Enhanced visual-inertial SLAM Using SuperPoint and semantic geometric dynamic feature detection

Visual-inertial simultaneous localization and mapping (VI-SLAM) is fundamental for unmanned driving and VR. However, traditional feature-based SLAM systems rely on hand-crafted features and lack dedicated methods for handling dynamic objects, which results in degraded performance under challenging conditions, such as violent motion, varying illumination, and dynamic environments. To handle the issues, we propose SuperDynaSLAM, an enhanced VI-SLAM that integrates SuperPoint which is a deep learning-based feature extractor with a two-stage dynamic feature point detection method. By replacing the traditional ORB extractor with SuperPoint, SuperDynaSLAM can extract more robust feature points under challenging conditions. Furthermore, by fusing semantic information and geometric constraints, SuperDynaSLAM can accurately detect moving objects and remove associated dynamic feature points. Experiments conducted on multiple datasets demonstrate that SuperDynaSLAM achieves more competitive performance compared with ORB-SLAM3 and other SLAM systems.