Abstract: The surface point cloud data of coal mine roadways extracted using 3D laser scanning technology has a large amount of redundant data. The existing point cloud data simplification methods have the problem of insufficient detail preservation in the processing of large-scale point clouds. In order to solve the above problems, a surface point cloud data reduction method for coal mine roadways based on secondary feature extraction is proposed. Firstly, the method performs denoising preprocessing on the collected original roadway point cloud data. Secondly, the method establishes a K-d tree and uses principal component analysis to estimate the denoised point cloud data to fit the normal vector of the neighborhood plane. Thirdly, the point cloud is preliminarily divided into feature and non-feature regions using a smaller normal vector angle threshold, retaining the feature regions and randomly downsampling the non-feature regions. Fourthly, based on the larger normal vector angle threshold, the feature region point cloud is divided into feature points and non-feature points. And voxel random sampling is conducted on the non-feature points. Finally, the method merges the two point cloud simplification results with the feature points to obtain the final simplified data. The simulation results show that under million data level point clouds and high precision conditions, this method achieves better results in feature preservation and reconstruction precision compared to curvature simplification methods, random simplification methods, and grid reduction methods. The average standard deviation calculated after 3D reconstruction can be about 30% lower than other methods under the same reduction rate.