Abstract: To address issues of precision degradation, and potential failure of LiDAR SLAM (simultaneous localization and mapping) in unstructured environments with limited satellite visibility, a robust LiDAR SLAM algorithm suitable for many types of LiDAR is proposed for ground robots. Firstly, the proposed non-hyperparameter ground initialization method is utilized to automatically segment ground points, while the geometric and intensity features of the remaining point clouds are extracted through PCA (principal component analysis). The extracted ground and feature points are fed into the LiDAR odometry module for three-stage direct feature matching, resulting in a high-frequency and relatively accurate inter-frame transformation estimation. In the LiDAR mapping module, the LiDAR inter-frame matching factor is constructed by intensity-enhanced geometric features registration, and ground points extracted from the front end are used to evaluate ground flatness and continuity for the construction of a ground constraint factor. Additionally, the IMU (inertial measurement unit) rotation and gravity constraint factors are constructed using the IMU pre-integration. Finally, the factor graph is constructed combining the above factors and the loop detection factor to conduct global optimization, obtaining an accurate robot state estimation and a global map in real time. Comparison with various methods in several challenging sequences collected by a ground robot demonstrates that the proposed algorithm achieves more accurate state estimation, and shows greater robustness in terrain-changed, large-scale and unstructured scenes.