Objective The point cloud denoising and 3D reconstruction of roadways serve as a key step in the digital modeling and analysis of roadways. However, the conventional algorithm based on single filtering fails to effectively remove the noise at varying scales from point clouds. Meanwhile, the existing 3D reconstruction algorithms suffer from low modeling accuracy and high susceptibility to distortion. These necessitate developing methods and technologies to obtain high-quality point cloud data and construct high-accuracy 3D models for roadways.

Methods This study proposed an adaptive classification-based point cloud denoising algorithm using neighborhood radius (R), minimum neighborhood point number (I_min), spatial threshold (σ_c), and feature preservation factor (σ_s). Accordingly, this study designed a deep-learning implicit surface reconstruction method based on signed distance functions (SDFs). By integrating a mean value method, an improved density-based spatial clustering of applications with noise (DBSCAN) algorithm, and an improved bilateral filtering algorithm, this study constructed a technical framework for classification processing. The integration algorithm could automatically analyze the noise types of point cloud data and then efficiently remove noise at different scales via an adaptive mechanism, thus achieving in-depth cleaning of main point cloud data. Then, the local regional features of the point clouds of a roadway were extracted using PointNet++, and local contextual information was learned using an introduced implicit neural network. As a result, the global SDF model was created. Finally, this study constructed a fine-scale 3D roadway model by combining the marching cubes algorithm.

Results and Conclusions Based on the experimental scene of the 1∶1 simulated roadway of the Zhangji coal mine in Anhui Province, this study explored the point cloud denoising and 3D reconstruction roadways in a multi-dimensional space. The results indicate that the integration algorithm developed in this study could adjust the denoising strategy dynamically according to the roadway scene and noise categories. This algorithm delivered adaptive optimization performance, yielding types Ⅰ and Ⅱ errors of 1.54 % and 5.37 %, respectively. Therefore, it can effectively remove large-scale, small-scale, and repetitive noise while preserving the features of main point cloud data. The reconstruction algorithm could reduce holes effectively while maintaining the accuracy and smoothness of the roadway model. Furthermore, it enabled the accurate characterization of the structural details of complex locations, with average, standard, and root-mean-square errors of the reconstructed roadway model of 0.037 m, 0.040 m, and 0.041 m, respectively. Therefore, the reconstructed model can meet the high-accuracy requirements of intelligent mine construction. This study will provide high-quality 3D data for the digital transformation and upgrading of mines, along with their intelligent and precise mining.