A Robust and Efficient Visual-Inertial SLAM for Vision-Degraded Environments

Visual challenging environments significantly threaten both the robustness and accuracy of SLAM (Simultaneous Localization and Mapping), resulting in unacceptable performance of unmanned systems. Many researchers propose various methods to overcome visual challenges, and we briefly categorize them into pure-vision-based and sensor fusion-based. The former typically has a better generalization to different vision cameras while accompanied by a performance ceiling due to single modal data. The latter generally achieves better robustness and accuracy with sophisticated sensors and hard-ware, which may be unaffordable on lightweight platforms. Most current SLAM methods cannot switch these two technical routes, resulting in unsatisfying performance. Focusing on this problem and dilemma, we propose a lightweight visual-inertial SLAM with a dynamic switching mechanism according to ambient perception. The system automatically switches to visual pose estimation, inertial propagation, and motion model algorithms according to the ambient perception quality. It effectively avoids redundant computations in good conditions and robustly keeps SLAM tracking in hard conditions. Extensive experiments show the feasibility and effectiveness of our method, with better performance compared with other state-of-the-art methods.