Abstract: Visual simultaneous localization and mapping technology has become a research hotspot in the fields of mobile robots and unmanned driving. At present, most visual SLAM assumes that the surrounding environment is static, and it is easy to fail when there are moving objects in the scene. To this end, this paper eliminates moving objects in the environment based on the semantic and geometric information of visual images to improve the robustness of the system. Specifically, on the basis of ORB-SLAM2, this work implements a lightweight network to obtain semantic information through a multi-threading mechanism to eliminate known types of dynamic objects, and designs a geometric detection module tightly coupled with semantic elimination to eliminate unknown type of moving objects. In order to balance the real-time performance and the robustness of map construction, when the processing capacity of the platform is insufficient, the strategy of only removing dynamic points for key frames is adopted. The experimental results on the dynamic environment dataset of TUM RGB-D show that the positioning accuracy of the algorithm in the dynamic environment is significantly improved compared with ORB-SLAM2; compared with other dynamic SLAM, the accuracy is also improved to a certain extent.