Two-stage random forest fusion for multi-source point clouds in underground 3D reconstruction

Camera sensors are highly sensitive to factors such as lighting, dust, fog, and lack of texture in underground environments, introducing measurement noise and significantly affecting 3D reconstruction accuracy. Existing point cloud regression and denoising models often struggle with the high-dimensional and unstructured nature of point clouds, increasing data processing complexity and limiting prediction accuracy. Furthermore, while fusing multi-line LiDAR and camera data can improve reconstruction accuracy, it introduces challenges related to data redundancy and computational burden. This paper proposes a fusion method using single-line LiDAR and depth cameras based on two-stage random forest regression. In the first stage, a dense 3D point cloud is constructed from LiDAR data, which serves as the reference for initial denoising of the depth camera point cloud using random forest regression. The second stage refines the denoising by incorporating spatial coordinates and FPFH descriptors through a second application of random forest regression. Experimental results demonstrate that the proposed method achieved at least a 77.42% improvement in noise suppression across five representative tunnel scenarios. In terms of dimensional accuracy, the maximum absolute error was reduced by 72.97% to 95.86%, while the average absolute error decreased by 71.91% to 96.92%, with final absolute errors remaining below 10 mm. These results confirm the method's strong generalization capability and high-precision performance in challenging underground reconstruction tasks. The associated dataset and source code are publicly available at: https://github.com/phn0315/LiDAR-Vision-fusion.git.