Surveillance Video Analytics for Automated Detection of Personal Protective Equipment using Deep Learning

As Malaysia continues its rapid development, ensuring worker safety at construction sites has become a paramount concern. Personal Protective Equipment (PPE) plays a critical role in preventing accidents, but manual monitoring of PPE compliance is often inefficient and prone to errors. This paper presents a video analytics system designed to detect PPE violations and automate people counting using state-of-the-art deep learning models, YOLO-NAS and YOLOv8. The system focuses on identifying PPE compliance by detecting helmets and safety vests with at least 50% overlap within the bounding box of detected individuals. Compliance violations are flagged, enabling real-time interventions. The system was tested on a variety of video samples captured from construction sites, achieving accuracy rates between 50% and 76% in detecting PPE violations and counting individuals. The varying accuracy values stem from challenges such as video quality and recording angles. Although there are these shortcomings, the proposed system is still capable of significantly improving the PPE monitoring process, reducing manual supervision errors, and enhancing overall workplace safety.