VSLAM faces challenges in dynamic scenes, including decreased localization accuracy and the presence of reconstruction artifacts. Although differentiable rendering-based SLAM techniques have significantly improved reconstruction quality, existing methods remain constrained by issues such as under-segmentation of semantic masks, missed detections of unknown moving objects, and unstable points in unreliable depth information and frequent occlusion areas regions. These limitations result in residual dynamic artifacts and static structure distortions in scene reconstruction. To address these challenges, this paper proposes a robust SLAM system based on 3D Gaussian Splatting. The system generates adaptive semantic masks using YOLOv8 and dynamic ratio analysis and implements a two-stage dynamic feature removal process by integrating epipolar constraints, DBSCAN clustering, and FastSAM instance segmentation. Furthermore, a planar map is constructed to provide structured scene perception, constrain pose optimization, and enhance stability in above regions. Additionally, an ellipsoid-thin-layer Gaussian initialization method is introduced, leveraging planar normal vectors to guide the Gaussian distribution and suppress distortions in above regions. Experimental results demonstrate that the proposed method effectively eliminates dynamic interference in dynamic scenes, achieving precise camera localization and high-quality dense reconstruction, with performance reaching state-of-the-art levels.