Most people think real-time computer vision needs a GPU and a cloud bill. So they're surprised when I tell them I ran autonomous person-tracking on an ESP32-CAM — a microcontroller that costs about the price of a coffee.
That constraint is the whole point of Netra, and it forced the kind of engineering you never learn when the hardware is doing the work for you.
Why tiny hardware is the interesting problem
Anyone can call a vision API from a beefy server. The moment you push intelligence onto a constrained device — kilobytes of RAM, no operating system luxury, a camera streaming over Wi-Fi — every assumption breaks. You can't be wasteful, because there's nothing to waste. And you learn what actually matters.
Netra is an ESP32-CAM streaming live video, processed through a YOLOv8 detection pipeline, driving pan-tilt servos that physically follow a detected person — all coordinated over MQTT, with a live React dashboard on top. It's a full edge-to-cloud stack: on-device firmware, edge detection, a Python inference server, and a real-time frontend, built as one cohesive system.
The details that make it work
- Streaming that doesn't stutter — MJPEG at ~15 FPS, 640×480, with PSRAM-backed dual buffering so the frame being read never collides with the frame being captured. A 4-second auto-reconnect recovers dropped streams on its own.
- A message bus, not hard wiring — MQTT decouples every component. The detector doesn't know how many things listen to it; the dashboard doesn't need a socket to the camera. That's what let me develop firmware, inference, and UI independently and have them find each other over topics.
- Scalable by design — ESP-NOW mesh handoff supports tracking a target across up to 8 camera nodes.
- Signal, not noise — a behavioral anomaly engine does multi-factor threat scoring and time-weighted patrol heat-maps, so it can tell "someone walked through at noon" apart from "someone is lingering at 3 a.m."
The lesson
Constrained hardware is the best teacher in engineering. It punishes every lazy shortcut and rewards you with an intuition for what's actually essential. Building vision on a $6 chip taught me more about systems design than any amount of cloud compute ever did — because the cloud lets you paper over the very decisions the microcontroller forces you to make.
The full architecture, wiring, and the tracking control loop are on the project page.
👉 See the build: www.divyakush.com/projects/netra
Divyakush Punjabi — Full-Stack & AI Systems Engineer
🌐 https://www.divyakush.com · 💼 LinkedIn · 💻 GitHub
