GAN-SLAM: A Visual Sparse SLAM System with Flow-Guided Inpainting for Robust Mapping in Dynamic Environments

In this paper, a dynamic visual Simultaneous Localization and Mapping (SLAM) system is proposed to enhance the accuracy and robustness of feature-based SLAM systems in highly dynamic environments. The proposed system goes beyond the assumption of static scenes prevalent in state-of-the-art SLAM systems and provides more information to the SLAM back-end. The system is capable of retrieving occluded dynamic object backgrounds with complex camera motions by incorporating flow-guided video inpainting techniques, which estimate background information from adjacent frames. Additionally, a parallel mechanism is proposed to coordinate the video inpainting module with the segmentation module, enabling an online solution for dynamic visual SLAM. The experimental results on TUM dataset demonstrate the significance of retrieving object backgrounds in dynamic environments.