Abstract: In view of the difficult modeling problems caused by complex environments such as roadway sealing, high dust, sparse geometric features and others in coal mine downholes, this artical proposes a set of roadway point cloud processing and 3D reconstruction schemes based on 3D laser scanning. Through handheld laser scanner, data is collected, after the registration of "coarse registration+fine registration", the point cloud is analyzed by adopting "statistical filtering+voxel filtering" preprocessing and a mixed strategy of "traditional geometric segmentation+deep learning segmentation". A high-precision model is constructed by combining Delaunay triangulation. The results show that this scheme can effectively overcome downhole disturbance, and the absolute errors of key dimensions (width, height, support spacing) of the model are all ≤ ±2 cm, accurately restoring the roadway topology and support distribution. The model can link personnel positioning and deformation monitoring data, and provide visualized support for coal mine safety management, intelligent mining, and training, which has significant engineering practical value.