With the increasing demand for autonomous operation of ground robots in complex environments, higher requirements are imposed on the accuracy and robustness of simultaneous localization and mapping (SLAM) technology. Aiming at the three core challenges of vertical drift, dynamic interference and cumulative error faced by light detection and ranging (LiDAR) SLAM systems in complex terrain environments including stairs, slopes, dynamic urban scenarios and long corridors, this paper proposes a LiDAR SLAM method for ground robots operating in complex Environments. The proposed method dynamically identifies reliable ground structures through principal normal analysis. It introduces ground constraints in flat regions to suppress vertical drift, and adaptively relaxes constraints in unstructured terrains such as stairs and slopes to mitigate the adverse impact of inappropriate constraints on pose estimation, reducing the vertical component of absolute trajectory error (ATE) by more than 85%. To tackle the interference of dynamic objects, dynamic targets are distinguished based on motion consistency across multi-frame point clouds, and sliding window optimization is employed to improve the efficiency of dynamic point removal. Meanwhile, a LinK3D descriptor based on local geometric structures is introduced to enhance the representation capability of environmental structural information. An adaptive loop closure triggering mechanism integrating fast retrieval of a bag-of-words model and precise Iterative Closest Point registration is adopted to realize efficient and accurate loop closure detection. Experiments conducted on a self-developed ground robot platform demonstrate that the proposed method outperforms mainstream baseline methods in most complex scenarios when loop closure is disabled. After enabling loop closure, its Absolute Trajectory Error is further reduced by approximately 8.7% compared with the version without loop closure. Relative to the best-performing comparative algorithm, the accuracy of the proposed method is improved by about 32.3% with loop closure activated, which fully validates its high precision and strong robustness in real-world complex environments.