Robust RGB-D SLAM in dynamic environment using faster R-CNN

Recent approaches for simultaneous localization and mapping (SLAM) have shown great success for static environment reconstruction. However, these methods suffer from performance degeneration in dynamic environment due to the unstable and ambiguous data associations. In this paper, we propose a robust and efficient method for SLAM in dynamic environment. To eliminate the influence of dynamic objects, we compute the difference between the current frame and keyframe by semantic and geometric information to detect the dynamic object. Then we pick up the reliable data association and get more accurate transform matrix. The experiments on the public dataset demonstrate that our proposed method outperforms the original method in dynamic environment.