Simultaneous localization and mapping (SLAM) is a technology widely employed in fields such as autonomous driving and augmented reality. Traditional visual SLAM methods assume a static environment, which compromises positioning accuracy and robustness in dynamic settings. To address this limitation, a dynamic SLAM algorithm based on instance segmentation and optical flow feature clustering is proposed. A thread for instance segmentation is introduced, utilizing YOLACT real-time instance segmentation to detect potential dynamic targets. Since directly removing the feature points of dynamic targets may result in an insufficient number of static feature points, which adversely affects pose estimation, an optical flow feature clustering algorithm is proposed to further refine the selection of static feature points. In addition, the enhanced and efficient local descriptor BEBLID is employed as a replacement for the original BRIEF feature descriptor to improve matching accuracy. The offline-trained BEBLID bag-of-words model is subsequently utilized to enable relocation and loop detection. Experiments conducted on the TUM dynamic scene dataset demonstrate the effectiveness of the proposed algorithm. When compared with ORB-SLAM2, the algorithm achieves an average reduction of 97.3% in the root mean square error of absolute trajectory error in dynamic environments, with a maximum reduction of 98.1%. Compared with other dynamic SLAM approaches, the proposed method also exhibits improved accuracy to a significant extent.