REA-LVIO: An Asynchronous LiDAR-Visual-Inertial Odometry Based on Multistate Constraint Kalman Filter

Light detection and ranging (LiDAR) provides accurate 3-D structure perception but degrades severely in geometrically featureless environments, while visual sensors offer rich texture information but are highly susceptible to illumination changes. To leverage their complementary strengths, in this article, we present REA-LVIO, a LiDAR-visual-inertial odometry system that asynchronously updates the state using LiDAR and monocular visual measurements within the multistate constraint Kalman filter (MSCKF) framework to ensure sensor-degradation robustness, efficiency, and accuracy. We introduce a tracking-stable and inactive feature joint update strategy. Specifically, farthest point sampling is utilized to select a subset of tracking-stable features with minimal correlation, which are then combined with inactive features for the visual update. A dual-layer augmented pose selection criterion is employed to ensure proper marginalization in open environments by leveraging both ego-motion and parallax. We also design a synchronous update strategy and compare it with the asynchronous strategy. To address the singularity issue of the covariance matrix in MSCKF, we derive a numerically stable and efficient Kalman gain formula utilizing QR decomposition. Experiments demonstrate that REA-LVIO achieves competitive accuracy and lower computational overhead than state-of-the-art LiDAR-visual systems, while ensuring continuous state estimation in complex environments.