Abstract
Ensuring consistent compliance with Personal Protective Equipment (PPE) requirements on construction sites is crucial for worker safety. Although deep learning-based methods already perform well in detecting non-PPE cases, there is still scope to further improve accuracy. Progress is hindered by the difficulty of building representative datasets: strict regulations mandate PPE usage, so genuine non-PPE instances are rare, even though such examples are essential for training robust detectors. To address this challenge, this study develops a Domain Adaptation (DA)-based Faster Region-Based Convolutional Neural Network (Faster R-CNN) for detecting five non-PPE categories: "Non-helmet", "Non-mask", "Non-glove", "Non-vest", and "Non-shoes". The proposed framework augments a standard Faster R-CNN with image-level and instance-level adversarial domain classifiers connected through gradient reversal layers, enabling the model to learn domain-invariant features while preserving detection accuracy. The approach leverages a fully labeled construction-site dataset alongside a general-context dataset, allowing the detector to exploit abundant surrogate non-PPE examples and alleviate the scarcity of real on-site violations. Among the evaluated backbones, ResNet-152 combined with comprehensive data augmentation and tuned hyperparameters achieved the best performance, reaching an mean Average Precision (mAP) of 86.84% on previously unseen construction-site images. Overall, the DA-enhanced detector outperformed conventional supervised baselines by up to 14 mAP points. These findings indicate that combining DA with systematic data augmentation improves the robustness of non-PPE detection under realistic construction-site conditions and provides a practical foundation for extending the approach to broader safety-monitoring applications.