RGB-D SLAM in dynamic environments with deep learning

Abstract

Traditional visual Simultaneous Localization and Mapping (SLAM) is mostly based on the assumption of static environment, which is susceptible to receive dynamic targets in dynamic environment, leading to the degradation of localization accuracy. In this paper, we introduce the instance segmentation network SOLOv2, which combined with motion consistency detection can effectively eliminate the dynamic feature points in the environment and improve the visual SLAM accuracy with the depth map hole repair algorithm. Tested on the TUM dataset, the positional estimation accuracy in dynamic environments is significantly improved compared to ORB-SLAM2.