GrazeAI — Livestock Computer Vision
GrazeAI applies trained YOLO11s object detection and Behavior Engine V3 temporal tracking to help farms observe group livestock postures (Eating, Lying, Standing) across images and barn/yard video.
Current Frame Behavior Classification
Output from the preserved POST /detect endpoint using trained YOLO11s weights.
Technical Transparency Note: These figures reflect a specific historical validation dataset on our trained YOLO11s weights (best.pt), not guaranteed production accuracy across all farms. Real-world performance depends on camera placement, lighting, occlusion, and species, and typically requires site-specific validation.
How GrazeAI Works Today
GrazeAI is implemented as a FastAPI computer-vision service powered by PyTorch and Ultralytics YOLO11s weights (best.pt), integrated directly into the FarmOS Core API.
- Three Behavioral Classes: Class 0:
Eating, Class 1:Lying, Class 2:Standing. - Image Inference Endpoint (POST /detect): Accepts uploaded livestock photographs and returns bounding boxes, confidence scores, and behavioral counts.
- Video Analysis & Behavior Engine V3 (POST /detect-video): Processes uploaded video files with track confirmation thresholds (
MIN_CONFIRM_FRAMES=15,SMOOTHING_WINDOW=9) to generate annotated MP4 outputs and timeline summaries. - Historical Validation Metrics: In historical evaluation, the model reported approximately 0.793 precision, 0.666 recall, 0.797 mAP@50, and 0.544 mAP@50–95. These metrics reflect a specific historical evaluation dataset rather than guaranteed production accuracy across all camera setups.
Explore a GrazeAI Pilot for Your Livestock Operation
Share your herd type, barn or feedlot layout, and whether you have existing IP cameras or sample video footage for a technical feasibility assessment.