A Unified Denoising Framework for Restoring the LiDAR Point Cloud Geometry of Reflective Targets

LiDAR point clouds of reflective targets often contain significant noise, which severely impacts the feature extraction accuracy and performance of object detection algorithms. These challenges present substantial obstacles to point cloud processing and its applications. In this paper, we propose a Unified Denoising Framework (UDF) aimed at removing noise and restoring the geometry of reflective targets. The proposed method consists of three steps: veiling effect denoising using an improved pass-through filter, range anomalies correction through M-estimator Sample Consensus (MSAC) plane fitting and ray projection, and blooming effect denoising based on an adaptive error ellipse. The parameters of the error ellipse are automatically determined using the divergence angle of the laser beam, blooming factors, and the normal vector along the boundary of the point cloud. The proposed method was validated on a self-constructed traffic sign point cloud dataset. The experimental results showed that the method achieved a mean square error (MSE) of 0.15 cm2, a mean city-block distance (MCD) of 0.05 cm, and relative height and width errors of 1.92% and 1.91%, respectively. Compared to five representative algorithms, the proposed method demonstrated superior performance in both denoising accuracy and the restoration of target geometric features.