Aiming at the problems of feature point matching accuracy degradation and map construction error increase caused by dynamic interference in visual SLAM in complex dynamic scenes, a dynamic visual SLAM algorithm combining semantic segmentation and sparse optical flow is proposed. Firstly, an adaptive thresholding strategy is introduced to effectively improve the algorithm′s ability to acquire feature points in complex environments. Secondly, the DY-Conv module is embedded into the U-Net semantic segmentation network and combined with the LK sparse optical flow field to achieve accurate detection and segmentation of dynamic objects, which effectively improves the feature matching accuracy and robustness of visual SLAM in dynamic scenes. Finally, the validity of the algorithm is verified based on the TUM dataset and real scenes. Experimental results show that the improved U-Net algorithm increases the average segmentation accuracy from 92.1% of the original algorithm to 94.5%. Meanwhile, the proposed semantic visual SLAM algorithm improves image processing speed by 60.13% compared to ORB-SLAM3, and enhances pose estimation accuracy by 43.75%, 77.33% and 64.00% on three high-dynamic sequence public datasets, respectively. Additionally, the dense 3D point cloud maps generated based on the TUM dataset and real-world scenarios further demonstrate that the proposed algorithm can effectively suppress the interference of dynamic factors, thereby improving the accuracy of map construction.