Consistency-Enhanced Tightly Coupled Visual-Inertial Odometry via Bounded Double Window Optimization

Abstract

Non-linear optimization in visual-inertial odometry (VIO) significantly enhances localization accuracy and robustness for autonomous robots and MAVs in dynamic environments. However, visual tracking instability, often induced by motion blur or abrupt lighting changes, frequently causes trajectory drift, compromising precise localization. The proposed method maintains visual reprojection constraints within an inner window, while applying marginalized relative pose constraints in an outer window, seamlessly linked by pre-integrated inertial measurements. To mitigate uncertainties commonly observed in low-cost sensors, our inertial pre-integration explicitly models axis misalignment, scale factors, and g-sensitivity. Crucially, as keyframes exit the inner window, reprojection costs are transformed into relative pose constraints with accurately computed full covariance matrices. Explicitly propagating this full covariance improves estimator consistency and prevents overconfident updates during sustained tracking failures. Furthermore, the bounded double-window architecture strictly limits the joint state dimension, ensuring computational scalability without the excessive overhead typical of global optimization methods. In addition, evaluations on the EuRoC dataset demonstrate that our algorithm outperforms widely-used methods, effectively handling highly noisy and visually blind segments.