Development and Optimization of a Construction Personal Protective Equipment (PPE) Detection Model on YOLOv8 Architecture

Abstrak

Workplace safety in the construction sector remains a critical issue due to frequent accidents caused by non-compliance with Personal Protective Equipment (PPE) regulations. Manual supervision is inefficient and prone to errors, necessitating an automated detection approach. The prior YOLOv5 version trained on the Construction Safety dataset from Roboflow-100, achieves a mean Average Precision (mAP@0.50) of 0.867. However, class imbalance, particularly the underrepresentation of "no-helmet" and "no-vest" categories, limited detection performance. This study improves the model by tuning hyperparameters for optimal training using grid search and applying data augmentation techniques to address dataset imbalance. Mosaic and Mixup augmentation technique is applied on the dataset. The augmented dataset is used to retrain YOLOv8, further optimizing detection accuracy. Results indicate an improved mAP@0.50 of 0.921, demonstrating enhanced performance in PPE violation detection. These refinements aim to strengthen workplace safety enforcement through more accurate and balanced PPE detection.