A Convolutional Neural Network Based Robust Automated Real-Time Image Detection System for Personal Protective Equipment

Abstract

Statistically, casualties in engineering workplaces often result from one of the following accidents: when people get stuck in the rotating machines, electric shocks or collision with heavy equipment. Most of these accidents can be prevented if the workers make proper use of personal protection equipment (PPE). This paper presents the design and implementation of a functional image detection system that takes a picture of an employee, analyses it, and determines the employee he is appropriately attired to enter a potentially hazardous workplace. This system can help to reduce the liability of company owners, by extension their costs, and can provide level of accident prevention. In this study, a convolutional neural network (CNN) was used to develop three sets of models, namely hard hat model, boot model, and vest model. These were used to detect the appearance of workers and determine if the PPE being worn was in compliance with the stipulated requirements for entry to a particularly hazardous workplace. To determine the performance of the system, each model was validated with two classes of image datasets: normal colour RGB (Red, Green and Blue) and grayscale image. The overall average accuracy of the system, in real-time implementation, then was calculated and determined to be 83.33%.