An Auditable Real-Time Safety Compliance Detection System using YOLOv8l and Web Integration

Abstract

Conventional approaches to enforcing Personal Protective Equipment (PPE) safety policies in workplaces depend heavily on human observation. This manual routine is inherently variable, demands significant manpower, and is prone to oversight. This paper describes a real time Safety Compliance Detection System that combines a tailored YOLOv8l object detector, a Euclidean-distance tracker, and a Flask web application with a PostgreSQL database. The detection algorithm was trained on a curated collection of 3,235 industrial site images with 14,341 annotations across seven PPE classes: helmet, vest, mask, boots, gloves, goggles, and face shield. A EuclideanDistTracker with a 100-pixel threshold tracked workers across video frames, assigning unique identifiers and enabling continuous compliance monitoring without duplicate alerts. The 7-class network produced an overall mean average precision (mAP@0.5) of 0.928 and maximum F1 score of 0.94 at confidence threshold of 0.548 on the test dataset. Moreover, the inference pipeline achieved 21.5 FPS on a Tesla T4 GPU. These outcomes establish the architecture as a highly accurate, continuous, and traceable tool for managing occupational hazards.