RD-LIVO: degeneracy-aware LiDAR–visual–inertial odometry with adaptive observation covariance
TL;DR
RD-LIVO is presented, a degeneracy-aware multi-sensor measurement framework that improves robustness through equivalent observation covariance regulation rather than estimator redesign and demonstrates improved absolute trajectory accuracy, reduced error dispersion, and cleaner maps in severely degenerate corridors where baselines fail or diverge, while maintaining online operation with acceptable computational overhead.
Abstract
In degenerate environments, conventional tightly-coupled LiDAR–visual–inertial odometry (LIVO) methods typically employ fixed observation noise models, which may assign unduly high weights to unreliable LiDAR and visual measurements and thereby exacerbate drift risk. This paper presents RD-LIVO, a degeneracy-aware multi-sensor measurement framework that improves robustness through equivalent observation covariance regulation rather than estimator redesign. Built upon a sparse–direct error-state iterated Kalman filter pipeline with unchanged photometric and point-to-plane residual formulations, RD-LIVO introduces a three-layer measurement credibility design for LiDAR point-to-plane constraints, sparse-direct photometric constraints, and sync-instant modality-level fusion regulation at LiDAR synchronization instants. Evaluations on public datasets against representative baselines including LVI-SAM, decoupled ablations, computational profiling, and real-robot tests with RTK ground truth demonstrate improved absolute trajectory accuracy, reduced error dispersion, and cleaner maps in severely degenerate corridors where baselines fail or diverge, while maintaining online operation with acceptable computational overhead.