A Monocular Dynamic SLAM Algorithm Based on Deep Learning

In the field of visual Simultaneous Localization and Mapping (SLAM), dynamic environment poses a significant challenge to the accuracy and robustness of systems, especially for monocular visual systems. Existing monocular SLAM algorithms tend to fail in tracking when dealing with dense or large-scale dynamic objects. To alliviate this problem, a real-time dynamic molocular SLAM algorithm based on SuperGlue and YOLO, named SuperGlue-YOLO-Dynamic SLAM(SYD-SLAM), is proposed. Initially, SYD-SLAM incorporates a novel weighted reprojection error during the monocular initialization, leveraging the strengths of SuperGlue. To this end, the YOLO algorithm was introduced during the uniform velocity model tracking phase to filter out dynamic points. Finally, a local map tracking algorithm is proposed by combining SuperGlue's descriptor and multi view geometry. Except for its fast monocular initialization speed, extensive experiments have shown the robustness and the real-time performance of SYD-SLAM in dynamic environments.