From Ultralytics YOLO Export to Edge Video Analytics with CosmoEdge

[samuel--hu]

Hi Ultralytics community,

We have been working on CosmoEdge, an open-source C++ edge AI engine for production video analytics. I wanted to share a small end-to-end workflow that starts from an Ultralytics YOLO model and turns it into a visual edge video analytics pipeline.

Demo

yolov8sdet.clspipeline.mp4

The demo walks through the path we use in practice:

1. Start from an Ultralytics YOLO model and export ONNX:

yolo export model=yolov8n.pt format=onnx

2. Import the ONNX model and labels into the CosmoEdge Model Repository.
3. Compose the pipeline visually: video decode -> YOLO detection -> tracking / classification -> zone and rule logic -> OSD and event records.
4. Run live OSD on edge hardware, with event integration available through MQTT / HTTP.

Workflow

CosmoEdge is not trying to replace model training or export. The idea is to make the deployment side more visual and production-oriented after a YOLO model is ready: model management, video source binding, pipeline orchestration, OSD, event records, and integration with downstream systems.

Benchmark snapshot

For the public ScenarioBench v1.0 data, the mixed NPU scenario combines pedestrian detection and no-safety-helmet detection on the same video channels:

Scenario	Stable channels	Concurrent scenario tasks	Target FPS	Result
No Safety Helmet	16	16	3	PASS
Pedestrian Detection	16	16	5	PASS
Pedestrian + No Safety Helmet	16	32	3	PASS
x86 CPU baseline	7	7	3	LIMITED

The benchmark reports this as verified stable capacity for the tested setup, not a claimed hardware upper bound.

Links:

• Repo: https://github.com/cosmo-wander-ai/cosmo-edge
• ScenarioBench v1.0 summary: https://github.com/cosmo-wander-ai/cosmo-edge/tree/main/docs/benchmarks/scenario-bench/v1.0
• Mixed NPU report: https://github.com/cosmo-wander-ai/cosmo-edge/blob/main/docs/benchmarks/scenario-bench/v1.0/pedestrian-helmet-mixed-npu/report.html

I would love feedback from people deploying YOLO on edge devices: which parts of the export-to-application path are still the most painful for your workflows?

[UltralyticsAssistant]

👋 Hello @samuel--hu, thank you for sharing your edge video analytics workflow with the Ultralytics community 🚀! This is an automated response to help point you to useful resources, and an Ultralytics engineer will also assist soon.

We recommend a visit to the Docs for new users and integrators, where you can find many Python and CLI usage examples and where many of the most common export, deployment, and inference questions may already be answered.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug it.

If this is a custom training ❓ Question, please provide as much information as possible, including dataset image examples and training logs, and verify you are following our Tips for Best Training Results.

Join the Ultralytics community where it suits you best. For real-time chat, head to Discord 🎧. Prefer in-depth discussions? Check out Discourse. Or dive into threads on our Subreddit to share knowledge with the community.

Upgrade

Upgrade to the latest ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8 to verify your workflow or issue is not already resolved in the latest version:

pip install -U ultralytics

Environments

YOLO may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLO Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

[glenn-jocher]

Thanks for sharing this, @samuel--hu. The export-to-application gap is definitely a common pain point for edge deployments, especially around runtime compatibility, postprocessing consistency, and multi-stream pipeline orchestration after yolo export. CosmoEdge looks aligned with the kind of ONNX-based deployment path covered in our Export mode docs and ONNX integration docs; if you run into any Ultralytics-specific export or inference mismatch, please open a focused issue with the model, export args, runtime details, and a minimal repro so we can take a closer look.

[samuel--hu]

Thanks Glenn, really appreciate the detailed feedback.

That export-to-application gap is exactly the area we have been working on with CosmoEdge. We are polishing a focused Ultralytics YOLO edge deployment guide that covers the path from YOLO export / ONNX import to pipeline composition, postprocessing / OSD, and multi-stream edge runtime validation.

Once the guide is stable, we’ll share the link here as a reproducible reference. And if we run into any Ultralytics-specific export or inference mismatch, we’ll open a focused issue with the model, export args, runtime details, and minimal repro as you suggested.

[glenn-jocher]

Sounds good, @samuel--hu — a reproducible guide with the exact ultralytics version, model, export args, ONNX/runtime versions, target hardware, and validation inputs/outputs would be especially useful for others comparing edge pipelines. Please feel free to share it here when ready, and open a focused issue for any Ultralytics export/runtime mismatch.

[samuel--hu]

Thanks Glenn, we’ve published the focused guide here:

https://github.com/cosmo-wander-ai/cosmo-edge/blob/main/docs/en/tutorials/06-ultralytics-yolo-edge/ultralytics-yolo-edge.md

It includes the exact ultralytics, ONNX, and ONNX Runtime versions, the YOLOv8n / YOLOv8n-cls export and import steps, target hardware/runtime details, validation input/output references, the pipeline package, OSD/event screenshots, and multi-stream ScenarioBench benchmark references.

We’ll keep the guide updated, and if we run into any Ultralytics-specific export/runtime mismatch, we’ll open a focused issue with the model, export args, runtime details, and a minimal repro as you suggested.

[glenn-jocher]

Thanks for sharing the guide, @samuel--hu. The added version pinning, validation references, runtime details, and benchmark context make it much easier for others to reproduce and compare the ONNX deployment path. Keeping the Ultralytics-specific boundary clear is appreciated too: if any export/output mismatch appears, a focused issue with the exact model, yolo export args, ONNX Runtime version, hardware, and minimal input/output comparison will be the best path for us to investigate.