Real-time motion removal based on point correlations for RGB-D SLAM in indoor dynamic environments

The traditional visual simultaneous localization and mapping (SLAM) systems have a high dependence on static world assumption, which makes them easy to fail to track in dynamic environments. In this paper, we propose a real-time dynamic visual SLAM system (RTDSLAM) based on ORB-SLAM3 to realize accurate pose estimation of the camera in indoor dynamic environments. We regard the static objects in the environments as a complete virtual rigid body and add two motion removal modules to handle the dynamic feature points without the aid of the camera's ego motion. The geometry-based motion removal module utilizes the point correlations and the structural invariance of rigid body to detect sparse dynamic feature points between two keyframes, and the clustering of depth images helps find the complete dynamic regions. Meanwhile, the template-based motion removal module uses template matching to fast track the known moving objects between ordinary frames. The dynamic feature points located on moving objects are removed and only static feature points are reserved for pose estimation. We evaluate our system on public TUM and Bonn datasets, and the comparison with state-of-the-art dynamic visual SLAM systems shows our advantages both in runtime and the accuracy of pose estimation. Besides, the test in real-world scenes shows the effectiveness of our system in dynamic environments.