System Overview & Architecture

MRAI is a network of roadside guardian nodes that watch a lane for crashes, obstruction, smoke or gas, and wet surface, and report what they see to a central operations service.

This document describes the deployed pilot architecture. It is not a certification document and makes no regulatory claim.

Raspberry Pi 3B running the vision model, a 5 MP camera, an ESP32 for sensor duty, HC-SR04 ultrasonic ranging, rain and MQ-series gas sensors, an OLED status panel, and a LoRa SX1276 radio for fallback links.

Sensor-gated keyframe inference: the ESP32 polls sensors at 20 Hz; a trigger wakes the Pi to capture a 3–5 frame burst, which YOLOv8n checks at 2–4 fps. The Pi 3B does not run continuous video inference.

Road-ready target: Pi 4/5 with a Coral or Hailo-8L accelerator for continuous 20+ fps, IP66 enclosure, JSN-SR04T sensor, 12 V LiFePO4 with solar.

Inference runs at the edge. Video frames are not transmitted; only derived state leaves the node.

Nodes POST JSON to the ingest endpoint over HTTPS, authenticated with a per-node device key issued at registration.

Unauthenticated packets are rejected. Keys can be rotated by an administrator without re-registering the node.

Accepted packets are persisted as telemetry readings and evaluated against incident rules. Incidents carry severity, node, state, and an audit trail of who acknowledged and resolved them.

A sweeper runs every five minutes and opens an incident for any node that has been silent beyond its window.