In railway infrastructure maintenance, rising upkeep cost by a declining labor force have become a critical challenge. Visibility inspections of obstruction warning signals—which require verification from distances of up to 800m—are highly labor-intensive, often performed at night and involving long walking distances. This paper proposes a novel visibility inspection system based on a 3D railway environment reconstructed from multi-view images captured by onboard cameras. The system generates a 3D point cloud using Structure from Motion (SfM) and constructs a photorealistic environment using 3D Gaussian Splatting enabling efficient and realistic rendering of complex railway scenes. Virtual viewpoints within the reconstructed space allow accurate remote visibility inspection. Furthermore, an integrated AI model enables automatic detection and evaluation of signal visibility without human intervention. Field tests demonstrate sufficient geometric accuracy and practical applicability. The proposed system reduces labor and operational costs while maintaining high inspection quality and supports the development of automated and scalable railway maintenance strategies.
Maeda et al. (2026) studied this question.