAP50	0.9950
mAP50-95	0.5355
Precision	1.0000
Recall	0.9357

Full Pipeline Output (223 frames)

Stat	Value
Average heads per frame	0.37
Max heads in one frame	3
Frames with 0 heads	154

---

Requirements

All notebooks run on Google Colab (T4 GPU). No local setup needed.

ultralytics
torch
PIL
matplotlib
numpy
PyYAML

Install in Colab:

!pip install ultralytics --quiet

---

How to Run

1. Upload PGH.zip (feeder dataset), PGH_head_labelled.zip (head dataset), and PoultryImage.zip (video frames) to your Google Drive root
2. Run notebooks in order: 01 → 02 → 03 → 04
3. Each notebook saves outputs (model weights, crops, CSV, graph) directly to Google Drive
4. Final outputs: PGH_head_counts.csv and PGH_feeding_graph.png

---

Project Structure

PGH-Chicken-Feeding-Count/
├── 01_train_feeder_seg.ipynb
├── 02_extract_roi_crops.ipynb
├── 03_train_head_detection.ipynb
├── 04_run_count_pipeline.ipynb
└── README.md

---

Author

Anas Bin Harun
Texas A&M University