Edge Embedded Deployment

Edge deployment moves vision onto cameras and gateways, so aisles keep counting during outages with small power and heat.

Why Does This Exist?

A store with 200 cameras cannot stream all video to cloud: bandwidth bills soar and a cut cable blinds theft alerts. Edge deployment exists to split work: cameras detect, gateways aggregate, cloud trains. Power caps at 15 watts, heat caps in ceilings, and updates must roll without bricking aisles.

Chips follow. Jetson serves GPU edges. Raspberry Pi serves CPU edges. This page is the split and fleet plan before chip picks.

Basics first: edge keeps raw pixels local and sends events plus crops. Heartbeats report health. Queues hold events during outages and drain after.

Think of It Like This

Branch offices with headquarters

Branches handle daily sales locally and send nightly totals up. Headquarters trains staff and sets prices. Internet cuts do not stop registers. Cameras are branches. Gateways are district hubs. Cloud is headquarters. The analogy stops at learning: branches follow playbooks, while edges also adapt thresholds locally within guardrails.

How It Actually Works

Profile per site: 8 cameras at 1080p need 4 TOPS for 15 fps each. Place per camera detect on Jetson Orin Nano at 15 watts, per gateway track and reidentify, per cloud retrain weekly. Cap models at 30 MB and 35 ms. Roll updates canary 5 percent then 50 percent with auto rollback on miss rate rise.

Heat breaks the assumption. Ceiling boxes at 45 C throttle 20 percent. Systems derate clocks 15 percent in thermal design and log throttle minutes per day.

Worked split load

Two hundred cameras times 15 fps equals 3,000 frames per second. Cloud alone needs 3,000 times 35 ms which equals 105 GPU seconds per second, about 4 A100s. Edge handles 28 of 30 fps locally and sends 2 crops per second per camera. Uplink falls 200 times 4 Mbps which equals 800 Mbps to 200 times 0.05 Mbps which equals 10 Mbps, an 80 times cut.

Watch Out For

Cloud metrics that hide edge throttle

Lab 28 fps becomes 21 fps in hot ceilings with missed theft windows. The symptom is green dashboards with field gaps. Fix it by testing in thermal chambers and by alerting on throttle minutes, not just fps averages.

The Quick Version

Edge splits detect on camera, track on gateway, train in cloud.
Keep pixels local, send events and crops only.
Our split cut uplink 800 Mbps to 10 Mbps, 80 times less.
Design for 45 C ceilings with derated clocks.
Roll canary with auto rollback on miss rise.