Accurate, effective, and reliable detection of Personal Protective Equipment (PPE) is a necessary component to ensure the safety of workers on construction sites. This work presents a deep learning system based on YOLOv5s that can effectively identify four PPE-related classes: vest, helmet, no vest, and no hat. The model was trained using a carefully curated dataset compiled from openly accessible industrial and construction site image archives, accounting for real-world variations in posture, lighting, and background. On static image evaluation, the model achieved a mean Average Precision (mAP@0.5) of 0.980, with individual class performance reaching up to 0.986 AP. An F1 score of 0.95 and an optimal operating confidence threshold of 0.513 demonstrate a strong balance between precision and recall. When deployed for live video inference on CPU-only systems, the model maintained real-time throughput (~38.7 FPS), making it well-suited for edge deployments where computational resources are limited. These results underscore the practicality of deploying lightweight yet high-accuracy PPE detection systems in real-world construction environments to enhance on-site safety compliance.