The construction industry has increasingly promoted the use of ICT (Information and Communication Technology) to address the declining workforce and improve working conditions. Despite technological advancements, the industry still heavily depends on the visual inspections performed by skilled technicians, leading to significant time and cost burdens for maintaining social infrastructure and surrounding areas. Consequently, there is a growing need for inspection systems leveraging deep learning.
Road inspections cover a variety of facilities, including road surfaces, slopes, drainage systems, and traffic safety installations. Road surface inspections typically involve identifying damage, subsidence, collapses, sinkholes, and fallen objects, with the latter being the most frequently detected issue. Slope inspections focus on detecting landslides, slope failures, and fallen or dead trees and branches that might obstruct traffic or pose threats.
While road surface abnormalities are relatively easy to spot from a driver's perspective, inspecting slopes is more challenging due to limited visibility while driving, increasing the risk of oversight. This study aims to develop an inspection system specifically targeting roadside trees and dead branches in mountainous areas, addressing these challenges.
The research focuses on a segment of National Route 17 in Gunma Prefecture, between Shibukawa and Mikuni Pass. In this mountainous area, typhoons often cause dead branches to get caught in roadside trees, creating hazards. Therefore, it is essential to detect and remove these branches in advance. Given the difficulty of observing roadside trees while driving, this study proposes improving the current inspection process by using YOLOv5, a deep learning-based object detection technology, to identify overhanging branches and employing image processing techniques to detect dead branches caught in them. This approach aims to enhance the efficiency and accuracy of roadside tree inspections.
As a result, through training with YOLOv5, we successfully detected overhanging branches. To identify dead branches entangled in the overhangs, we utilized the rotation-invariant phase-only correlation method, adjusted the viewing angles, divided the images into 9 and 16 segments, and calculated the similarity using Imgsim and the phase-only correlation method. The results showed the potential for detecting dead branches by evaluating the coefficient of variation in similarity as calculated by Imgsim.