SSF-SLAM: Real-Time RGB-D Visual SLAM for Complex Dynamic Environments Based on Semantic and Scene Flow Geometric Information

Most existing SLAM algorithms rely on static world assumptions and perform poorly in complex dynamic environments. In order to improve the accuracy and robustness of SLAM in complex dynamic environments, based on ORB-SLAM2, this paper proposes a SLAM system that combines semantic information and scene flow geometry information (SSF-SLAM). Semantic information is the core of robot scene understanding and cognition. Firstly, a lightweight object detection module is constructed and the acquired semantic information is innovatively coupled with multi-view geometry to achieve rapid and accurate dynamic object recognition. Then, a novel clustering module of scene flow geometry information based on depth and density is designed, which can effectively reduce the limitation of geometric constraints and realize fast and accurate calculation of geometric dynamic regions. In addition, a semantic mapping module is also built to generate 3D point clouds and 3D semantic objects to help mobile robots understand scenes in actual tasks. In SSF-SLAM, the object detecting module and semantic mapping module are integrated into a single thread and run in parallel to ensure the real-time performance of the system. Finally, the method was tested on various public data sets and real-world environment, the results showed that compared with other advanced methods, SSF-SLAM performed better in terms of timeliness, accuracy, and robustness.