Abstract: We introduce a dynamic simultaneous localization and mapping (SLAM) algorithm, combining deep learning-based semantic segmentation with visual SLAM. As a conclusion, this algorithm utilizes voting based on three categories of objects and semantic loop closures to effectively mitigate the impact of dynamic objects on SLAM system performance, while enhancing localization and mapping accuracy. Firstly, the semantic objects are classified into three categories: static, potentially dynamic, and certainly dynamic. Then voting method is employed based on reprojection depth error to identify the motion states of these semantic objects, thereby negating the influence of moving targets.Additionally, we employ a semantic similarity loop closure optimization method to enhance loop closure detection robustness.Experimental results on the TUM RGB-D dynamic dataset and the KITTI dataset demonstrate that our algorithm reduces the average absolute trajectory error by 57. 13% and 23. 39% compared to the ORB-SLAM3 algorithm, respectively, confirming its robustness in dynamic scenes.