Who Came Through the Gate? Vehicle Recognition for Secure Areas

Two Stages, Two Places

Fast trained models run where the camera is. A far more capable vision-language model runs in the cloud, and only sees the few frames worth sending. Splitting the work this way keeps most of the video on the premises, keeps working when the connection drops, and keeps the cost of the cloud model in proportion.

Stage one, on the premises

Powoflow Sense Mesa, or the Powoflow Sense AI computer for heavier camera loads, runs trained detection models on each camera's video. They recognise that a frame contains a vehicle, follow it from frame to frame, and keep the best view of it. An empty yard, a tree in the wind or the same parked car hour after hour goes no further.

A frame only moves on when it passes a few simple gates: recognition is turned on for that camera, the frame contains a vehicle, the vehicle is new rather than the one already seen, and the camera is outside its cooldown. Daily limits cap how much is sent.

Stage two, in the cloud

The vehicle is cropped from the full-resolution frame. Dedicated plate models find the plate and read it. A vision-language model then looks at the vehicle crop and returns a structured description: make, model, colour and type. It is instructed to describe the vehicle only, never people, faces or anything that identifies a person.

One Vehicle, Many Sightings

Each vehicle becomes a record of its own, so the second visit adds to the first instead of starting again. The record keeps a few reference images (the first capture, the view from each camera, the latest and the best), the plate with its read confidence, and every sighting with its camera and time.

Make, model and colour show how many sightings agree. One model read on one frame can be wrong; three that agree are worth more, and the record says which you have.

Disagreements Are Shown, Not Hidden

Plates are read from real camera angles in real light, and a character can come out wrong. When two reads of the same vehicle disagree, Powoflow keeps the more confident one as the plate and shows the other as a disputed read, so a person decides rather than the software quietly picking one.

Use It Where You Have the Right To

A number plate identifies a vehicle, and through it, usually a person. In many countries, including across Europe under GDPR, that makes plate reads personal data. Before you turn recognition on:

• Point cameras at your own ground — your gates, yards, depots and restricted areas, not public roads or neighbours.
• Tell people — signs at the entrance that cameras read plates, and why.
• Write down the purpose — for example, knowing which vehicles entered a secure area. Run a data protection impact assessment where your law requires one.
• Decide how long you keep it — and delete what you no longer need.
• Limit who can see it — sightings are only visible inside your own account; give access to the people who need it.

Used that way, vehicle recognition answers a narrow and useful question: who came through the gate, and when.