With the outstanding performance of deep learning technology in semantic segmentation, object detection and other aspects, researchers have gradually tried to apply it to the field of three-dimensional scene semantic segmentation and extend it to LiDAR point cloud data, hoping to improve the automatic and real-time processing level of 3D scene semantic segmentation, and the combination of deep learning methods can effectively improve the system processing efficiency and improve the final SLAM construction accuracy.especially semantic segmentation, which is one of the main directions of deep learning to improve the LiDAR SLAM system.At present, the extraction of ground point cloud by traditional liDAR SLAM algorithm is still a simple geometric method, and the extraction accuracy of ground point cloud is low.To solve the problem of low accuracy of ground point cloud extraction, this paper proposes a method of LiDAR SLAM point cloud segmentation extraction algorithm based on LEGO-LOAM algorithm and lightweight point cloud network RangeNet++,which implements the effective classification of LiDAR ground point clouds, and then improves the extraction quality of feature point clouds and the mapping accuracy of liDAR SLAM.