Visual recognition using artificial intelligence (safety detection in the workplace using artificial intelligence)

Abstract

Numerous studies have concluded that non-compliance with personal protective equipment (PPE) requirements dramatically affects the workplace’s safety level. Though currently available strategies for detecting compliance of PPE requirements could provide rapid and fast detection of helmets, it lacks the capability of detecting other PPE such as vests, gloves, and masks. Furthermore, the lack of dynamic user interfaces further complicates the deployment and application of such techniques. Therefore, the objective of this project is to design a real-time PPE monitoring system that is both efficient and accurate. To achieve this, different you only look once (YOLO) models were tested and benchmarked against one another, and YOLOv5s was selected for its accuracy and detection speed. After selecting the model, multiple experiments such as hyperparameters fine-tuning, model structure modification and data augmentation were performed to increase detection accuracy further. Meanwhile, a novel dataset was constructed containing 3414 high-resolution images with 28,977 instances across 8 different classes. With the new dataset, the trained model obtained a 69.5% mean average precision at 32 frames per second. In addition, a flexible graphical user interface was developed to enable users to customise detection features as well as the camera source. Finally, a geofencing function was also implemented to allow users to customise the precise monitoring areas.