Send a camera frame to the cloud, run detection, send a command back — and by the time your servo moves, the person has already walked out of frame.
We treat latency like a number to optimize later. For anything that touches the physical world, it's a design constraint you decide up front. A few hundred milliseconds is invisible in a web request and an eternity for a motor trying to track a moving target.
If the intelligence lived next to the sensor instead of in a datacenter, the loop would close in real time. Detect, decide, actuate — all local, no network in the critical path. The cloud is still useful, just not for the part that has to happen now.
That's the architecture behind Netra: an ESP32-CAM streams frames into a local YOLOv8 pipeline that drives pan-tilt servos over MQTT, so tracking reacts immediately instead of waiting on a round-trip. The dashboard and heavier analytics can live elsewhere — the reflex can't.
Draw the boundary first: what has a physical consequence stays at the edge; what's inherently remote can be remote. Get that line right and everything downstream gets simpler.
More edge and systems work → www.divyakush.com
Top comments (0)