Recognize an enrolled person and a registered vehicle

Everything CV-Scope does so far is anonymous. This guide switches on the two licensed modules, face recognition of enrolled identities and vehicle plate recognition, and shows how a rule can be limited to one enrolled person or one registered plate. It uses a webcam or IP camera for the person, because the faces in the sample clips are about 20 px tall, too small to identify.

Read docs/recognition.md first if you administer the installation: it explains the licence model, the roles and what is stored.

Step 1. Unlock the modules

1. Open Recognition in the sidebar. Without a licence it shows one page with both modules as Not licensed.
2. Ask the vendor for a licence. They will want the licensee name and, for a licence bound to this computer, the id shown under This machine’s id.
3. Click Create the first administrator token. The token is shown once and stored in this browser; keep a copy. Everything on the Recognition pages needs such a token.
4. If the vendor’s public key is not already configured, paste it under Trust an issuer public key and click Trust key.
5. Paste the licence JSON under Install a licence file and click Validate and install. Both modules now read Licensed and the sidebar shows People, Vehicles, Recognition events and Settings.

Step 2. Install the models

1. Open Models and scroll to Recognition models.
2. Install YuNet face detector and SFace face embeddings (about 37 MB together, MIT and Apache-2.0 licences, CPU).
3. Install YOLOv9-t licence plate detector (384 px) and CCT-S v2 global plate OCR (about 13 MB, MIT). They need ONNX Runtime; the setup script installs it.
4. Open Hardware. The panel Recognition modules: effect on camera capacity estimates how many streams this computer can still process with recognition on. On the development machine the estimate dropped from about 10 streams to 4 at 10 processed frames per second.

Step 3. Enroll a person with the guided capture

The guided capture works like enrolling a face on a phone: the person sits in front of the camera, the screen asks for one thing at a time, and each view is kept by itself. Nobody has to press a button per picture.

1. Open Recognition → People, type Research Subject 001 as the display name and E-001 as the reference id, then click Create profile.
2. The profile opens on the Guided (live camera) tab. Choose your camera and click Start guided capture. The camera picture appears in a circle with a ring around it, one segment per view.
3. Follow the instruction in large type above the circle. It changes as the person moves: Look straight at the camera, Move a little closer, More light on the face, please, Move into the circle, Only one person in front of the camera, please. When everything fits, it says Hold still… and the bar fills; a moment later the view is captured, its ring segment turns green and the next instruction appears.
4. The plan is five views: front, turn to your left, turn to your right, chin down (what a camera above the person sees) and chin up. An arrow inside the ring points the way. Afterwards you can add Lighting (change the light in the room, the capture waits until it really differs) and Back of head, which is kept as an appearance reference only and never produces a template.
5. Click Finish and check quality. The result shows the status, the quality, the views covered and how consistent the pictures are with each other. Five views of the development machine’s test pattern gave quality 0.87 and consistency 0.64. An enrollment with a missing front view, too few views or inconsistent pictures stays insufficient and is never matched.

Notes on the guided capture:

• The left, right, chin-down and chin-up views are judged against this person’s own front view, so a camera mounted off to the side or a model that reads a different neutral pose does not make every picture wrong.
• If a different face appears halfway through, the capture is refused with This does not look like the same person as the front view.
• The camera is opened only while the capture runs and is shared with the live view; a running experiment keeps its camera, and the guided capture then uses that run’s picture.
• Single pictures is the second tab: pick a view, capture one frame or upload a file, and read the same guidance. Use it for network cameras where nobody watches the screen, or to enroll from existing photographs.

Step 4. Add another look

One capture session records the person as they are today. Recognition gets steadier when the profile also holds the appearances they really turn up in.

1. Open the profile again. Above the capture tabs, the row of Look buttons shows First enrollment. Click + Add another look, type a name such as Glasses and click Start this look.
2. Change the appearance for real: put the glasses on, the hard hat on, or move to the evening light. Run the guided capture again; the ring starts empty, because each look keeps its own views.
3. Click Finish and check quality. The summary now lists both looks with their own agreement and how well the new look matches the first enrollment. A look that does not match is reported as a problem, which is what stops someone else from being added by mistake.
4. Every look is used when matching, so the person is recognized either way. A look can be deleted on its own; Re-enroll still clears the whole profile and starts over.

Step 5. Check that recognition actually works

Open Recognition → Test recognition. Nothing on this page is stored.

1. Live camera: pick the camera and click Start live test. Stand in front of it. Each face gets a box and a label: the name and the similarity when recognized, ? name for a possible match, or Unknown or Insufficient quality with the reason. The label follows the person about four times a second.
2. Walk to where the camera really watches from. If the label turns into Insufficient quality, the face is too small, too dark or too blurred there: move the camera closer or raise the resolution rather than lowering the thresholds.
3. A picture: upload a photograph of the person, or take one frame from any camera, and read the ranked candidates with their similarity. On the development machine, the photograph a profile was enrolled from came back as Recognized at 1.00 against a Recognized threshold of 0.46.
4. Enrollment check: every profile is compared with its own pictures and with everyone else. Read the verdict and the advice column: Only one look points back to Step 4, Only 2 of 3 views to the missing views, and a Mix-up risk warns that two enrolled people are close enough to be confused, which is the moment to add views and looks for both.

