Accuracy
Every sighting passes through several models. Independent readers agree on the plate, a fingerprint confirms the vehicle, and each result shows its evidence.
The models
Several models work in turn on every image. Each has one job, and each result is checked by the next step before it reaches your record.
Illustrative: the plate and confidence shown are an example, not a real reading.
A detection model finds each vehicle and its plate in the image. The readers then work on the plate itself, not on the whole picture.
Independent readers read the same plate. They are different models, so they tend to make different mistakes. Where they agree, the reading stands.
A UK plate has a fixed shape, so a position that must be a letter cannot be a number. Neuralis uses this to settle look-alikes such as 8 and B, or O and 0.
A vision model works out make, model, colour and small details. These feed the vehicle fingerprint and are compared with outside records where they are available.
Verification
A single reading can swap an 8 for a B or an O for a 0. Neuralis compares independent readings character by character and uses the shape of a UK plate to settle look-alikes.
When the readers disagree and the shape cannot decide, the record is flagged for review. The vehicle's fingerprint is a further check.
Two of three agree on B. 8 and B look alike, and this position on a UK plate is always a letter, so B wins.
Vehicle fingerprint
A plate can be misread, covered, dirty or copied. Neuralis also builds a fingerprint of the vehicle itself from every image, and checks each new sighting against the ones before it.
Every sighting already has an image. Neuralis finds the vehicle in it and studies the whole vehicle, not just the plate.
The vehicle is described by its make, model, colour and the small details that set it apart from others of the same kind.
The fingerprint is compared with earlier sightings and gets a match strength, with the matching details shown.
A plate that arrives with a different fingerprint from its own history is flagged, with both sightings side by side.
The same fingerprint under two plates is linked and sent for review.
When a plate is covered, dirty or out of frame, the fingerprint can still point to the vehicles it is most likely to be.
Every vehicle carries a description built up over time, so a search shows what it looks like as well as where it has been.
The result is a strength, not a yes or no. Each answer shows which details matched and which did not, so you can see why it was flagged, or why it was not.
Strong match
Partial match
You decide what gets flagged. The level that raises a flag is yours to set, and a person makes the final call. Fingerprints describe vehicles only, and follow the same access and retention settings as the rest of your record. In development, and shown here with illustrative examples.
Intelligence you can trust
Neuralis never asks you to take a result on trust. Whatever it tells you, it also shows the sightings behind it.
A link between two vehicles shows how often they arrived together. A flag shows the readings that raised it. A pattern shows the visits that make it up.
The make and colour seen in the image are compared with the vehicle's record where one is available. A mismatch is flagged for review.
Readings and matches come with a confidence level. When the data is thin, Neuralis says so instead of giving a confident answer.
Neuralis raises the question and shows the evidence. Your team makes the call, and you set the level at which something is flagged.
Early access
We are working with a small number of early customers, starting with sites that already run Hikvision plate-reading cameras.
How it starts