Artificial Intelligence System Based Personal Protective Equipment Detection for Construction Site Safety using YOLOv8

Abstract

This paper addresses the limitations of current deep learning models for detecting Personal Protective Equipment (PPE) on construction sites, where performance enhancement is crucial. This paper use You Only Look Once (YOLO) architecture, focusing on ten categories: \'Hardhat\', \'Mask\', \'NO-Hardhat\', \'NO-Mask\', \'NO-Safety Vest\', \'Person\', \'Safety Cone\', \'Safety Vest\', \'machinery\', \'vehicle\'. A new high-quality dataset, named PPE dataset from Roboflow, was created, comprising 1,330 images that reflect real construction environments, various poses, angles, distances, and multiple PPE types. Among the evaluated models, YOLO v8 achieved the highest mean Average Precision (mAP) of 87.55%, while YOLO v8 demonstrated the fastest processing speed at 52 images per second on a GPU. The study involved training a model using 2,934 images and validating it with 816, resulting in a 95% mean Average Precision (mAP). It underscores the significant role of artificial intelligence in enhancing safety management and occupational health within the construction sector. This research serves as a foundation for future advancements in AI-driven safety measures, addressing the urgent need for innovative strategies to minimize workplace risks and elevate compliance standards in the industry.

Introduction

Context and Problem:
The construction industry is one of the most hazardous sectors globally, with high accident and fatality rates often linked to non-compliance with Personal Protective Equipment (PPE) usage (helmets, vests, boots). Over 70% of fatal incidents involve PPE violations. There is a critical need for reliable, efficient computer vision systems to automatically detect PPE compliance in real-time, improving safety monitoring on-site.

YOLO Framework Evolution:
The YOLO (You Only Look Once) family of object detection models has evolved since 2016, improving speed and accuracy through versions YOLOv1 to YOLOv7. The current project leverages the latest YOLOv8 model, known for superior performance in real-time object detection tasks.

Project Goals:

• Develop a fast, accurate, and scalable object detection system using YOLOv8, OpenCV, and TensorFlow.

• Detect PPE usage in images and live video with adaptability to various lighting and environmental conditions.

• Optimize the system for low-resource devices (edge/embedded systems).

• Provide an easy-to-use interface and seamless integration with existing applications.

Motivation:
YOLOv8’s advanced capabilities benefit many applications such as surveillance, autonomous vehicles, and medical imaging. Developing expertise in this technology offers significant career opportunities and supports innovations in safety monitoring and automation.

Objectives for PPE Detection:

• Ensure safety compliance by detecting PPE in real-time.

• Mitigate risks by identifying non-compliance quickly.

• Automate monitoring and generate actionable reports.

• Support training and emergency responses by verifying proper equipment usage.

Literature Review Highlights:

• Deep learning methods, including YOLO variants, have demonstrated high accuracy in PPE detection and other safety-related applications.

• Comparative studies show YOLO models outperform others like SSD and Faster R-CNN in precision, speed, and real-time detection capabilities.

Methodology:

• The YOLOv8 model was trained on a large Roboflow PPE dataset (~330,000 images) with dedicated splits for training, validation, and testing.

• Data preprocessing involved resizing, cropping, labeling, and augmentation to enhance model robustness.

• Model training and validation followed iterative fine-tuning based on performance metrics like mean Average Precision (mAP).

Software and Tools:

• YOLOv8 framework with PyTorch backend.

• Python environment with OpenCV for image processing.

• Roboflow platform for dataset management and annotation.

Results and Evaluation:

• YOLOv8 achieved the highest precision (0.923), recall (0.87), and mAP@0.5 (0.822) compared to other models (Faster R-CNN, SSD, YOLOv5, YOLOv7).

• Precision-Recall curves demonstrate YOLOv8’s balanced accuracy in detecting PPE across varying thresholds.

Outputs:

The system successfully detected PPE in various real-world images, demonstrating its practical utility for ensuring safety compliance on construction sites and other hazardous environments.

This work highlights the potential of cutting-edge deep learning models like YOLOv8 to enhance occupational safety through automated, real-time PPE detection, improving monitoring, compliance, and ultimately worker protection.

Conclusion

The research indicates that the current dataset is limited, suggesting a need for a larger training set, particularly for safety shoes. Additionally, the testing phase was constrained to just 300 images, which may not represent all real-world scenarios. Employing data augmentation techniques could broaden the dataset to better reflect various construction site conditions. These findings not only confirm the practical utility of YOLOV8 but also point out areas for future enhancements to improve PPE detection accuracy. Among the models evaluated, YOLOv8 demonstrates the highest precision at 0.923, indicating that it has the most accurate positive predictions compared to the other models. This is complemented by a recall of 0.87, suggesting that YOLOv8 effectively identifies a significant proportion of true positive instances. Consequently, its mAP0.5 score of 0.822 is the highest in the table, underscoring its overall superior performance in object detection tasks.

References

[1] Lingard, H., 2013. Occupational health and safety in the construction industry. Constr. Manag. Econ. 31, 505– 514. https://doi.org/10.1080/01446193.2013.816435 [2] Pinto, A., Nunes, I.L., Ribeiro, R.A., 2011. Occupational risk assessment in construction industry – Overview and reflection. Saf. Sci. 49, 616–624. https://doi.org/10.1016/j.ssci.2011.01.003 [3] Waehrer, G.M., Dong, X.S., Miller, T., Haile, E., Men, Y., 2007. Costs of occupational injuries in construction in the United States. Accid. Anal. Prev. 39, 1258–1266. https://doi.org/10.1016/j.aap.2007.03.012 [4] Memon, M.H., Abas, N.H., Sohu, S., 2023. Ranking the critical causes of accidents in construction projects of Sindh: Perspective of safety professionals. IOP Conf. Ser. Earth Environ. Sci. 1205, 012035. https://doi.org/10.1088/1755-1315/1205/1/012035 [5] Sehsah, R., El-Gilany, A.-H., Ibrahim, A.M., 2020. Personal protective equipment (PPE) use and its relation to accidents among construction workers. Med. Lav. 111, 285–295. https://doi.org/10.23749/mdl.v111i4.9398 [6] Ahmed, M. I. B., Saraireh, L., Rahman, A., Al-Qarawi, S., Mhran, A., Al-Jalaoud, J., Al-Mudaifer, D., Al-Haidar, F., AlKhulaifi, D., Youldash, M., &Gollapalli, M. (2023). Personal Protective Equipment Detection: A Deep-Learning-Based Sustainable Approach. Sustainability, 15(18), 13990. https://doi.org/10.3390/su151813990 [7] V. Kumar, V. Goel, A. Amoriya and A. Kumar, \"Object Detection Using SSD,\" 2023 5th International Conference on Advances in Computing, Communication Control and Networking (ICAC3N), Greater Noida, India, 2023, pp. 743-748, doi: 10.1109/ICAC3N60023.2023.10541704 [8] Ferdous M, Ahsan SMM. PPE detector: a YOLO-based architecture to detect personal protective equipment (PPE) for construction sites. PeerJComput Sci. 2022 Jun 17;8:e999. doi: 10.7717/peerj-cs.999 [9] W. E. I. B. W. N. Afandi and N. M. Isa, \"Object Detection: Harmful Weapons Detection using YOLOv4,\" 2021 IEEE Symposium on Wireless Technology & Applications (ISWTA), Shah Alam, Malaysia, 2021, pp. 63-70, doi: 10.1109/ISWTA52208.2021.958742

Copyright

Copyright © 2025 Nirupama Malik , Amiyajyoti Nayak. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.