Netra — AI Surveillance System

Autonomous person-tracking surveillance on constrained hardware, with real-time detection and an anomaly-scoring dashboard.

Autonomous person-tracking surveillance on constrained hardware: an ESP32-CAM streams into a YOLOv8 detection pipeline driving pan-tilt servos over MQTT, with anomaly scoring and patrol heat-maps surfaced through a FastAPI and React console.

Hardware: ESP32-CAM
Detection: YOLOv8
Transport: MQTT

What it had to solve

Surveillance that only records is evidence after the fact. Making it act — follow a person, notice something unusual — normally means a camera with a computer behind it. Doing it on an ESP32-CAM means the whole loop, detection through to servo movement, has to close on hardware with a fraction of the memory and no GPU, over a transport that tolerates an unreliable link.

The decisions that shaped it

A closed perception-to-motion loop

The ESP32-CAM streams into a YOLOv8 detection pipeline whose output drives pan-tilt servos. Detection is not the end of the system, it is the input to a control loop — the camera physically follows what it finds, which is what makes the tracking autonomous rather than an overlay on a recording.

MQTT because the link is not reliable

Commands and telemetry move over MQTT, a publish-subscribe transport built for constrained devices and intermittent networks. On this hardware that matters more than throughput: the loop has to degrade rather than stall when the link does.

Anomaly scoring on top of detection

Raw detections say a person is present. Anomaly scoring is the layer that decides whether that is worth attention, which is what turns a stream of positives into something an operator can act on.