The waiting room is the least observed space in the hospital.

Machine perception on the ceiling cameras a hospital already owns. The video stays inside the building.

The camera they already ownA ceiling dome, 149 mm across and 102 mm deep. It is already installed and already pointed at the room. Today it records to a box nobody watches.
Its wide lensThe wide view that lets one camera cover a whole waiting room is the same wide view that bends every straight line in it. Undoing that bend is the first thing the software does.
The image sensorA rectangle a little over 5 mm across. Everything the system will ever know about the room arrives through it, heavily compressed, at a low frame rate.
One cable, one wayThe video goes down the wire the camera is already on. That cable was pulled years ago and it already carries the camera’s power. Nothing new goes into the room.
Where the inference runsOne box in the hospital’s own rack. It finds people in the frame and reads their posture, then follows them from one frame to the next. The security team can put a hand on the hardware.
A second stage has to agreeTwo stages read different evidence. Nothing reaches a person until both of them say the same thing, and most candidates never do.
One alert leavesAn event and a timestamp cross the network boundary. The frame stays behind, and so does the clip it came from.

The path of one alert

Today it reports two things a camera can honestly see: a person who is down and has not gotten up, and a patient who left the room without being discharged. Every alert arrives with the frame and the timestamp attached.