RetiGlue-SLAM: A Robust Visual SLAM System Designed for Low-Light and Complex Environments

Abstract

Visual simultaneous localization and mapping (SLAM) often degrades in low-light environments because of insufficient brightness, reduced contrast, and loss of texture, which cause sparse features, accumulated mismatches, and unstable trajectories. This paper presents RetiGlue-SLAM, a visual-inertial SLAM framework that retains the third version of ORB-SLAM (ORB-SLAM3) as the system backbone and improves the front end through task-driven coupling of three components: Retinexformer for low-light image enhancement, SuperPoint for learned local feature extraction, and LightGlue for robust cross-frame matching. The enhanced images and learned correspondences are integrated into the tracking thread, whereas the geometric back-end and inertial optimization of ORB-SLAM3 remain unchanged. Experiments on low-light sequences from EuRoC and the night-time Old Town sequence of 4Seasons show improved feature matching, tracking completion, and localization accuracy relative to the baseline. A two-stage ablation separates the learned front end from the full RetiGlue-SLAM pipeline, and a module-level runtime analysis reports a front-end latency of 41 ms per stereo frame (24.39 FPS). The results support RetiGlue-SLAM as a benchmarked low-light visual-inertial localization framework; field deployment and end-to-end real-time operation remain outside the scope of this study.