Dynamic Scene Semantics SLAM Based on Semantic Segmentation

Abstract

Simultaneous Localization and Mapping (SLAM) have become a new research hotspot in the field of artificial intelligence applications such as unmanned driving and mobile robots. Most of the current SLAM research is based on the assumption of static scenes, and dynamic objects in the indoor environment are inevitable. The assumption based on static scenes greatly limits the development of SLAM and the application of SLAM system in real life. At the same time, the semantic segmentation is added to the SLAM system to generate a semantic map with semantic information, which can enrich the understanding of the mobile carrier to the environment and obtain high-level perception. In this paper, we combine the visual SLAM system ORB-SLAM2 and PSPNet semantic segmentation network, and propose a PSPNet-SLAM system, which uses optical flow and semantic segmentation to detect and eliminate dynamic points to achieve dynamic scenes semantic SLAM. We performed experiments on the TUM RGB-D dataset. The results show that compared with other SLAM systems, PSPNet-SLAM can reduce the camera pose estimation error in indoor dynamic scenes to different degrees and improve the camera position estimation accurately.