Yuhox Edge.
An on-site box that watches video and raises alerts automatically. Install it and try it right away; paste a license code offline to expand.
How to start
See the price and order
After payment, the license appears under Console → My licenses. Nothing is issued until the money arrives.
Collect the license in the software
In System settings → License, click “Get license with Yuhox account”. You do not copy a license code.
Machines without internet
Use Console → Offline activation / transfer. The license code is a secret; it is masked by default.
What it solves
The hard part of video analytics was never whether the model runs — it's that the field isn't the lab: lighting changes, angles are skewed, cameras hide behind someone else's platform, and the machine room allows no internet. Yuhox Edge solves these together: inference on the box, algorithms swapped as plugins, a gateway that can reach streams behind other platforms' walls, false alarms fed back for retraining, and one firmware for public cloud, on-premises and fully offline.
Before and now
| Before | Now |
|---|---|
| Once the model is installed it is locked. False alarms on site just have to be lived with. | False-alarm feedback, labeling, fine-tuning, and an evaluation gate. After changes pass, they go to the box over OTA. |
| Cameras are registered on a third-party platform; the box cannot reach them directly | The access gateway either logs into the platform with an account and pulls streams automatically (Dahua ICC / AnnEye are done), or takes hand-entered addresses per stream, normalizes everything to stable RTSP and hands it to the engine in batches — not a single on-site device needs touching |
| On-premises means a separate build. Offline means yet another build. | One box firmware. Three deployment models differ only in configuration. |
| Cutting false alarms by swapping models over and over just makes things messier | Zone, time window, duration, consecutive frames, and confidence threshold are all configurable. Use rules first to bring false alarms down |
| Alerts only, no handling — reports go nowhere | Grading, confirmation, false-alarm marking, closing — every step records who clicked it, and alerts can push to the customer's existing platform |
Core capabilities
- 01
All public algorithms are visible
Intrusion, line crossing, absence, smoke and fire, hard hats, illegal parking and the rest of the public algorithms are all in the catalog, and the list can be checked right there in the admin console. A trial applies only 1 at a time; buying opens more according to your allowance. At expiry only the overage is paused, the configuration stays, and renewing brings it back.
- 02
Offline licensing
One system, one certificate: paste it into the admin console and you are done. It works with no internet on site, and we do not have to do any operations alongside you.
- 03
Three deployments
Public cloud / on-premises / fully offline ferry, same box firmware.
- 04
Ingest gateway
Cameras hidden behind an NVR or a third-party platform are converted by the gateway into stable RTSP and ingested in bulk. Streams that come in through the gateway count against the camera allowance and are not billed again.
Who it is for
Plant, construction site and campus safety management
Hard hats, area intrusion and the like need someone watching, but people cannot watch everything
Sites where data must not leave the premises
On-premises or fully offline; video does not go to the cloud
Sites that already have many cameras
Do not want to replace hardware; just want analysis on existing video
Why this one
Sites change
The data loop is the real difference between Yuhox Edge and a "buy a model, install it, done" approach. The model has to follow the site, or after three months nobody looks at it.
Rules first, then the model
Most false alarms are suppressed with rules: duration, consecutive frames, time window, region. Jumping straight to swapping the model is the most expensive and slowest way.
Offline is a first-class citizen
USB ferry for models and alerts; fully usable in a pure offline environment. Not an online edition with a few features cut.
Someone else can write the algorithm
Plugins follow a fixed C ABI contract; headers, examples and scaffolding ship with the SDK, so third parties can integrate their own algorithms without waiting on our schedule.
Same job, two ways
The left column is how most sites work today — no vendor names. The right column is what the current version actually does.
| Topic | Common practice | Yuhox Edge |
|---|---|---|
| The model does not match the site | Installed once and frozen: live with false alarms, or swap models forever | Rules first to cut false alarms, then fine-tune on returned data and push back to the box |
| Cameras sit behind someone else's platform | Replace cameras, or wait for the platform to open up | The ingest gateway logs into that platform and turns it into stable RTSP. Site cameras stay |
| The machine room cannot reach the internet | A separate offline SKU with half the features cut | The same firmware: public cloud, private, or USB sneakernet |
| Alerts with nobody on them | Only a popup | Grading, confirmation, false-alarm marking and closing all leave a trail, and can push back to the customer's existing platform |
Key parameters
| Box platforms | NVIDIA Jetson / Rockchip RK3588 / Windows x64 |
|---|---|
| Engine | C++17 + GStreamer |
| Plugin interface | C ABI (headers and examples ship with the SDK) |
| Deployment modes | Public cloud / private EdgeHub / fully offline ferry |
| Ingest gateway | Go single file, zero dependencies |
| Open APIs | Alert push API |
Up front
What is the recognition accuracy?
We will not give a number detached from your site. The same model varies a lot under different lighting, angles and occlusion. The real approach is to run it on your site for a week, fine-tune on the returned data, then talk accuracy.
How many streams can one box handle?
It depends on box model, resolution, frame rate, and algorithm complexity. Jetson and RK3588 differ a lot. We have to price it from your camera points; there is no one-shot quote.
Can it connect to our current video platform?
It depends on whether your platform can yield a video stream. For Dahua ICC / AnnEye we built account login, automatic device listing and stream pulling; for Hikvision, Uniview, ONVIF devices or national-standard gateway streams, anything that produces an RTSP URL connects the same way. Alerts then push back to your platform through the open API.
How do you upgrade models fully offline?
USB ferry package with models and config; takes effect on import. Alarm data can be carried out the same way for return.
One account, multiple systems
If you use several Yuhox systems, one account covers them all. No separate accounts or passwords per system.
Start with a site visit
Tell us the number of sites, camera models and whether there is internet access. We will judge how many boxes you need and whether a gateway is required.