QA-MSCKF: A Statistically Adaptive Measurement Update Method for Visual-Inertial Odometry

Accurate and robust visual–inertial odometry (VIO) is essential for real-time localization on resource-constrained robotic platforms. However, in practical Multi-State Constraint Kalman Filter (MSCKF)-based VIO, fixed measurement-noise models often become inconsistent with the actual residual statistics under nonstationary sensing conditions, leading to degraded accuracy and compromised consistency. Furthermore, this mismatch can manifest across different temporal scales—where slow cross-frame statistical drift and abrupt withinframe residual bursts coexist—making single-scale covariance tuning or simple outlier rejection insufficient. To address these issues, we propose QA-MSCKF, a statistically adaptive update method that constructs an effective measurement covariance through dual-timescale adaptation in the projected residual space. Specifically, temporal innovation statistics guide global covariance-scale calibration for persistent cross-frame drift, while row-wise reliability assessment of projected residuals reduces the influence of localized within-frame contamination. When such contamination becomes severe, a bounded update safeguard further limits unstable corrections without altering the standard MSCKF propagation model or feature-elimination geometry. Experiments on public visual–inertial benchmarks, outdoor vehicle sequences, and an OpenVINS integration demonstrate improved innovation calibration, enhanced robustness to measurement-noise mismatch, and reduced long-tail trajectory errors under challenging residual conditions, with limited update-stage overhead.