MYIS-SLAM: A Manhattan World-Based RGB-D SLAM With Plane-Based Incremental Segmentation

Most simultaneous localization and mapping (SLAM) systems perform well in texture-rich environments but suffer from significant rotational drift in texture-limited scenes. Moreover, real-time performance, accuracy, and efficiency in map segmentation remain challenging. To address these problems, we propose a novel real-time RGB-D SLAM system based on the Manhattan World (MW) assumption. This system detects Manhattan frames (MF) based on lines and planes, and performs pose estimation based on the Manhattan coordinate system in MF. For non-MF, constraints are added to minimize reprojection errors for parallel and perpendicular lines and planes. The proposed method is well-suited for texture-limited environments and can effectively reduce the long-term trajectory drift caused by rotation estimation. For mapping and segmentation, we propose an improved real-time incremental plane-based segmentation method, which determines segmentation boundaries using geometric cues from depth images and subsequently refines label association through pairwise plane-label confidence. The accuracy of the proposed method is evaluated both on public datasets and in real-world scenarios, which exceeds that of the state-of-the-art methods.