To tackle the challenges of adaptability and precision in VSLAM for dynamic environments, we propose a joint refined semantic-geometric approach that improves SLAM’s performance across various dynamic settings. Our method integrates semantic segmentation networks with morphological processing to extract stable boundary features from potential dynamic objects accurately. By restoring depth information and applying geometric constraints that account for camera motion, we facilitate the precise identification and removal of dynamic objects. Additionally, we exploit static scene information to inpaint the background areas occluded by dynamic objects, thus enabling complete scene reconstruction. Quantitative evaluation using dynamic sequences from the TUM dataset reveals a significant reduction in RMSE for both high and low dynamic sequences compared to ORB-SLAM2, DynaSLAM, Yolo-SLAM and Blitz-SLAM. Specifically, there is 95.23%–32.30% decrease in RMSE for high dynamic sequences and 44.7%–5.52% decrease for low dynamic sequences, respectively. These results demonstrate the method’s enhanced adaptability and localization accuracy across different levels of dynamic scenes. Furthermore, the dense reconstruction maps derived from the Static Background Inpainting process offer more complete static scene information than original maps, providing adequate technical support for autonomous localization and mapping of robots in dynamic environments.