YOLOv4-based Deep Learning Approach for Personal Protective Equipment Detection

Abstract

Detecting Personal Protective Equipment (PPE) has become essential for assuring worker safety and regulatory compliance in numerous industries. This study presents a deep learning approach for PPE detection using the YOLOv4 architecture. The primary objective is to develop a robust model capable of identifying ten PPE classes. The dataset used for training and evaluation consists of 2,605 images for training, 114 images for validation, and 82 images for testing, with checks performed to prevent data leakage. The proposed model architecture is based on YOLOv4 and comprises 225 layers. It incorporates convolutional layers, spatial pyramid pooling, skip connections, and data augmentation techniques to enhance detection performance. The model is trained for 100 epochs using Stochastic Gradient Descent (SGD) optimization with a learning rate of 0.01. Evaluation metrics, including precision, recall, and mean Average Precision (mAP), are employed to assess the model's effectiveness. Experimental results demonstrate the model's proficiency in detecting certain PPE classes, such as Mask and machinery, with high precision and recall scores. However, challenges are encountered in accurately detecting the absence of safety items and localizing vehicles. Precision-recall curves reveal trade-offs between precision and recall for safety-related objects, while precision-confidence and F1-confidence curves indicate performance improvements at higher confidence thresholds. A comprehensive analysis of class-wise performance metrics reveal that the vehicle and Person classes exhibit higher box, object, and classification losses, indicating difficulties in accurate localization and classification. Conversely, the Mask class achieves the highest precision, and the machinery and Mask classes demonstrate strong recall performance. This study contributes to the advancement of PPE detection by presenting a deep learning approach using YOLOv4 and conducting a thorough performance analysis across various PPE classes. The findings highlight the importance of detailed performance evaluation to identify class-specific challenges and guide future research efforts in enhancing PPE detection accuracy and robustness. The proposed approach can be integrated into real-world safety monitoring systems, promoting worker safety and compliance in industrial settings.

Indexing terms: n, , , ,