An online robust loop-closure correction method for multisensor odometry

Abstract

Multisensor odometry combines LiDAR, visual and inertial measurements to provide accurate local motion estimation, but accumulated drift still degrades long-term trajectory and map consistency. Introducing an online loop-closure backend raises two practical concerns. Incorrect constraints may corrupt global optimization, while direct application of backend corrections may cause discontinuities in the published pose. This paper presents an online robust loop-closure correction method for LiDAR–visual–inertial odometry without modifying the front-end estimator. LiDAR submaps provide the geometric evidence for loop registration, whereas the fused visual–inertial–LiDAR trajectory supports candidate retrieval, motion-consistency evaluation and online pose correction. Candidate constraints are generated by complementary global and local retrieval, verified by registration and statistical tests, and activated only after consistency with the odometric chain, agreement among induced corrections and observation support are established. The optimized correction is then released progressively in SE(3). Experiments on 28 sequences from HILTI, M2DGR, MARS-LVIG and NTU-VIRAL reduce mean APE RMSE by 13.81%, 8.72%, 8.14% and 2.09%, respectively. Ablation results show that statistical and consistency screening suppress excessive loop activations, while progressive output reduces peak single-step translation and rotation changes by 59.6% and 67.3%. The corrected pose remains available at approximately 10 Hz.