Abstract
Safety is an important aspect of every mining operation, in particular in underground conditions. Personal Protective Equipment (PPE) is considered the last line of defense as it helps minimize exposure to hazards for workers. Frequently, access control to different mine levels and proper PPE use assessment relies either on human monitoring, which can be difficult in large operations, or in wireless systems such as RFID technology integrated into wearable devices, which is not available in smaller operations. This work presents the development and validation process for a surveillance camera-based system that can provide real-time assessment of proper PPE use and workers access to different areas in underground mines, which can be highly valuable when planning to minimize hazards and to properly plan for rescue operations when needed. The approach considered takes advantage of state-of-the-art machine learning algorithms, operating over off the shelf components, suitable for mining environments, to create a platform that can be easily integrated into underground mines. The workflow considers the creation of a representative dataset containing images of workers of diverse ages, genders and ethnic backgrounds using different brands of PPE, as well as the selection of proper computer vision processing algorithms, such as convolutional neural networks (CNN), suitable to identify proper PPE use and number of workers in particular regions of a mine. The proposed design was validated with the implementation of a functional prototype at the underground San Xavier Mining Laboratory at the University of Arizona, showing that computer vision represents a valuable tool to provide automatic and timely safety information from existing surveillance hardware in underground mines.