What it is.
Semaphor runs an object detector — a YOLOX-based model executing locally on the monitoring machine — on the original crop of every camera pane it separates from the screen. It looks for the two things guards care about most: people and vehicles. Detections are grouped into numbered, clip-local tracks, so the guard sees “person 01 walked from the gate to the door” rather than a flicker of boxes. It is the React stage of the Semaphor loop, and it decides which moments move forward to an alert.
Why it matters.
Motion is not a person.
Classic motion detection alerts on anything that changes: headlights sweeping a lot, rain on a lens, a flag in the wind, the sun coming out from behind a cloud. Guards learn to clear those alerts on reflex. Object detection asks a better question — is there a person or a vehicle in this scene? — and that is the question post orders are actually written around. We compare the two approaches in AI video analytics vs motion detection.
The quiet threats are the still ones.
A person who walks in and then waits — sitting on a stairwell landing, standing in a dark corner of a garage, crouched beside a vehicle — produces almost no motion. A motion-only system forgets them within seconds. Semaphor keeps them supported.
Works on the tiles you actually have.
The detector runs on panes cropped from a real video wall: small tiles, compressed streams, older cameras. It sees what your guard sees, and it sees it on every tile at once.
How it works.
Detect.
Each camera pane is sampled on a schedule and the detector runs on the original crop, including its edges, finding people and supported vehicle classes.
Track.
Successive detections are associated by motion into numbered, clip-local tracks. IDs are per camera view — tracking, not identity recognition.
Recover.
When a detection drops for a moment, pixel-continuity recovery and regional rechecks keep the track alive instead of starting over.
Clip-local track IDs.
Each person or vehicle gets a short numbered label within its camera view — 01, 02, V01. Those labels let guards and the AI talk about the same object across frames: “person 02 remained near the entrance.” They are deliberately local and anonymous. Semaphor does not use facial recognition and does not try to establish who anyone is. See security and privacy.
Regional rechecks for still people.
Detectors can lose confidence on a person who is seated, partly hidden or not moving. Semaphor runs targeted rechecks on the region where a tracked person was last seen, so a person sitting in a waiting room or leaning against a stairwell wall stays supported instead of silently disappearing from the track list.
Pixel-continuity recovery.
Short detection gaps happen: a person passes behind a pillar, a car moves through shadow. Semaphor bridges brief gaps using optical flow and color continuity, so one person crossing a lot remains one track rather than three.
From detection to event.
A person or vehicle detected with sufficient confidence proposes a candidate event — even on the first sample, without any movement. Substantial pixel changes can propose events too. Those candidates go to event routing and duplicate suppression, then to an alert-first card in the guard inbox, then to an AI scene report. The primer how object detection works on security cameras covers the underlying technology.
Where it matters.
- Parking lots — vehicles arriving after hours, people walking between parked cars, someone lingering by a driver’s door.
- Stairwells and corridors — a person sitting on a landing or waiting in a corridor long after the building has emptied.
- Building lobbies — seated visitors in a waiting area and after-hours entries through the front door.
The recorded demo on the home page shows detection overlays and a track inspector on real research footage from a stairwell, a parking area and a waiting room.
Frequently asked questions.
Does person detection run in the cloud?
No. The detector runs locally on the monitoring machine. Only the optional AI scene review can use hosted inference, and that travels over verified HTTPS with authentication.
Is this facial recognition?
No. Semaphor detects that a person or vehicle is present and follows it with an anonymous, per-camera track number. It does not recognize faces or identify individuals.
What happens when someone stops moving?
Regional rechecks look again at the area where the person was last detected, so a seated or still person keeps their track instead of vanishing the moment they stop.
See detection running on your own video wall: book a live demo or start a pilot.