RPSD-SLAM:Robust RGB-D SLAM Method Integrating Point-Line Information with Semantic Data in Dynamic Environments

Current neural network-based SLAM methods typically require high computational resources to pre-define dynamic objects, and often face issues such as insufficient reliable points remaining after removing dynamic objects and underutilization of environmental information. In this paper, we propose a novel SLAM system that integrates semantic information with point-line fusion which is build on ORB-SLAM2. We use semantic information and feature point states to identify dynamic regions, and utilize depth information to distinguish foreground regions. Further, we implement grid-based feature point re-extraction in regions with rich texture information and extract line features from the scene. By combining scene point-line constraints, we eliminated the negative impact of dynamic features on the SLAM system.