Answers are grounded in this paper and cite their sources.
This study addresses the critical challenge of balancing robust feature representation and computational efficiency in loop closure detection for visual simultaneous localization and mapping (SLAM) systems under sudden illumination variations, as well as the issue of temporal misalignment between image sequence matching and real-world scenes. To tackle these problems, we propose an illumination-adaptive, hierarchical, and lightweight framework for robust loop closure detection in Visual SLAM. Its core architecture comprises three synergistic modules: a Swin-Tiny Transformer-based feature extraction module, which enhances illumination robustness by leveraging mid-level structural features integrated with sliding window self-attention; a dimension-reduced network-based vector locally aggregated descriptor generation module, designed to optimize both matching speed and storage overhead while preserving discriminative power; and a motion-constrained Fast Dynamic Time Warping (FastDTW) sequence matching module that improves temporal alignment between image sequences and physical scenes by incorporating robot motion parameters. Extensive evaluations on the New College and City Centre datasets–characterized by significant illumination variations–demonstrate that the proposed framework achieves precision rates of 80.6% and 78.3% at 50% recall, respectively, with a per-frame processing delay of only 18 ms. Compared with Oriented FAST and Rotated BRIEF (ORB)-SLAM3, our framework yields a 22% improvement in precision, validating its superiority in balancing robustness, efficiency, and lightweight performance for complex illumination scenarios.
No takes yet. Share an insight, caveat, or question.
Xian et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: