Simultaneous localization and mapping (SLAM) is the key to autonomous robot navigation. However, traditional SLAM systems are typically designed for static environments, when dynamic objects are present, dynamic feature points can lead to incorrect data associations, reducing accuracy and reliability. Existing solutions still face challenges such as undetected potentially dynamic objects and an insufficient number of useful feature points when dynamic objects dominate the scene. To overcome these limitations, this study proposes a vision SLAM system based on ORB-SLAM2. Firstly, yolov8 object detection is utilized to provide semantic information, which is combined with depth information for depth constraints to generate dynamic masks; next, a quadtree-based uniform allocation of feature points is implemented based on dynamic probability, ensuring the removal of dynamic feature points while preserving more useful features; finally, optical flow tracking is utilized to detect and reject feature points on potentially dynamic objects. In which the dynamic mask is combined with keyframes to realize motion segmentation, thus constructing clean and dense point cloud maps. Experimental results on the TUM and Bonn datasets demonstrate that, compared to ORB-SLAM2, the average localization accuracy improves by over 90% in highly dynamic scenes while maintaining reliable performance in relatively static environments. Additionally, the improved system achieves real-time performance and outperforms other state-of-the-art methods in its category.