I’d want it too, but from what the camera docs say it’s a big change: Hermetry passes the live video through only while someone watches, doesn’t decode or re-encode it, and stores nothing. There’s no frame in the platform to run a model on. So in the meantime it goes at the edge, which is where I’d put it anyway for 9 cameras: no upstream bandwidth for 9 streams all day.
What we run on a logistics yard: a Jetson Orin Nano pulling the cameras’ RTSP streams locally, a person/vehicle model on it, and a small script that sends results to Hermetry as normal telemetry every 10 s:
{"yard.people": 2, "yard.forklifts": 1, "nogo.people": 0}Then alarm rules and charts work as for any device, e.g. nogo.people > 0 with a time-window condition for after hours on a Rules card rule. The cameras stay in Hermetry for live view, so the person who gets the alarm can open the camera.
Cheaper first step: many cameras do their own motion or line-crossing detection. If yours are supported for camera events, switch events on for the endpoint and you get event.motion as 0/1 readings on the camera device, with ready alarm rules on the profile (motion is off by default). Not object detection, but it costs nothing.