SGS-SLAM is a semantic visual SLAM system based on Gaussian Splatting, which integrates appearance, geometry, and semantic features through multi-channel optimization. It addresses the oversmoothing limitations of neural implicit SLAM systems in high-quality rendering, scene understanding, and object-level geometry. The system introduces a unique semantic feature loss that compensates for the shortcomings of traditional depth and color losses in object optimization. A semantic-guided keyframe selection strategy prevents erroneous reconstructions caused by cumulative errors. Extensive experiments show that SGS-SLAM achieves state-of-the-art performance in camera pose estimation, map reconstruction, precise semantic segmentation, and object-level geometric accuracy, while ensuring real-time rendering capabilities. SGS-SLAM uses a 3D Gaussian representation for scene modeling, enabling fast and real-time camera tracking and scene mapping. It leverages 2D semantic maps to learn 3D semantic representations expressed by Gaussians, providing high-fidelity reconstruction and optimal segmentation precision. Semantic maps provide additional supervision for parameter optimization and keyframe selection. The system employs a multi-channel parameter optimization strategy where appearance, geometric, and semantic signals contribute to camera tracking and scene reconstruction. SGS-SLAM uses these diverse channels for keyframe selection during tracking, focusing on actively recognizing objects seen earlier in the trajectory. SGS-SLAM provides a highly accurate disentangled object representation in 3D scenes, laying a solid foundation for downstream tasks such as scene editing and manipulation. It enables dynamic moving, rotating, or removal of objects in the map in real time by grouping Gaussians based on semantic labels. The system outperforms neural implicit semantic SLAM systems in terms of rendering speed, reconstruction quality, and segmentation accuracy. It enables precise editing and manipulation of specific scene elements while preserving the high fidelity of the overall rendering. The integration of semantic representation allows for more accurate scene interpretation and object-level geometry. SGS-SLAM is evaluated on both synthetic and real-world datasets, demonstrating superior performance in mapping, tracking, and semantic segmentation. It achieves state-of-the-art results in reconstruction quality, depth L1 loss, ATE error, and mIoU scores. The system is efficient and effective in both mapping and segmentation processes, with a focus on real-time performance and high-quality rendering. The method is suitable for downstream tasks such as scene editing and manipulation, offering solid prior for robotics or mixed reality applications. Limitations include reliance on depth and 2D semantic signal inputs, and large memory consumption for large scenes. Future research aims to address these limitations.SGS-SLAM is a semantic visual SLAM system based on Gaussian Splatting, which integrates appearance, geometry, and semantic features through multi-channel optimization. It addresses the oversmoothing limitations of neural implicit SLAM systems in high-quality rendering, scene understanding, and object-level geometry. The system introduces a unique semantic feature loss that compensates for the shortcomings of traditional depth and color losses in object optimization. A semantic-guided keyframe selection strategy prevents erroneous reconstructions caused by cumulative errors. Extensive experiments show that SGS-SLAM achieves state-of-the-art performance in camera pose estimation, map reconstruction, precise semantic segmentation, and object-level geometric accuracy, while ensuring real-time rendering capabilities. SGS-SLAM uses a 3D Gaussian representation for scene modeling, enabling fast and real-time camera tracking and scene mapping. It leverages 2D semantic maps to learn 3D semantic representations expressed by Gaussians, providing high-fidelity reconstruction and optimal segmentation precision. Semantic maps provide additional supervision for parameter optimization and keyframe selection. The system employs a multi-channel parameter optimization strategy where appearance, geometric, and semantic signals contribute to camera tracking and scene reconstruction. SGS-SLAM uses these diverse channels for keyframe selection during tracking, focusing on actively recognizing objects seen earlier in the trajectory. SGS-SLAM provides a highly accurate disentangled object representation in 3D scenes, laying a solid foundation for downstream tasks such as scene editing and manipulation. It enables dynamic moving, rotating, or removal of objects in the map in real time by grouping Gaussians based on semantic labels. The system outperforms neural implicit semantic SLAM systems in terms of rendering speed, reconstruction quality, and segmentation accuracy. It enables precise editing and manipulation of specific scene elements while preserving the high fidelity of the overall rendering. The integration of semantic representation allows for more accurate scene interpretation and object-level geometry. SGS-SLAM is evaluated on both synthetic and real-world datasets, demonstrating superior performance in mapping, tracking, and semantic segmentation. It achieves state-of-the-art results in reconstruction quality, depth L1 loss, ATE error, and mIoU scores. The system is efficient and effective in both mapping and segmentation processes, with a focus on real-time performance and high-quality rendering. The method is suitable for downstream tasks such as scene editing and manipulation, offering solid prior for robotics or mixed reality applications. Limitations include reliance on depth and 2D semantic signal inputs, and large memory consumption for large scenes. Future research aims to address these limitations.