AI-driven road inspection with SUD-ROAD: High-resolution LiDAR benchmark and a novel cross-dimensional semantic segmentation pipeline

Aging transportation infrastructure worldwide demands innovative artificial intelligence (AI) solutions for maintenance and monitoring. In this paper, we introduce SUD-ROAD, a new high-resolution dataset and methodology aimed at modernizing road infrastructure management through AI-driven inspection. SUD-ROAD is a specialized subset of the Santiago Urban Dataset, spanning 1635 meters of urban roadway and containing 57 million 3D LiDAR points labeled into seven semantic classes (road pavement, lane lines, other road markings, manhole covers, drains, cracks, and patching). Exploiting the near-planarity of road surfaces, we project the 3D point cloud onto 2D grids, allowing state-of-the-art image-based models to replace more complex 3D networks. A ConvNeXt segmentation model trained on these 2D representations attains a mean Intersection-over-Union of 0.74 and overall accuracy of 0.97, accurately detecting both large-scale assets and fine-grained defects critical for early intervention. We also analyzed the impact of intensity and geometric properties on segmentation effectiveness across different categories. By enabling real-time, AI-driven condition assessment, our approach supports proactive repairs, extends asset life, and reduces life-cycle costs—advancing the broader goal of safer and more sustainable transportation infrastructure. The dataset can be accessed at the following repository: https://github.com/msqiu/SUD-Road.