ICF-SLAM: Instance-guided loop closure and asynchronous historical correction for filter-based LiDAR–inertial SLAM

Abstract

Simultaneous localisation and mapping (SLAM) is a key technology for mobile robots and autonomous systems to obtain real-time pose information and environmental spatial structure, and its measurement accuracy directly affects localisation reliability and map consistency in large-scale scenarios. However, in environments such as public buildings and long-range inspection tasks, where repetitive structures and long-range revisits are common, filter-based LiDAR–inertial SLAM systems often struggle to generate reliable loop constraints and to consistently correct historical trajectories after loop closure. To address this issue, this paper presents an asynchronous bypass framework that augments a filter-based LiDAR–inertial front end with instance-guided loop closure and historical correction. The visual instance front end generates masks for instance targets that are static, stable, and repeatedly observable in the scene, and projects these masks onto the LiDAR scans to form compact instance sub-clouds, thereby providing spatial priors for loop-candidate association and coarse-to-fine registration initialisation rather than relying on any fixed semantic category. Once a loop is accepted, the correction is propagated through the historical frames: near-end frames are re-estimated from an inertial measurement unit increment prior, and far-end frames are propagated by time-based pose interpolation. Experiments on eight self-collected and public sequences demonstrate that ICF-SLAM achieves reliable loop detection, improves global localisation accuracy and historical-map consistency, and preserves real-time front-end operation. The proposed method achieves a macro-averaged maximum F1 score of 0.998 and reduces the macro-averaged ATE translation RMSE by 38.6% relative to FAST-LIO2 and by 32.4% relative to the loop-only variant.