PointCleaner: Dynamic Manifold Path Optimization for LiDAR Point Cloud Denoising

Light detection and ranging point cloud denoising is a critical challenge for constructing high-precision digital surface models in complex terrains, primarily due to the strong coupling between noise and terrain features. Current methods neglect manifold topological constraints and the non-stationary distribution of noise, leading to geometric distortions and residual noise. To address this, we propose PointCleaner, a point cloud denoising framework based on dynamic optimization of manifold paths. The core idea is to construct an optimal dynamic manifold path from noisy points to the true terrain surface for precise denoising. Specifically, our approach includes: (a) multi-scale geometric awareness by combining hierarchical neighborhood sampling with a dynamic graph convolutional network to extract complex terrain features adaptively; (b) manifold constraints and normal joint optimization to enhance local geometric consistency and preserve terrain topology; and (c) a path selector for dynamic path optimization, combining a global manifold projection path with a local gradient field to adjust the denoising path. Experimental results show that PointCleaner significantly outperforms existing methods on the HFUT-OP360 dataset, reducing chamfer distance and point-to-manifold metrics by 13.30% and 16.20%, respectively. Digital surface model reconstruction further demonstrates its advantage in preserving key terrain features and fine details, whereas experiments on an ISPRS dataset validate its excellent generalizability and robustness.