FusionLVIR: a multi-sensor mobile mapping framework with global anchoring and confidence-aware map maintenance for degenerate built environments

Accurate long-range simultaneous localization and mapping (SLAM) in degenerate built environments such as tunnels, corridors, and repetitive structures remains challenging because weak observability causes drift accumulation and unreliable historical points progressively degrade map geometry. FusionLVIR, a multi-sensor mobile mapping framework built on the sequential LiDAR-camera-IMU estimator of FAST-LIVO2, is proposed to address these issues. RTK position observations are incorporated into the sequential error-state iterated Kalman filter of FAST-LIVO2 to provide intermittent global anchoring, while a depth- and confidence-aware map self-cleaning strategy selectively replaces unreliable historical points to preserve local geometric consistency. Experiments on a 3.2 km self-collected six-sequence dataset, supplemented by two public sequences for external validation of the map-maintenance module, yielded a mean leave-one-out aligned control-point RMSE of 0.031 m across Seq01–Seq04 and a mean loop-closure error of 0.026 m across four loop sequences. Controlled ablations showed that RTK global anchoring improves long-range trajectory consistency, while confidence-aware map self-cleaning reduces indoor loop residuals and map thickening and improves agreement with independently measured structural dimensions. Sensitivity analysis supported the selected thresholds, while both public sequences showed consistent reductions in local map-maintenance indicators, further supporting the applicability of the proposed strategy to long-range spatial data acquisition and automated mapping in degenerate underground and built environments.