STAC-LIO: Robust LiDAR-Inertial Odometry in dynamic environments via spatiotemporal consistency perception and adaptive observation noise covariance

Abstract Aiming at the challenge of balancing suppression of dynamic interference and the preservation of effective observation constraints in complex dynamic scenes for LiDAR-Inertial Odometry (LIO), a dynamic robust LIO framework based on spatiotemporal consistency perception and adaptive observation noise covariance (STAC-LIO) is proposed. In this framework, an active collaborative mechanism for mutual feedback enhancement between dynamic perception and state estimation is featured. It can adaptively adjust observation confidence, improving environmental adaptability of LIO in complex scenes. Within this framework, the proposed dynamic point detection algorithm based on temporal occupancy and core extended neighborhood refinement-compensation mechanisms (TOP-CEN) quantifies the dynamic-static attributes of point clouds. At the state estimation layer, the dynamic-static attributes provided by TOP-CEN are introduced into the adaptive robust iterative extended Kalman filter (AR-IEKF). They guide the adaptive inflation of observation noise covariance, thus suppressing dynamic interference while retaining effective observation constraints. The posterior pose output by AR-IEKF is further fed back to TOP-CEN, providing a more reliable spatiotemporal consistency benchmark for its dynamic-static attribute discrimination. This reduces dynamic point misjudgment caused by pose estimation errors. This mutual feedback enhancement mechanism combines perception-assisted localization and localization-enhanced perception. It helps improve the completeness of dynamic point detection and the stability of state estimation. We compared TOP-CEN and STAC-LIO with 7 dynamic point removal methods and 4 LIO algorithms, respectively, on 20 sequences from 5 datasets. Compared with state-of-the-art online methods, TOP-CEN achieves an average relative improvement of 8.09% in the F1 score on the SemanticKITTI dataset. For localization accuracy, STAC-LIO achieves maximum improvements of 42.47% and 12.20% on the KITTI and UrbanLoco datasets, respectively. We also conducted real-world experiments using a mobile robot platform. Overall, STAC-LIO improves localization accuracy and robustness in complex dynamic environments while constructing static maps.