Autonomous driving has gradually become a research hotspot in recent years, the limitations of traditional visual SLAM systems in dynamic environments are becoming increasingly prominent. To overcome this limitation, this paper propose YGM-SLAM, an enhanced SLAM framework built upon ORB-SLAM3. It integrates a lightweight YOLOv11-Seg model for real-time semantic mask generation and dynamic keypoint removal. In addition, In addition, we propose a depth-aware adaptive thresholding mechanism and integrate it into the multi-view geometry framework, enabling robust identification and removal of feature points associated with dynamic objects while maintaining reliable static observations. Evaluation on highly dynamic sequences from the TUM dataset demonstrates that YGM-SLAM enhances localization performance, reducing ATE by more than 88.24% compared to ORB-SLAM3.These results demonstrate its strong practicality and potential for deployment in real-world dynamic environments.