• object detection
• deep learning
• YOLO
• convolutional neural networks
• personal protective equipment

The continuous adoption of Personal Protective Equipment (PPE) in dangerous working sites has been proven to reduce the risk of accidents and their severity. Automated systems can monitor workers' compliance and the correct use of PPE, such as helmets and safety vests, fostering safety in both indoor and outdoor environments.
Image analysis can be applied in real-time to a video stream in order to check whether a worker is wearing PPE and raise an alarm in case of danger. Machine learning, and in particular deep learning, is a mature technology that can effectively perform real-time object detection, even on low-power devices such as smartphones or small single-board computers, if empowered by a Vision Processing Unit (VPU).
The aim of this thesis is to expand a smart system for PPE detection by enabling the detection of safety vests and gloves, in addition to helmets.
Different models based on Convolutional Neural Networks (CNNs), including YOLOv4 and YOLOv4-tiny networks, are compared and deployed on a Raspberry Pi 4 connected to an Intel Neural Compute Stick 2 (NCS2).
The training process and fine-tuning of the models is carried out on a public dataset which we manually labeled, adding the missing PPE as well as chests and hands. Performance are assessed in terms of accuracy and detection speed, using also a test set composed by frames of a realistic use-case.