Advancing Sustainability in Civil Infrastructure through Semantic Segmentation of Road Surface Using 3D Point Cloud Data

To advance effective road maintenance, we introduce a specialized sub-dataset based on the Santiago Urban Dataset (SUD), named SUD-ROAD. Unlike current public urban point cloud datasets that only include roads and road markings, our dataset provides detailed segmentation of road surfaces. Covering 1,635 meters and containing 57 million points, it has seven different classes: Roads, Lane Lines, Road Markings, Manhole Covers, Drains, Cracks, and Road Patching, allowing for a thorough evaluation of road conditions. Based on the flatness of road surfaces, we transformed 3D point clouds into 2D images, thus reducing the complexity of the segmentation task from three dimensions to two dimensions. We used the ConvNeXt model for training, and the results showed excellent segmentation performance. This confirms the effectiveness of converting 3D data directly to 2D and attests to the reliability of our SUD-ROAD dataset. We also analysed the impact of intensity and geometric properties on segmentation effectiveness across different categories. These results demonstrate the value of the SUD-ROAD dataset in improving sustainability in civil infrastructure.