DRSO-SLAM: A Dynamic RGB-D SLAM Algorithm for Indoor Dynamic Scenes

Abstract

In order to improve low position accuracy and insufficient robust performance of service robots in indoor dynamic scenes, a dynamic RGB-D simultaneous localization and mapping (SLAM) algorithm based on semantic information and optical flow (DRSO-SLAM) is proposed. Firstly, the Mask R-CNN semantic segmentation network is used to obtain semantic information in indoor dynamic scenes, establishing a priori semantic information database of objects and the system's rough self-motion estimation is performed based on the established prior semantic information database. Secondly, the feature points optical flow field is calculated to track the dynamic objects optical flow, the pixel-level semantic information obtained through the semantic segmentation network serves as mask information in the optical flow tracking and the fundamental matrix is further calculated. Finally, the epipolar geometry is used to filter out the actual dynamic feature points and only static feature points are retained in the tracking, local mapping and loop detection threads. Experiments are carried out in high and low dynamic scenes of the TUM dataset. Compared with the ORB-SLAM2, the root mean square error of DRSO-SLAM in high dynamic scenes is improved by an average of 95.02% and it also has considerable positioning accuracy in low dynamic scenes. Experimental results have shown that DRSO-SLAM can effectively improve the position accuracy and robustness of SLAM system in dynamic indoor scenes.