To address the limitations of manual supervision and cloud-based monitoring systems in high-risk industrial environments, this study presents an edge-based personal protective equipment (PPE) detection system and analyzes its performance characteristics. The proposed system is implemented on a low-power edge platform by combining a Raspberry Pi 5 with a Hailo-8 neural processing unit (NPU), and employs a YOLOv8s object detection model optimized through post-training quantization (PTQ). By converting the trained FP32 model to an INT8 representation, the model size is reduced by approximately 54.7%, enabling efficient deployment on resource-constrained edge hardware. Experimental evaluations show that the optimized system achieves an inference speed of 32.99 frames per second (FPS) while maintaining an mAP@0.5 of 0.8817 on the test dataset. In addition, system-level analysis indicates that offloading inference to the NPU significantly reduces CPU utilization and thermal load compared to CPU-only execution. Qualitative experiments conducted under low-resolution and partial occlusion conditions further demonstrate that the system is capable of detecting workers and PPE items in practical construction site scenarios. These results suggest that the proposed edge-based configuration provides a feasible reference for deploying PPE monitoring systems without reliance on cloud servers, while maintaining stable performance under constrained computational environments.
Lee et al. (Sat,) studied this question.