Personal protective equipment detection using artificial intelligence

Abstract

Multiple studies conducted by Singapore’s Ministry of Manpower highlight the vast number of injuries occurring in industrial workplaces annually. Despite the existence of laws designed to reduce injuries by enforcing the usage of personal protective equipment (PPE), there is still a significant risk of workplace accidents due to non-compliance from workers. With the recent advancement in efficient object detection models and the widespread utilisation of surveillance cameras in workplaces, this study proposes the development and implementation of an accurate and efficient real-time PPE detection system. Through comprehensive research and comparison analysis conducted on various object detection models, YOLOv8 was streamlined to be utilised as the baseline model due to its accuracy and advantages in inference speed. Additionally, the expansion of a pre-existing PPE dataset to increase the total number of samples from 9,886 to 12,981 images and the number of classes from 11 to 12 classes was carried out to improve the detection model’s ability to generalise unseen data with more efficiency. With various data pre-processing and augmentation strategies explored to refine the overall performance of the detection model, a PPE detection system utilising the YOLOv8 model was achieved with a mean Average Precision at 0.5 intersection over union threshold (mAP@0.5) of 93.1 % along with an inference speed of 19.4 milliseconds (ms).