NVIDIA Innovation Lab Application

Real-time flock intelligence, built for the barn.

Nesteye combines computer vision, edge AI, and adaptive barn controls to detect risk, estimate flock weight, and intervene before preventable loss occurs.

Read the full project brief

The project

A continuous view of flock conditions.

Poultry production still relies on periodic manual observation. Nesteye turns the visual signals already present in a barn into timely, practical decisions.

Piling detection

Identify dangerous flock density as it forms, before compression becomes a mortality event.

Continuous weight

Estimate flock weight from overhead imagery without interrupting normal barn activity.

Earlier health signals

Read posture, movement, behavior, and thermal patterns to surface disease and mortality risk.

Adaptive intervention

Use LED and sound cues to disperse piling and help restore healthy flock movement.

5: Active commercial pilot barns in Texas and Washington
<30 ms: Target inference latency per frame
Edge first: Private operation without a cloud dependency

Development is informed by advisors from UGA and Texas A&M. Training data combines public research papers and datasets with Nesteye's own privately collected barn data. Synthetic data expands coverage across camera height, angle, lighting, flock density, and behavior.

Technical approach

From camera frame to barn-level action.

The perception-and-response loop runs locally, keeping the system resilient and flock footage on site.

Camera feeds → vision models → risk signal → LED and sound intervention

Model layer

Perception: Custom CNN and YOLO instance segmentation
Weight: Body-area and keypoint features with EfficientNet
Health: ConvNeXt behavior, posture, and thermal classification

NVIDIA platform

Simulation: NVIDIA Omniverse and Isaac Sim
Training: NVIDIA TAO Toolkit
Deployment: NVIDIA DeepStream and TensorRT on Jetson Orin Nano

Innovation Lab support would accelerate simulation fidelity, model optimization, and the path from active pilots to a scalable commercial deployment.

Fourteen-week plan

A direct path to field validation.

Weeks 1–5

Setup and simulation

Configure the environment, establish evaluation baselines, and generate varied synthetic barn scenes.

Weeks 5–9

Model fine-tuning

Train and calibrate segmentation, weight, behavior, posture, and thermal models.

Weeks 9–11

Edge optimization

Build the DeepStream pipeline and optimize inference with TensorRT for Jetson.

Weeks 11–14

Field validation and delivery

Measure performance in active barns, document results, and package the deployment.

Deliverables

A deployment-ready package.

Continuous flock-weight model
Piling detection and intervention system
Disease and mortality risk module
Synthetic poultry-barn dataset
Field evaluation report
Jetson deployment package

What follows

The same edge platform can extend into longitudinal flock health scoring, additional species, automated welfare reporting, and privacy-preserving learning across multiple barns.

Download the project brief