Abstract: Simultaneous Localization and Mapping (SLAM) is one of the most attractive research directions in the field of mobile robots. SLAM is able to construct maps and provide localization information for the robots in unknown environments. Most existing SLAM algorithms perform well in static environments. However, regarding the environment with moving objects such as vehicles and pedestrians, there exist dynamic points in the laser point cloud, leading to the problem that the registration accuracy of point clouds can be adversely affected among different reference frames. To address this issue, this paper improves the front-end feature extraction and back-end loop detection modules of SLAM via using the geometry information of laser point cloud. By removing the dynamic points, the precision of mobile robot positioning and mapping is enhanced in dynamic environment. First, in order to improve the accuracy of front-end feature extraction, a stepwise ground segmentation method is proposed, which performs ground point extraction based on point cloud height information. Next, the Random Sample Consensus method is applied to elaborate ground segmentation on the corrected point cloud. Subsequently, a seed-growing clustering method is utilized to extract non-ground dynamic points based on height thresholds, followed by feature extraction and registration. Finally, for the back-end loop detection module, the traditional spatial relationship-based loop detection method is replaced with point cloud descriptors, so as to reduce the cumulative error and improve the sensitivity of loop detection. The effectiveness of the proposed method is validated through visualization simulation and accuracy assessment.