DCT-DSLAM: Discrete Cosine Transform Towards the DYNAMIC SLAM

This study addresses the challenges faced by visual SLAM in dynamic environments and proposes a scheme to enhance Dyna-SLAM systems through the utilization of Discrete Cosine Transform (DCT) masks. The proposed approach tackles the issues of occlusion misidentification and unstable feature points prevalent in traditional Mask R-CNN by devising a compact vector representation based on DCT for precise prediction of dynamic regions. we have integrated the DCT encoding mechanism into the Dyna-SLAM architecture, transforming dynamic masks into low-dimensional vectors, thereby reducing computational complexity and storage requirements. Experiment results demonstrate significant performance improvements on datasets such as TUM RGB-D, showing a reduction of 34.9% in average tracking frame time and a decrease of 19.7% in root mean square error. These outcomes validate the effectiveness and practicality of our system when dealing with visual dynamic SLAM scenarios.