ABSTRACT
The construction industry increasingly adopts ICT to counter workforce decline and improve working conditions, yet it remains reliant on skilled technicians for visual inspections, resulting in significant time and cost burdens. Challenges such as limited visibility during inspections, especially of slopes and roadside trees, highlight the need for deep learning-based systems. This study focuses on developing an inspection system for detecting roadside trees and dead branches in mountainous areas, specifically along National Route 17 in Gunma Prefecture, where typhoons frequently cause dead branches to become entangled in overhanging branches, creating hazards. Using YOLOv5, the system detects overhanging branches, while image processing with Imgsim identifies dead branches by dividing images into 9 and 16 segments and evaluating similarity variation coefficients. Results demonstrate the potential of this approach to improve inspection accuracy and efficiency, addressing limitations in current practices and enhancing safety in social infrastructure maintenance.
