AVC-Init: aggregated visual constraints for decoupled initialization in visual-inertial navigation

Abstract

Accurate and stable initialization is critical for Visual-Inertial Odometry (VIO). Existing methods often formulate visual constraints at the individual-observation level, so their computational cost grows with the number of feature measurements. They may also maintain 3D landmarks in initialization, which increases the state dimension and complicates the trade-off between accuracy and stability. This paper presents AVC-Init, a decoupled visual-inertial initialization framework based on Aggregated Visual Constraints (AVC). AVC-Init first estimates the gyroscope bias from aggregated rotational constraints between co-visible keyframe pairs. Preintegration merging is adopted to reuse inertial information across non-adjacent pairs. The framework then estimates the initial velocity and gravity with global-position constraints in either a loosely coupled or tightly coupled formulation. For joint refinement, Visual-Inertial Epipolar Adjustment (VI-EA) aggregates feature observations into structureless epipolar constraints. VI-EA removes 3D landmarks from the optimization state and enables low-dimensional joint refinement. Finally, depth-weighted multi-view triangulation recovers landmarks from the refined poses. Experiments on simulated data, EuRoC, TUM-VI, and a self-collected outdoor dataset show that AVC-Init achieves higher success rates and lower errors than the tested baselines in most settings.