Real-Time Construction Safety Gear Detection Using YOLOv4 with Darknet

Safety in the construction industry is a major concern. This is especially true in developing countries such as Bangladesh which are witnessing massive development projects. There are numerous hazards in the construction industry, many of them life threatening. Most developed countries endeavor to reduce the damages, losses, injuries and deaths resulting from construction accidents by preventing, eliminating, and bypassing the probable occurrences. Unfortunately this is not the case in many developing countries such as Bangladesh which lack a robust safety system and every year thousands of workers are injured or die from accidents on construction sites. Lack of training and knowledge about the safety gear equipment are among the top five causes of these misfortunes. To achieve safety in construction sites, in this study we propose a deep learning based model to actively monitor in real-time the proper wearing of safety gear such as hard hats, gloves, face masks, vests, harnesses and boots by the construction workers. We implement our model using YOLOv4 (You Only Look Once Version 4) with Darknet. Based on the results of experimental tests, the model proved to have 86.93% mean average precision, which was effective for identifying safety gear correctly. Using YOLOv4 and Darknet, these pieces of equipment can be detected and classified simultaneously.