What This Reference Design Is
Fall detection that runs entirely on the edge device. A fixed indoor camera, pose estimation on the device's own accelerator, a per-person tracker, and a temporal state machine that decides whether someone went down. When it fires, the device publishes one MQTT JSON message — subscribe to it from Home Assistant, a nursing-call system, an NVR or your own service.
The pose model is YOLO11-Pose (YOLOv8s-Pose on Hailo-8) at 640² input. Each person keeps a track ID while they move around the room, and their state runs normal → suspected → fallen → recovering. The fall_event flag is set only on the transition into the fallen state, so an automation triggers once per fall instead of repeating for as long as someone is on the floor.
Inference, tracking, the fall decision and the MQTT broker all sit on the detector host, so no video leaves the device — what crosses the network is a JSON message of a few hundred bytes. That is what makes it work without internet access and without a per-camera cloud subscription. The trade-off: the fall has to happen inside the frame. Starting the detector on someone already lying down reports the posture and raises no event. This is an assistive alert, not a certified medical or life-safety system.