Semi-Direct SLAM with Manhattan for Indoor Low-Texture Environment

Simultaneous Localization and Mapping (SLAM) with the incorporation of the Manhattan World (MW) assumption has been significantly discovered in recent years. While previous methods relied on the MW assumption to estimate camera rotation accurately, they faced limitations due to the requirement of a suitable planar environment. These constraints restricted the applicability of such systems. To overcome these limitations, we propose a novel approach that addresses the strict requirements of MW-based systems and significantly enhances tracking robustness in low-texture scenes. Our system leverages planar information in the environment to identify the presence of an MW scene. By decoupling the process, we achieve drift-free rotation estimation when the system detects an MW scene. Simultaneously, utilizing a semi-direct approach that combines point and line features to estimate translations in MW scenes, while performing full-camera pose estimation in non-MW scenes. Furthermore, we introduce a more precise loop closure detection strategy by exploiting the relative relationship between the Manhattan axes (MA) and line features in the scene. This strategy enhances the accuracy of identifying loop closures, which are crucial for SLAM systems. To evaluate the performance of our approach, we conducted experiments using public benchmarks. The results demonstrate improved pose estimation and loop closure performance compared to state-of-the-art methods. Overall, our proposed method alleviates the strict requirements of previous MW-based systems, enhances tracking robustness in low-texture scenes, and achieves improved performance in terms of pose estimation and loop closure detection.