Detection models, faces, plates and semantic search
Updated 2026-10-22
Pick a bigger detector and a faster rate, raise Loitering events, recognize the people you add, read license plates against a watchlist, and search snapshots by describing them.
Everything on this page runs on your own server. Nothing is sent to Anthropic or anywhere else, and no AI key is needed. Each feature is off or on its default until you change it, and each one says why it is not running when a model is missing.
Models
Open System → Detection. The Models table lists every model Sentinel can use, with its size, license and state. Click Download to fetch one; progress shows in the table and in Admin → System Dashboard → Current Tasks, and the file is checked against a pinned SHA-256 before it is used. Cancel stops a download and deletes the partial file. Delete removes a downloaded model; the bundled detector and the detector in use can't be deleted.
To install a model by hand (an offline site, for example), copy the file into the models directory named under the table. It counts once its SHA-256 matches.
| Model | Used for |
|---|---|
| YOLOX-Tiny (bundled) | Person, vehicle and animal detection, as before |
| YOLOX-S | A bigger detector that finds smaller and farther objects; best with OpenVINO or Core ML |
| OSNet | Appearance search ("Find similar") |
| YAMNet | Audio classes: glass break, scream, gunshot, alarm, dog bark, car horn, siren, baby cry |
| YuNet and SFace | Faces: finding a face, then recognizing it |
| PaddleOCR detection and recognition | License plates: finding the text, then reading it |
| CLIP image and text | Semantic search |
Advanced: a detector you trained yourself can go in as custom.onnx with a custom.json beside it naming the family (yolox or ultralytics), the input size and the labels. It then appears in the model picker.
Detector and rate
- Detector model lists the installed detectors. A new model takes effect when Sentinel restarts; until then the card says Restart to apply.
- Detection rate is how many times a second each camera is checked while there is motion. Auto is 1 a second on the CPU and 5 with an accelerator. More passes follow moving objects better and use more CPU. The change applies at once.
- The status line shows the model, engine, device, the rate in use and the time per frame.
Detections of one moving object now share a track. In Search, Same object on an event card lists every detection of that track.
Loitering
In Admin → Devices → Cameras, edit a camera and set Loitering after to a number of seconds (0 turns it off). A person who stays in view that long raises one Loitering event, a Warning alarm, per visit. It needs object detection on the camera.
Faces and people
- Download Faces (find) and Faces (recognize) under System → Detection → Models.
- Open Admin → Security → People and turn on Recognize faces on Person events. Without the models the switch stays off and the page names the missing file.
- Faces that match nobody appear under Unassigned faces. Use New person to name someone from a face, or Add to person to add another face to someone you already have. A few clear faces from different cameras work better than many from one.
- For each person choose what a match does: Name them in the alarm (the Person alarm keeps its severity and says who it is), Don't alarm (for family and staff; the alarm is closed with a note), or Critical watchlist alarm.
What is stored, and privacy. For every face it finds, Sentinel keeps a list of numbers that describes the face (a vector) and a small crop of the face, on your server only. Unassigned faces are deleted after the number of days you set (7 by default). Deleting a person deletes all of their faces; their past events keep the name. Face recognition is off by default. Some places regulate biometric data like this, and some require consent or notice before you use it. Check what applies where your cameras are. Sentinel does not give legal advice.
License plates
- Download Plates (find) and Plates (read).
- Open Admin → Security → Plates and turn on Read plates on Vehicle events. Each Vehicle event's snapshot is read, and the plate shows on the event card.
- Add plates to the watchlist with a name and what a sighting does, as for people. A plate of 6 or more characters also matches when one character is misread. Click a plate to see where it was last seen.
Accuracy depends on the camera. The plate has to be sharp and at least a few dozen pixels wide in the snapshot; a wide overview camera across a parking lot rarely reads plates. In Search, the Plate field finds events by plate; the start of a plate is enough.
Semantic search
Download the two Semantic search models. In Search, type what you're looking for in Describe it, for example a white van in the driveway or a person with an umbrella. Results are snapshots that look most like the description, best first, with a bar for how close each one is; the time range and cameras come from the filter bar. New events are indexed as they arrive. To include older events, click Index recent events above the results (it covers the newest 50,000 events with a snapshot; progress also shows in Current Tasks). When the models are missing the box is greyed out and its tooltip says why.
What each feature needs
| Feature | Needs | Turned on in |
|---|---|---|
| Bigger detector | YOLOX-S; an accelerator for 5 passes a second | System → Detection |
| Loitering | Object detection on the camera | Camera settings |
| Audio classes | YAMNet and a camera with a microphone | Camera settings (audio detection) |
| Faces | YuNet and SFace | Admin → Security → People |
| Plates | Both PaddleOCR models and a camera that sees plates close up | Admin → Security → Plates |
| Semantic search | Both CLIP models | Search (on when the models are present) |
People, Plates and the Detection card are on the web dashboard. The Mac and iPhone apps show the Watchlist alarms these features raise, with the name or plate in the alarm details.
More guides, including Sentinel Cloud setup and users and roles, are in the knowledge base for signed-in users.