Denoising of Noisy Point Clouds Using Normal-Guided Cylindrical Neighborhood and Bilateral Weighting

Point clouds acquired by low-cost laser scanning systems have a problem of high noise, which makes the point cloud appear as thick and geometric features blurred, while existing denoising algorithms either fail to maintain a balance between denoising and shape preservation or incur excessive computational cost. To address this issue, this paper proposes a shape-preserving denoising algorithm based on normal-guided cylindrical neighborhood and bilateral weighting. Specifically, the proposed method first optimizes the PCA-initialized normals of the point cloud by integrating curvature-based feature detection and bilateral weighting. Subsequently, a cylindrical neighborhood is constructed for each point along the optimized normal direction. Finally, a bilateral weighted projection mechanism that jointly incorporates spatial and normal features is employed, whereby the aggregated projection of neighboring points drives the displacement of the central point along the normal direction, thereby achieving point cloud denoising. Experiments are conducted on synthetic datasets and real scanned datasets. The results show that, for synthetic data denoising, the proposed method achieves the best or second-best performance in 25 out of 30 experiment cases across different models and different noise levels. For real scanned data, the section views and reconstructed mesh models demonstrate that the proposed method outperforms popular algorithms in removing complex noise while preserving geometric features. In addition, the proposed method demonstrates excellent computational efficiency, capable of denoising at a speed of processing one million points every 2.4 s, and achieves acceleration of processing speed by six times compared to the fastest competitive algorithms.