Detection of Unsafe Behavior in conveying Vehicle Parts using Computer Vision

Abstract

Deep Learning (DL) has experienced notable growth in various applications, which highlights its use in vision systems for object detection. The present work proposes a proof of concept for detecting unsafe acts in a vehicle assembly plant. The employment of Convolutional Neural Networks (CNNs) for either object or event detection was studied, and a vision system specifically trained for real-time detection of unsafe acts carried out by personnel while conveying car body parts was implemented. The intention of this research is to prevent workplace accidents and promote safety in the production environment by creating a personalized dataset composed of images that capture some incorrect ways of loading the car body doors, labeled as unsafe acts. For this purpose, a YOLOv8 DL model was trained to recognize unsafe behaviors, and after the test execution, the system efficiently identified safe and unsafe acts. Therefore, the proposal is feasible to be deployed to improve surveillance in daily operations, deliver automated reports for decision-making, and establish countermeasure actions.