Face and plate recognition

Known objects are matched by appearance: clothing, build, a vehicle's shape and color. Face and license plate recognition add stronger evidence. Where a face or plate can be read, it decides who the visit was; elsewhere, appearance still answers. That makes rules like these dependable:

• Jenny in the garage: a person known by face as Jenny, on the Garage camera
• Unknown license plate: a vehicle whose plate is on no known vehicle

Both are off by default and are turned on camera by camera. They run on the recorder itself, after each visit is recorded, on the same picture the event shows. That is the still cut from the full-resolution recording, when it is sharper than the detection stream. Nothing is sent off the recorder, and live detection is not slowed.

Face and plate recognition is a licensed feature. Where the recorder's license does not include it, turning faces or plates on, adding a known person or vehicle, enrolling a face or plate, and tagging an event as someone are refused. What is already on and enrolled keeps matching.

Install the models

The face and plate models are downloaded once, on the recorder:

npm run setup:recognition

Each file is checked against a pinned SHA-256 before it is used. Until the models are installed, Settings → Recognition says so, and settings can still be saved.

Turn it on

An administrator opens Settings → Recognition.

Faces. Check Recognize faces of enrolled people, read the notice, and confirm that the site may use face recognition. Then choose the cameras that see faces up close, such as a doorway, gate, or garage.

License plates. Check Read license plates and choose the cameras that see plates nearly head-on, close enough to read. Add your own vehicles' plates on Known objects so they are never reported as unknown.

Saving re-reads the last day of people and vehicles on the chosen cameras.

warning

Face recognition processes biometric data. Many places require posted notice, written consent from each person enrolled, or a published retention policy, and some forbid it in certain settings. Check what applies to the site before turning it on. EverNVR records who turned it on and when.

Enroll people and vehicles

On Known objects:

• A person: under Face, use Add a photo with a clear photo of that person alone, facing the camera. The photo is not kept; only the face feature taken from it is. Tagging an event as that person on a camera with face recognition also enrolls the face in that event's picture. An enrolled face must be at least 48 pixels across and face the camera; otherwise the recorder says why and enrolls nothing. Several enrollments from different days and angles work best.
• A vehicle: under License plates, add the plate. Tagging a vehicle event also adds the plate it read, when the read was confident. A plate belongs to one vehicle.

Remove a face or plate there at any time. Forgetting an identity deletes its faces and plates.

What is kept

	Kept	Not kept
Faces	A face feature (128 numbers) for each enrollment of an enrolled person. For each event, the outcome: known (and who), unsure, unknown, or unusable.	Features of anyone not enrolled. Uploaded photos.
Plates	The plate text and read confidence with each event, for as long as the event is kept. Your known vehicles' plates.	Nothing else.

Turning face recognition off deletes every enrolled face feature. To turn it back on, people have to be enrolled again. Backups made while faces were enrolled contain them until those backups age out.

How a visit is decided

Faces. A face smaller than about 40 pixels across is ignored. A face turned too far from the camera, such as one seen from the side, is reported as turned away or too small and decides nothing. When a picture holds several faces, the one at the middle of the visit's picture is used. Faces at the edges belong to other people's visits.

Face reading	Meaning
Known	Matches an enrolled person. The visit is known by face.
Unsure	Close to an enrolled person, but not close enough to name them.
Unknown	A clear face that is not anyone enrolled.
Unusable	Turned too far from the camera to compare.

Plates. A read is used when the plate reader is reasonably sure of every character. A read that matches a known plate, or is one character off from it, names the vehicle, because readers often confuse characters such as 8 and B. A plate is reported unknown only when it is read with high confidence and differs from every known plate by at least three characters. Anything in between is an unsure read, so a misread of your own car does not raise an unknown-plate alert.

The thresholds are starting values. They have been tried on one site's cameras, not tuned across many. Confirm alerts from the footage. To correct a wrong name, use the Not button on the event (for example Not Jenny). That also removes any face enrolled from that event.

Rules

In an object or departure rule's Who, under Face and plate recognition:

Choice	Fires when
Someone known, by face	The visit was named by an enrolled face. Choose the person under In particular.
A face nobody enrolled	A clear face matched nobody enrolled.
A vehicle known by its plate	The visit was named by a known plate.
An unknown license plate	A confident plate read matched no known vehicle.

These are decided a moment after the arrival, once the picture has been read, and each fires at most once per visit. Notifications name who was recognized and how, or the plate that was read. MQTT event messages carry the same face and plate fields.

Search by plate

On Events, type all or part of a plate into the Plate filter. Selecting a plate in an event's panel lists every visit with that plate.