SkyGuard-AI

Construction sites and industrial facilities remain among the most hazardous workplaces, and many incidents are linked to missing personal protective equipment (PPE), unauthorised entry into restricted areas, and workers moving too close to operating machinery. Conventional safety audits depend on periodic manual walkthroughs and fixed CCTV cameras, both of which leave spatial and temporal blind spots. SkyGuard AI is a drone-based safety audit system that combines an autonomous patrol flight with a real-time computer-vision pipeline, geofencing logic, a telemetry-aware backend, and a live web dashboard.

The vision component uses a YOLOv8n detector fine-tuned in three phases on 12,691 images merged from three public Roboflow datasets and remapped to six classes: helmet, no-helmet, vest, no-vest, person, and machinery. Five violation types are generated from these detections: no-helmet and no-vest directly from the detector, person-down through a bounding-box aspect-ratio heuristic, restricted-zone entry through a Shapely point-in-polygon test, and machinery proximity through a haversine distance in GPS mode or a pixel-distance fallback. The first-pass PPE model (ppe_v1) achieved an mAP@0.5 of 0.816 on its validation split, and the six-class models full_v1 and full_v2 achieved 0.694 and 0.687 respectively on the held-out test split. The weakest classes are no-vest, person, and machinery, which are also the least represented in the merged dataset.

The flight software was developed and tested in PX4 Software-In-The-Loop simulation with Gazebo and QGroundControl. Python MAVLink scripts read live telemetry, upload a four-waypoint patrol mission, switch the vehicle to mission mode, and trigger Return-to-Launch. A FastAPI backend stores violations in SQLite, aligns each violation with the nearest telemetry fix, and pushes updates to a React dashboard over WebSockets. Physical drone assembly, onboard Raspberry Pi inference, and real flight footage are pending, and all proxy or simulated results are labelled as such in this report.

Keywords: Construction Safety, PPE Detection, YOLOv8, Unmanned Aerial Vehicle, PX4 SITL, MAVLink, Geofencing, FastAPI, Real-Time Monitoring