Dense 3D Neural Map Reconstruction Only Using a Low-Cost LiDAR

Abstract

Scanning plays a vital role in civil engineering, automation construction, and remote sensing, which is the basis for high-level tasks like Building Information Modeling (BIM) and construction quality control. However, static scanning requires a significant amount of manual labour and time. Mobile scanning has reduced the time and post-processing work compared to traditional static scanning. However, the point cloud density in mobile scanning is often lower than that of static scanning. To address this challenge, we propose a novel approach that only relies on a cheaper LiDAR to enhance scanning resolution. Our approach employs explicit LiDAR odometry for odometry estimation and AI-based neural 3D dense mapping. The approach combines the high-precision registration of current explicit LiDAR odometry with the strategy of jointly optimising global temporal information through neural map representations. As a result, it achieves high accuracy in localisation and increased density in mapping. These case studies on building scenes demonstrate that this method significantly improves the point cloud density of original scans.