HBA-VIO: High-Precision Direct Sparse Monocular Visual-Inertial Odometry via Hybrid Bundle Adjustment

Visual-inertial odometry (VIO) is a core perception module for autonomous robots, where high-precision pose estimation is essential for reliable navigation. However, low-cost setups equipped with inertial measurement units (IMUs) and monocular cameras suffer from cumulative trajectory drift. In this article, we propose HBA-VIO, a high-precision direct sparse monocular VIO system based on hybrid bundle adjustment (HBA). First, to address initialization challenges in IMU-based systems, a multistage visual-inertial initialization strategy is proposed, consisting of static initialization and coarse-to-fine visual-inertial alignment. This ensures precise calibration between monocular visual odometry and inertial measurements across diverse scenarios. Second, for high-accuracy trajectory estimation, we design an HBA framework integrating backward-decoupled preintegration residuals, inertial-only prior residuals, cross-frame preintegration residuals, and a scale refinement module, which effectively refines the global scale and suppresses cumulative trajectory drift. Extensive evaluations are performed on real-world challenging scenarios. Comprehensive experimental results demonstrate that HBA-VIO achieves state-of-the-art accuracy, validating the effectiveness of each technical component and the system’s strong robustness to real-world disturbances.