AI-Driven Extraction of Road Geometry and Asset Inventory from Mobile LiDAR Point Clouds

Abstract. The rapid urbanization and rising traffic volumes strain transportation infrastructure, demanding efficient road design auditing and asset management. Conventional manual surveys are labor-intensive and lack holistic three-dimensional context. This research presents an end-to-end methodology combining mobile LiDAR with an AI model to automate extraction of road geometric parameters and inventory features. Mobile LiDAR data from a Bengaluru corridor was preprocessed using Trimble Business Center, applying a statistical outlier removal filter and progressive morphological ground segmentation. A custom PointNet++-based deep learning architecture with hierarchical set abstraction layers was trained on manually labelled point cloud subsets ( 45 million points, 10% labeled) to classify roads, poles, vehicles, trees, and buildings. The model achieved 0.86 mean Intersection-over-Union (mIoU) and 92.4% overall accuracy on semantic segmentation. Key parameters—lane width (8.099 m), road length (44.383 m), zebra crossing dimensions (7.336 m), and pole height (7.890 m)—were accurately extracted. The automated workflow reduced manual processing time by 85% (from 40 to 6 hours per km), improving repeatability and scalability across urban corridors. Results confirm that the proposed AI-driven workflow significantly reduces manual effort while providing high-accuracy datasets for infrastructure planning. This study demonstrates the transformative potential of integrating mobile LiDAR and AI, offering a scalable tool for safer, more sustainable transportation systems.