Pavement Point Cloud Upsampling Based on Transformer: Toward Enhancing 3D Pavement Data

The accurate representation of pavement in three-dimensional (3D) space is pivotal for road infrastructure management. However, the sparse measurement from affordable equipment, such as 2D LiDAR, has limited the fully utilization of 3D pavement data. This paper aims to address this challenge by exploring the point cloud upsampling for pavement area. Specifically, Pavement-PU, a novel data-driven model is designed to enhance the quality and efficiency of dense point cloud generation task. Utilizing a transformer-based feature extraction module coupled with an integrated point upsampling strategy, Pavement-PU significantly improves the density, uniformity, and accuracy of point clouds derived from sparse and irregular initial scans. Through rigorous testing on the public dataset and a specially curated Pavement3D dataset, the model demonstrates substantial improvements over existing methods in terms of both quantitative metrics and qualitative assessments. Ablation studies further validate the impact of our architectural choices, confirming the effectiveness of the innovative structures implemented within the network. Our research paves the way for more effective and efficient methods in pavement maintenance and monitoring, leveraging advanced techniques for practical, real-world applications in pavement management.