IBR-SLAM: visual SLAM based on improved BiSeNet with RGB-D sensor

Abstract

Abstract Visual Simultaneous Localization and Mapping (VSLAM) is the key technology of mobile robots’ localization and mapping. At present, the VSLAM system has high robustness in static environments, but it will cause feature point mapping errors in dynamic environments, which will affect the robustness of the system. To improve this situation, this study proposes a dynamic robust SLAM framework IBR-SLAM. This framework combines enhanced semantic segmentation and multimodal geometric constraints. The system acquired images by RGB-D camera, extracted semantic information of images through improved BiSeNet and used this information combined with the geometric constraints in the adaptive model to determine the dynamic region. In the dense mapping thread, the point cloud in the dynamic region is removed, so as to construct an accurate static global point cloud map. At last, the proposed system is tested on two datasets, TUM and Bonn, and compared with ORB-SLAM3, the absolute trajectory error is improved by 97.33% and 89.79% respectively. The results show that IBR-SLAM maintains high robustness in various dynamic scenarios.