Point level confidence-driven optimization for LiDAR-Inertial SLAM mapping in dynamic environments
TL;DR
This work proposes a novel method for constructing point-wise observation confidence by integrating geometric consistency, free-space reasoning, and temporal stability, which retains the observability of geometric constraints while effectively mitigating the impact of dynamic interference, thereby enhancing mapping accuracy in dynamic scenarios.
Abstract
Laser-inertial SLAM systems operating in dynamic environments still face significant challenges related to localization drift. Conventional approaches rely on semantic segmentation or hard-thresholding techniques. However, semantic methods demand substantial computational resources, limiting their deployment in cost-sensitive or real-time applications. Hard-thresholding strategies may eliminate valid observations in sparse structural features or limited evidence, undermining system observability. Therefore, this work proposes a novel method for constructing point-wise observation confidence by integrating geometric consistency, free-space reasoning, and temporal stability. First, the system derives confidence metrics that are consistently applied throughout front-end registration and back-end factor graph optimization. Then, a confidence gating mechanism combined with a confirmation-counting strategy is introduced to suppress the entrenchment and propagation of dynamic artifacts. Finally, the proposed approach is validated through experiments on M2DGR and KITTI datasets. Results demonstrate that, on the Street_08 sequence, the proposed method reduces APE-RMSE by 70.8%, 63.4%, and 26.8% compared with LeGO-LOAM, LIO-SAM, and Removert, respectively. Crucially, the system retains the observability of geometric constraints while effectively mitigating the impact of dynamic interference, thereby enhancing mapping accuracy in dynamic scenarios.