Personal Protective Equipment use inspection, real time surveillance with YOLO

Abstract

Two situations are necessary to improve the safety and health conditions of people in industries: A proper use of Personal Protective Equipment by workers who develop their jobs into industrial areas and a continuous inspection promoted by employers. In this paper, we review the best known object detectors for prediction tasks, evaluating the features that are needed for a real-time surveillance system that can provide reliable predictions in a safety and health context. We present a tool based on Computer Vision which make predictions about if people are using their respective High Visibility Vest and Hard Hat. Additionally, the same model predicts if a person is wearing a face cover or there is an absence of this protector on the person's face. Our dataset was built by collecting images from available online datasets and we added own images. It has six classes which were defined by the combination of four instances. For detection task, we explored and evaluated different versions of You-only-look-once (YOLO). The best performance for a real time surveillance was reached by YOLO v5-nano, 125.3 FPS. On the other hand, the most important outcomes in terms of average precision were achieved by YOLO v4 and YOLO v5-small with 0.892 and 0.916 respectively.