Enhancing Mobile LiDAR Point Clouds for Urban 3D Reconstruction: A Comparative Study of Denoising Techniques and a 3D Reconstruction-Based Evaluation Framework

By Xinyi He

Abstract
In recent years, the demand for small-scale, high-precision urban datasets has grown significantly, yet remains constrained by high costs and limited accessibility. Mobile Laser Scanning (MLS) technology offers a promising solution due to its operational flexibility and efficiency. However, MLS point clouds face challenges like data sparsity and significant noise, inadequate evaluation metrics, and processing capabilities for complex scenarios. These issues severely limit MLS applications in urban environments, making MLS quality enhancement and evaluation innovation essential.

This study proposes an integrated data processing workflow for MLS point clouds in urban environments. Utilizing the PRISMA systematic literature review method, this study selected and compared the denoising performance of 18 algorithms on MLS point clouds. The quality evaluation method for denoising comprises two key steps. The first step involves a comprehensive scoring system and visual interpretation of preliminary screening. The second step introduces an innovative quantitative evaluation framework based on 3D reconstruction to assess the performance of the denoising effects. To validate the performance of this workflow, this study conducted experiments using roof and road datasets from Bildtsestraat Street in Leeuwarden, the Netherlands. The experiments evaluated different denoising algorithms for their effectiveness and robustness with urban MLS point clouds. The experimental results show that (1) Normal-based algorithms (including normal-guided filtering, normal-based bilateral filtering, and normal voting tensor) achieved consistently superior performance on both datasets. (2) The method combining random sampling with Bayesian regression demonstrated adaptability, effectively handling urban scenarios with different noise levels. (3) In contrast, voxel-based and patch-based KNN methods showed limited reliability, performing acceptably in simple road scenarios with low noise but significantly worse in complex roof scenarios.

By screening denoising algorithms, validating cross-scenario applicability, and establishing a novel evaluation framework, this study provides important references for urban point cloud analysis. The research results have positive significance for promoting the practical application of MLS technology in urban

Keywords: Mobile Laser Scanning (MLS); Point Cloud Denoising; Urban 3D Reconstruction