OPR-SLAM: A Semantic SLAM with Occluded Point Recovery for Dynamic Environments

As a fundamental technology in robotics, Simultaneous Localization and Mapping (SLAM) has been widely researched in recent years. However, most of existing SLAM systems assume the word is static. This assumption greatly limits the applications of them in real-word environment. In this paper, we present a semantic visual SLAM based on ORB-SLAM2, named as OPR-SLAM, that integrates the capabilities of dynamic object detection and occluded point recovery. First, a neural network is employed to obtain the pixel-level segmentation of objects. Then, dynamic objects are identified according to geometric information. The feature points belong to dynamic objects and the points with large reprojection error are filtered out. Finally, camera motion is estimated by static feature points. Furthermore, the background feature points occluded by dynamic objects are recovered to improve the performances of relocalization and loop closing. The experimental results on the TUM dataset demonstrate that OPR-SLAM has a competitive performance on loop detection, and outperforms other state-of-the-art SLAM systems in terms of tracking accuracy in dynamic environments.