LiDAR Simultaneous Localization and Mapping (SLAM) systems face critical challenges in complex environments, where interference from moving objects leads to point cloud registration failure, localization drift, and trajectory loss. Rather than treating semantic segmentation, LiDAR–inertial odometry, and loop closure as independent modules, the proposed Semantic-LIO establishes a semantic reliability pipeline across the whole SLAM system. In the front end, semantic labels are used before feature extraction to remove dynamic-class points and prevent them from contaminating scan-to-submap registration. In the back end, static semantic instances are organized into triangle-topology descriptors for loop-candidate retrieval, and the semantic confidence is further propagated into the information matrix of loop-closure factors. Compared with methods that exploit semantics mainly in isolated registration, map labeling, or loop-retrieval modules, the main distinction of Semantic-LIO lies in the coupled use of semantic information for point filtering, topological loop retrieval, and confidence-aware graph optimization. Experimental results on the KITTI odometry dataset and a self-collected dynamic campus dataset demonstrate that the proposed method achieves a mean APE RMSE of 0.77 m. The self-collected dynamic campus dataset records the lowest root-mean-square error of 1.23 m and relative pose error (RPE) of 1.11 m while maintaining real-time operation above 10 Hz.