A Robust Deep Learning Architecture for FireFighter PPEs Detection

Abstract

Personal Protective Equipment (PPE) is one of the primary defence mechanisms to reduce the exposure of the personnel to hazardous environments. It's significantly important to Fire Fighters as they are constantly exposed to dangerous elements such as fire, gas or chemicals. Unfortunately, in real-time emergencies, such as fires, it is very difficult to identify if a responder using PPE is fully equipped to reduce any accidents in the workplace or even coordinate response actions due to the high pace of the situation. A lack of a unified Fire Fighting PPE image dataset was also observed, which makes the task of training Machine Learning (ML) models to solve this problem a challenge. To that end, we first create a general purpose FireFighter Equipment Detection dataset. We then propose to utilise the widely used YoloV5 Deep Network architecture to detect different PPE components in real-time. This work leverages the pretrained YoloV5 model, using transfer learning to fine-tune the model using the created detection dataset that contains targeted Fire Fighter PPE images. By employing the pre-trained model which requires substantially fewer training samples, we were able to achieve a considerably good performance on the Fire Fighter PPE object detection. The proposed method can distinguish four different PPE components such as a Helmet, Gloves, Mask or Insulated protective cloth, achieving high detection efficiency which is experimentally established.