Enhancing Worker Safety: Real-Time Automated Detection of Personal Protective Equipment to Prevent Falls from Heights at Construction Sites Using Improved YOLOv8 and Edge Devices
Doil Kim, Shuping Xiong
Abstract
Personal protective equipment (PPE), including helmets, harnesses, and lanyards, is pivotal in preventing falls from heights at construction sites. However, ensuring consistent and correct usage of PPE presents a significant challenge. To address this issue, this study introduces an enhanced You Only Look Once, version 8 model (YOLOv8), a computer-vision-based AI model tailored for real-time multiclass PPE monitoring on portable edge devices. A pioneering large-scale multiclass PPE data set is curated to facilitate model training. Balancing detection accuracy with a lightweight design, we augment YOLOv8 through the integration of the coordinate attention module, ghost convolution module, transfer learning, and merge-nonmaximum suppression. The proposed model surpasses the original YOLOv8 and state-of-the-art models, showcasing improved accuracy and reduced computational cost. Deployed on the edge device Jetson Xavier NX, the model achieves precise PPE detection (mAP50: 92.52%) in real-time, operating at 9.11 frames per second. These findings establish a robust foundation for the efficient and real-time automated safety monitoring of construction sites, promising substantial enhancements to worker safety and data privacy within the construction industry.