Industrial Safety Helmet Detection Using Single Shot Detectors Models and Transfer Learning

Abstract

Personal safety is concerned to be as a crucial part for the industrial workers while working in an industrial environment. Industries provide personal protective equipment to their workers to ensure their safety, similarly the workers are also meant to wear and follow all the regulations regarding the personal protective equipment's (PPEs) provided to them. Our study provides the methodology to detect the industrial safety helmet using the surveillance cameras. In this study, we have trained two different single shot detector models i.e., Single Shot Detector (SSD) MobilenetV2 and Single Shot Detector (SSD) Resnet50 and used transfer learning methodology to detect the industrial safety helmet. We have utilized a publically available dataset from Kaggle website and utilized that dataset for the purpose of training the models. Furthermore, the models evaluation is done based on these parameters i.e., classification loss, localization loss, regularization loss and total loss. However, we concluded that the SSD Mobilenet V2 performs better than SSD Resnet50 model based on loss parameters. For SSD Mobilenet v2 we achieved a classification loss of 0.11, localization loss of 0.05, regularization loss of 0.15, and a total loss as 0.32 respectively. Moreover, the graphs for the loss of each model has also been studied.