Been running this in my homelab for a few months. Here's what's actually in the stack:

Hardware: NVIDIA Jetson Orin Nano 8GB, 2x IP cameras, phones on LAN as additional camera sources.

What it does: Real-time person detection and tracking across all cameras simultaneously. When someone disappears from one camera's view, the system renders a ghost overlay on the other cameras showing where they should be, based on a self-calibrating homography it learned from watching when both cameras saw the same person.

Performance numbers on Jetson:

YOLOv8n via TensorRT FP16: sub-10ms inference

Adaptive Kalman filter per tracked person: 6-state [x, y, z, vx, vy, vz]

Per-prediction homography projection: effectively 0ms (one 3x3 matrix multiply)

10 viewers connected = 1 GPU inference pass (singleton pipeline)

The architecture that made this work on edge hardware: one shared perception pipeline runs detection, tracking, and fusion once per tick, then broadcasts pre-serialized msgpack snapshots.…