Abstract: 3D reconstruction of underground roadways is an important approach for mine surveying. 3D laser scanning combined with Simultaneous Localization and Mapping (SLAM) technology enables roadway scanning and 3D reconstruction. However, in underground environments with sparse geometric features, there exist problems such as insufficient point cloud registration accuracy and low efficiency. Taking the auxiliary haulage roadway in the southeastern section of the south wing of Ningtiaota Coal Mine under Shaanxi Coal and Chemical Industry Group as the engineering background, this study conducted high-precision modeling of underground roadways. Considering the characteristics of the underground environment of Ningtiaota Coal Mine, control points were arranged on the roadway roof to reduce environmental interference on their positions. A GOSLAM-GSJH12 handheld 3D laser scanner was used to collect point cloud data. By constructing known point constraints and performing nonlinear optimization on the point cloud coordinates based on the known control point coordinates, point cloud drift was corrected. Denoising algorithms such as Wavelet Decomposition and Non-Local Means, along with the deep learning segmentation algorithm based on PointNet++, were applied to remove noise in the point cloud data. Roadway point cloud features were extracted using an improved Harris3D corner detector and the Random Sample Consensus (RANSAC) algorithm. Point cloud registration was performed by fusing data from 3D LiDAR and the Inertial Measurement Unit (IMU), enabling high-precision map construction. The Delaunay Triangulation algorithm was adopted to construct an irregular triangular mesh model of the underground roadway, and multi-stage optimization was used to achieve fine 3D reconstruction, which was finally presented via a visualization platform. The research results can be integrated with technologies such as the Internet of Things, big data, and artificial intelligence to realize intelligent mine management and decision-making.