Dynamic Objects Recognizing and Masking for RGB-D SLAM

Abstract

Simultaneous Location and Mapping (SLAM) has been applied widely in robots and computer field. However, most traditional SLAM systems based on static scene assumption cannot meet the accuracy requirements in dynamic environments. In order to improve the positioning accuracy of the robot in the dynamic environment, this paper proposes a stable RGB-D SLAM approach based on ORB-SLAM3. First, the camera frame is divided into static area, potential dynamic area, and priori dynamic area. Then the mask of the priori dynamic area generated by pixel-level semantic segmentation is used to obtain an approximately accurate camera initial pose, and the multi-view geometry technology is combined to identify the potential dynamic area. Finally, the features in static area are used to complete the tracking trajectory and camera pose. Experiments on the public dataset TUM demonstrate that the proposed method in dynamic environment have better performance than ORB-SLAM3.