Teaching it while you test

While the live test runs it asks about the person on screen, at most once every 20 seconds per person.

1. Is this Research Subject 001? Click Yes, correct. The verdict is recorded and, because Keep the picture when I confirm a match is ticked, that frame becomes an enrollment picture in the look Live confirmations. Confirm a few times in the light and the position of the real camera and the profile learns the site, not just the enrollment desk.
2. Click No when it is wrong. Say who it really is, and that person gets the picture instead; or choose Not enrolled — register. The verdict is kept either way, and the counts above the picture show how often the installation is right here.
3. A face nobody matches is not identified. The bench asks Not recognized. Register this person?; with their agreement, click Register and teach the views, type the name, and the guided capture runs in place on the same camera, asking for front, left, right, chin down and chin up. Close it and the test recognizes the new profile at once.
4. A picture that does not look like the profile is refused: This face matches … only at 0.21. Add it anyway only when you are certain; that override is written to the audit trail. None of this moves a threshold by itself, and it is not model training: it adds enrollment pictures, which is what makes the difference in practice.

Step 6. Register a vehicle

1. Open Recognition → Vehicles and register plate ABC12345 with the description Site car and the group Staff.
2. In Test the plate format layer, type DXB 12S67 and click Normalise. With the uae format enabled in the settings it becomes DXB12567 with one confusable substitution; with only generic it stays DXB12S67. Choose the formats of your site under Settings → Plate recognition and keep generic last.

Step 7. Rules that fire only for them

1. Create an experiment for the camera that sees the person, with Person, and draw a zone Zone A covering the area where they will stand.
2. Add a rule: WHEN Person, RECOGNITION Specific enrolled person → tick Research Subject 001, ENTERS Zone A, mode Remains in zone for… more than 30 seconds, CREATE EVENT Extended Zone A presence.
3. Add a second rule with RECOGNITION Anonymous person and COUNT AS Unknown visitor to see the difference.
4. For vehicles, create an experiment on a camera with Car and Truck, draw a gate Entrance Gate, and add: WHEN Car, PLATE Specific plate → ABC12345, CROSSES Entrance Gate, RECORD AS Vehicle Entry.
5. Click Save and Start. A rule with a recognition condition needs the module to be licensed and enabled; otherwise it stays inactive and the experiment’s Resolved configuration says so.

Step 8. Watch the recognition happen

1. While the run is active, open Recognition → Recognition events. The Diagnostics of active runs panel lists every track: whether a face was detected, its quality and why a frame was rejected, how many observations were usable, the best frame, the candidate identity with its similarity and runner-up, and the result (Not decided, Recognized, Possible match, Unknown, Insufficient quality).
2. Open Live. Because this browser holds a recognition token, the video label reads #12 person = Research Subject 001 and the panel Recognized now lists each track with its name and confidence. In the Scene Builder the live overlay shows the same name. A browser without a token sees the box and the anonymous id only, on the same stream.
3. When the person has stayed 30 seconds, the event Extended Zone A presence appears under Data with entity: enrolled_person in its context. The stored event holds the id only; the Identity column resolves it to the name for this browser, and the CSV, JSON and Parquet exports stay anonymous. The full recognition detail (similarity, quality, model version, camera, track, frame) is on the Recognition events page.
4. If the experiment records video, open the run’s Analysis → Video tab (set From entry to exit on the experiment to get one clip per visit): the clips are listed with the event that started each one and play in place, and the events inside the clip jump the player to the moment. The Watch link on every event row in Data opens the same clip on the Video review page. Recorded video never has names drawn into it, whoever watches it.
5. For the vehicle, the diagnostics show the raw OCR of each read, the temporal consensus (ABC12?45 becomes ABC12345 once enough reads agree), the normalized plate and the registered vehicle; the crossing is recorded as Vehicle Entry.

Step 9. Clean up like an administrator

1. Recognition events: filter, export as CSV (administrators only; the export is audited) or delete single events.
2. Settings → Retention: recognition events and unregistered plate reads are removed after the configured days.
3. People: open the profile and click Delete profile. The templates are removed at once and the encrypted enrollment images with them (or after the configured grace period).
4. Settings → Audit trail lists who did what: profile created, enrollment finalised, licence installed, export created, profile deleted.
5. Settings → What this installation stores (the ordinary Settings page) now lists the modules and their state, so the privacy list stays honest.

Notes

• Faces smaller than the minimum size (40 px by default) and plates below 14 px are never identified; the diagnostics say Insufficient quality. Move the camera closer or raise the resolution rather than lowering the thresholds far.
• The Recognized threshold, the margin to the runner-up and the number of observations before a decision are under Settings → Advanced settings. Calibrate them on your own site with a few known people.
• Unknown people always stay anonymous tracks. The module cannot search for people who were not enrolled and never infers age, gender or any other attribute.
• When recognition looks wrong on site, go to Test recognition before changing thresholds: it says whether the picture, the enrollment or the threshold is the problem.