Enhancing Construction Site Safety through Automated PPE Monitoring: A Computer Vision-based Approach

Abstract

The construction industry is known as one of the most accident-prone industries, where conventional Personal Protective Equipment (PPE) compliance checks remain laborious and inadequate for real-time safety monitoring. Although several sensor-based and AI-driven systems exist, many remain limited in scope, difficult to scale, or unsuitable for low-compute environments. This study developed a practical, multi-class, real-time PPE monitoring framework using an optimized YOLOv5s model aligned with industry insights. A PRISMA-based literature review and a semi-structured survey of 74 professionals identified critical PPE items and implementation barriers, informing dataset design and model requirements. Trained on hybrid datasets, the model achieved a mean average precision (mAP) of 86.5% at an Intersection over Union (IoU) threshold of 0.5, along with 91.4% precision, 87.5% recall, and an inference speed of approximately 140 frames per second (FPS). The resulting framework provides a scalable and easily deployable solution that integrates with existing CCTV systems to automate PPE compliance monitoring on construction sites.