This study aims to develop a practical LiDAR–inertial Simultaneous Localization and Mapping (SLAM) system for industrial mobile robots operating in dynamic, repetitive and resource-constrained warehouse environments. Rather than treating dynamic object handling as isolated point removal, the proposed system formulates it as dynamic uncertainty propagation across perception, state estimation and back-end optimization.
The method integrates a point-cloud dynamic mask for feature-level hard rejection, temporal 3D-Intersection over Union (IoU) covariance modulation for uncertainty-aware LiDAR updates and a PSA-ICP two-stage loop-registration module within multifactor graph optimization. Dynamic information is first used to reject high-risk point-cloud features, then encoded in the Iterated Error‑state Kalman Filter (IESKF) measurement covariance and finally propagated to graph optimization through covariance-aware odometry factors and dynamically filtered loop-registration inputs.
With the dynamic threshold set to t = 0.5, the proposed dynamic-object evaluation achieves an F1 score of 0.886. On hall_02, street_04 and street_08, the proposed method obtains absolute pose error root‑mean‑square error (RMSE) values of 0.145 , 0.376 and 0.163 m, respectively, reducing RMSE by 8.23%, 7.16% and 7.39% compared with the strongest dynamic baseline on each sequence. Additional self-collected warehouse experiments and Jetson Nano profiling further demonstrate improved map consistency and embedded feasibility.
Unlike conventional dynamic SLAM approaches that treat dynamic-object removal as an isolated preprocessing operation, the proposed method formulates dynamic-object handling as a continuous uncertainty-propagation process spanning front-end perception, state estimation and graph optimization.
