Abstract: Visual-inertial odometry (VIO) is widely applied in UAV and robotics navigation. However, existing VIO systems often struggle with robustness when feature point matching quality is inconsistent. Most VIO algorithms assume a constant covariance matrix for the noise term in the observation model, overlooking variations in feature point matching quality. To address this issue, this paper proposes a VIO localization method based on the multi-state constraint Kalman filter (MSCKF), incorporating normalized cross-correlation (NCC) for dynamic covariance adjustment. This method constructs a new observation model by introducing tracking errors for feature points at the pixel level, with NCC employed as a metric for quantifying feature point matching quality. By calculating the NCC of the feature matches, the method adjusts the covariance matrix of the observation noise dynamically to reflect variations in the matching quality. This approach leads to more accurate and robust results, especially in complex environments with significant differences in feature point matching quality. Comparative experiments conducted on the EuRoC open dataset and a real-world underground parking lot dataset demonstrate that, compared to the traditional MSCKF algorithm, the proposed method reduces planar root mean square error (RMSE) by 41.8% and 33.8%, respectively. Furthermore, compared to the VINS-MONO algorithm, the proposed method reduces planar RMSE by 26.3% and 24.7%, respectively. These results demonstrate a significant improvement in both the robustness and localization accuracy of VIO systems under challenging feature matching conditions.

Key words: VIO, MSCKF, NCC, matching quality, observation noise