A 3D Point Cloud Dataset for Road Infrastructure and Evaluation of Semantic Segmentation Algorithms

Accurate perception of road infrastructure is crucial for applications like autonomous driving and intelligent transportation systems. However, progress in 3D reconstruction and analysis is hindered by the scarcity of specialized, annotated point cloud datasets for diverse road environments. Methods: To address this gap, this study constructed a comprehensive 3D point cloud dataset specifically targeting road infrastructure. We employed a ground mobile scanning vehicle and drone-borne LiDAR for synergistic data acquisition, capturing three distinct road types: national highways, municipal roads, and rural roads. Following multi-source data fusion, approximately 100 million points covering roughly 7 km of roadways were manually annotated with fine-grained semantic labels. Algorithm Evaluation: Considering the complexity and scale of road scenes, we systematically evaluated the performance of representative semantic segmentation algorithms (PointNet, PointNet++, and RandLA-Net) on this novel dataset using the Mean Intersection over Union (mIoU) metric. Results & Conclusion: Experimental results demonstrated that RandLA-Net significantly outperformed others, achieving a remarkable average mIoU of 91.09% across all three road types. It exhibited strong capabilities in recognizing key road features, particularly traffic signs and light poles. This study successfully provides both a valuable, large-scale annotated point cloud dataset for road infrastructure and establishes RandLA-Net as an effective and efficient semantic segmentation framework for such complex scenarios.