YOLOv4 algorithm for the real-time detection of fire and personal protective equipments at construction sites

Abstract

Many difficulties are encountered during evacuation from construction sites in hazardous situations, which may lead to severe fatalities. These fatalities, especially caused by fire, may be significantly reduced by ensuring personal protective equipment (PPE) compliance of construction site workers and fire detection through proper surveillance. Thus, the detection of PPEs, fire and injured or trapped persons, can greatly assist in the reduction of fatalities and economic loss. This article presents a novel approach towards the detection of fire and PPEs to assist in the monitoring and evacuation tasks. This work utilizes the YOLOv4 and YOLOv4-tiny algorithms based on deep learning for carrying out the detection task. A self-made dataset has been utilized to train the model in the Darknet neural network framework. Moreover, a comparative analysis with previous works has been carried out in order to endorse the real-time efficacy of the proposed work. The results verify the strength of YOLOv4 algorithm in real-time detection and surveillance at construction sites with maximum mean average precision (mAP) of 76.86 %.