Fast Personal Protective Equipment Detection for Real Construction Sites Using Deep Learning Approaches

Area of Science:

• Computer Vision
• Artificial Intelligence
• Construction Safety Engineering

Background:

• Existing deep learning Personal Protective Equipment (PPE) detectors have limitations in variety and performance for real construction environments.
• Improved PPE detection is crucial for enhancing worker safety on construction sites.

Purpose of the Study:

• To develop and evaluate deep learning models for detecting multiple types of PPE in realistic construction settings.
• To introduce a new, high-quality dataset (CHV) for training and benchmarking PPE detection models.

Main Methods:

• Training and evaluating eight You Only Look Once (YOLO) deep learning detectors on a custom dataset (CHV).
• The CHV dataset comprises 1330 images with diverse construction site conditions, including various PPE classes (helmets, person, vest), angles, and distances.
• Performance comparison based on mean Average Precision (mAP) and inference speed (FPS).

Main Results:

• YOLO v5x demonstrated the highest performance with an 86.55% mAP.
• YOLO v5s achieved the fastest detection speed at 52 FPS on GPU.
• Models trained on the CHV dataset outperformed other deep learning approaches on similar datasets, with a minor 7% accuracy decrease for helmets on blurred faces.

Conclusions:

• The developed YOLO-based detectors, trained on the CHV dataset, offer improved performance for real-world PPE detection in construction.
• The novel CHV dataset is publicly available to advance research in construction safety monitoring.