GCENet: A geometric correspondence estimation network for tracking and loop detection in visual–inertial SLAM

Establishing robust and effective data correlation has been one of the core problems in visual based SLAM (Simultaneous Localization and Mapping). In this paper, we propose a geometric correspondence estimation network, GCENet, tailored for visual tracking and loop detection in visual–inertial SLAM. GCENet considers both local and global correlation in frames, enabling deep feature matching in scenarios involving noticeable displacement. Building upon this, we introduce a tightly-coupled visual–inertial state estimation system. To address challenges in extreme environments, such as strong illumination and weak texture, where manual feature matching tends to fail, a compensatory deep optical flow tracker is incorporated into our system. In such cases, our approach utilizes GCENet for dense optical flow tracking, replacing manual pipelines to conduct visual tracking. Furthermore, a deep loop detector based on GCENet is constructed, which utilizes estimated flow to represent scene similarity. Spatial consistency discrimination on candidate loops is conducted with GCENet to establish long-term data association, effectively suppressing false negatives and false positives in loop closure. Dedicated experiments are conducted in EuRoC drone, TUM-4Seasons and private robot datasets to evaluate the proposed method. The results demonstrate that our system exhibits superior robustness and accuracy in extreme environments compared to the state-of-the-art methods.