Development of a robot vision system for improving workplace safety in construction sites

Abstract

Despite the massive potential of Artificial Intelligence (AI) in improving workplace safety, AI has been largely under-utilized in construction sites. The aim of this report is to utilize Object Detection to assist site supervisors in solving two key challenges, which contribute significantly to workplace accidents - the inappropriate usage of Personal Protective Equipment and the difficulty in predicting forward collisions. This work details the elaborate techniques used to construct the image dataset, which is needed to train the Object Detection model, YOLOv2 Darkflow. This will constitute the overall training procedure. The predictive phase is systemically detailed, with the introduction of mathematical functions used and the thorough breakdown of the different tasks in various scenarios. The scenarios are then individually accounted for, with an explanation of the corresponding flow chart and a comprehensive breakdown of the results. Through the introduction of orientation-based detection, the trained predictive model could solve these challenges efficiently, proving its potential and necessity to improve workplace safety in construction sites.