SynOne Vision — VLM Video Recognition

Conventional video analytics draws a box around “a person”. SynOne Vision uses a vision-language model that understands a scene and answers in words, so it judges situations: how many people are in the lift, whether there is a trolley, whether it is loaded, whether it belongs in this lift — or whether the guard has been away from the desk too long. It uses the cameras you already have, and the model runs on a GPU on site, with no per-call cloud API.

• Buildings that separate passenger and goods lifts
  • Office
  • Retail
  • Industrial

  Platform and hand trolleys or bulky cargo in a passenger lift raise an alert, while cleaning carts and similar can be allowed; a goods lift carrying people without a trolley can be flagged too.

• Committees and property firms that need eyes on the guard desk

  A guard away from the desk beyond a set time, or apparently asleep during night hours beyond a set time, each raises an alert with a snapshot; looking down at a phone does not count as sleeping.

• Sites with CCTV that do not want to buy AI cameras

  It reads the RTSP streams of your existing IP cameras; add one machine with a GPU, and the rules are tuned to the site.

Situations and alarm rules are set and tuned by us to each site’s needs. Cameras must see the whole lift car or desk. The production configuration is a Windows machine with an 8 GB NVIDIA GPU, with the model using about 5 GB of video memory; check other hardware with us first.

• Situations, not just boxes

  The model reports how many people are in the car, whether there is a trolley and what kind (platform, two-wheel, folding cart, cleaning cart…), whether it is loaded, whether there is bulky cargo, plus a one-line description — ignoring people outside the door, in the lobby or in mirror reflections.

• No change, no question

  Thumbnails are compared first and the model is asked only when something changed — at most every 2 seconds per lift, with periodic re-checks even without change. The GPU idles most of the time.

• Majority vote against one-off errors

  State changes only when most of several consecutive answers agree, and repeat detections in the same lift within 30 minutes merge into one alarm.

• Snapshots and an audit trail

  Alarms pop up in the SynOne alarm centre with a sound, snapshot and AI description, and “acknowledge” is recorded; the change log filters by date or alarms only.

Vision uses the SynOne platform’s camera streams and alarm centre; situations can be added per site.

• Model
  What it covers

  Open-source vision-language model (Qwen3-VL 4B) on an on-site GPU
  Notes

  No cloud API fees; no video sent to cloud AI.
• Lift cars
  What it covers

  People, trolley and type, loaded or not, bulky cargo; alerts for trolleys in passenger lifts and goods lifts used without one
  Notes

  Each camera is set as passenger or goods lift; allowed trolley types are configurable.
• Guard desk
  What it covers

  Away from the desk beyond a set time; apparent sleeping in night hours beyond a set time
  Notes

  Crop area, minutes and night hours are configurable; tell guards before enabling it.
• Output
  What it covers

  AI panel on the lift page (recognition view or live video, red outline), alarm centre, change log with snapshots
  Notes

  Recognition, lift alarms and desk alarms have separate switches; alarms are off by default.
• Production configuration
  What it covers

  Windows machine, 8 GB NVIDIA GPU; the model uses about 5 GB of video memory
  Notes

  Under about a second per frame; with many cameras, frames queue in turn.

Do we need AI cameras?

No. It reads the RTSP streams of existing IP cameras, provided they see the whole lift car or desk. You also need a machine with an NVIDIA GPU to run the model.

Is video uploaded to the cloud?

No video goes to any cloud AI service — the model runs on the site’s own GPU. Alarm snapshots are stored on the site server and visible only to platform accounts with camera rights.

Can it recognise other situations?

A vision-language model is told in words what to judge, so a new situation does not require retraining. Lifts and the guard desk are in production today; tell us what else you need and we will assess it, set it up and test it on site before it goes live.

How accurate is it?

It depends on camera angle, lighting and the situation, so we do not quote a universal percentage. On rollout, recognition runs with alarms off while every situation is tested on site, and alarms are enabled only once the judgements meet the need.

Licensing

Sold under a perpetual licence for one site and one server, bound to that server’s hardware by a digitally signed licence file. Copying it to other machines or sites, reselling or lending it, reverse engineering it or circumventing the licence check is prohibited, and infringements are pursued under the Copyright Act. Full licence terms (Chinese) →