Abstract: Simultaneous localization and mapping (SLAM) is a critical technology for robot autonomous navigation that enables robots to navigate unknown environments by constructing maps while locating their positions. However, most existing SLAM algorithms only perform well in static environments because dynamic objects, such as vehicles and people, introduce dynamic points into the LiDAR points cloud that degrade the accuracy of points cloud registration and lead to cumulative errors in localization and mapping. To address these issues concerning SLAM in dynamic environments, this study proposes a novel LiDAR SLAM algorithm that integrates dynamic points removal and enhance loop closure detection. The proposed algorithm overcomes the critical challenge associated with SLAM in dynamic environments, that is, precise separation of ground and dynamic points, through a three-step ground segmentation process that minimizes the false removal of static objects near the ground. The innovative features of the proposed algorithm include the segmentation of ground structures, the clustering of dynamic objects at the front-end, and the incorporation of optimization factor graphs into loop closure detection at the back-end. At the front-end, a three-step ground point segmentation method is used to reduce point cloud registration errors caused by dynamic points. Firstly, a coarse ground extraction is performed using height-based filtering and voxel grid analysis to correct point cloud distortion due to sensor installation, miscalibration, or motion chattering. Secondly, a refined ground plane fitting is achieved using the random sample consensus (RANSAC) algorithm, which iteratively optimizes the ground model by evaluating inlier points. Thirdly, non-ground points are processed using the growth clustering method in the height threshold seed selection region to identify and remove dynamic objects, such as vehicles and people. The above steps mean that dynamic points can be removed during the feature points extraction period and points cloud registration. These improvements significantly increase the robustness of LiDAR odometry in dynamic environments. At the back-end, a scan-context-based geometric descriptor is employed to enhance the environments representation accuracy by encoding multi-layer height differentials in polar coordinates. The subsequent projection of a keyframe points cloud into a 2D (Two-dimensional) polar grid achieves rotation-invariant feature encoding with height variation quantization. Furthermore, a simulative lateral translation is introduced to improve descriptor sensitivity under lane-changing environments, which means that the detection loop closure candidates can be identified by calculating the cosine similarities between descriptors. This overcomes the accumulated drift in traditional spatial-relationship-based methods and enables efficient and accurate detection, even in repetitive or evolving environments. Experimental validation using the M2DGR street_08 sequence and KITTI 04 sequence demonstrated the superiority of the proposed method. Compared to other state-of-the-art approaches, such as LeGO-LOAM, LIO-SAM, and Removert, this method achieved maximum root mean square error (RMSE) reductions of 29.8% compared to the M2DGR street_08 sequence and maximum RMSE reductions of 42.7% compared to the KITTI 04 sequence. These results confirmed that the proposed method effectively enhanced the global consistency and localization precision of LiDAR SLAM in dynamic environments.