In safety-critical applications such as Personal Protective Equipment (PPE) monitoring and hazard detection, real-time object detection on edge devices is essential. However, most high-performance object detection models are designed for powerful hardware, making their deployment on resource-constrained IoT devices challenging. This study focuses on evaluating lightweight object detection models optimized for edge computing, specifically targeting models with fewer than 3 million parameters. We introduce a novel, labeled PPE detection dataset named PPE-Det, consisting of 5,000 images captured at varying distances to enhance detection robustness. The dataset is designed to detect helmets and three types of coveralls (red, blue, yellow) commonly used in industrial environments, providing a valuable benchmark for assessing detection models under real-world conditions. Several models, including YOLOX-Nano, YOLOv8n, YOLOv9t, YOLOv11n, and MobileNet-SSD, are compared based on accuracy, latency, and computational efficiency. The models are deployed and tested on a Raspberry Pi 5 using PyTorch and NCNN frameworks to assess real-world performance. Experimental results indicate that YOLO-based models achieve the highest mean Average Precision (mAP) scores, with YOLOv9t and YOLOv11n demonstrating superior accuracy and efficiency. Performance evaluation across different deployment formats highlights trade-offs between accuracy and computational cost, where lightweight models such as MobileNet-SSD and NanoDet-M provide faster inference at the expense of reduced detection accuracy. Our findings offer valuable insights for optimizing object detection on IoT edge devices, contributing to safer and more efficient industrial and security applications.