To address the limitations of the static assumption in visual SLAM for dynamic real-world applications,a visual SLAM algorithm is proposed, which is based on Fourier-Mellin transformfor high dynamic environments. It involves Fourier-Mellin transform for motion compensation, employs frame differencing for motion mask generation, utilizes the short-term dense connection network for semantic segmentation to identify potential moving objects,combines motion and object masks to obtain the final object motion region, and eliminates the corresponding feature points in that region. Finally, the pose accuracy is optimized based on stable static feature points. Experimental results demonstrate a reduction of over 95% in absolute trajectory error and relative pose error compared to ORB-SLAM2, and over 30% compared to DS-SLAM. These evaluate its excellent localization accuracy and robustness in complex dynamic scenes. The impact of motion blur and lighting changes on motion detection is effectively mitigated,and the limitations of traditional dynamic SLAM in detecting non-prior motion objects are overcome.