A Consistency-Improved Visual-Inertial Odometry for Localization in NTN-Assisted IoT Applications

Our previous work1 introduced a visual-inertial odometry (VIO) method that leverages long-tracked drifted features, which is able to perform precise localization in complex environments. However, a key limitation is its reliance on computationally intensive large sliding window filter (SWF) to maintain performance. By analyzing the observability of the original system, we shown that the consistency of the SWF is disrupted due to the discrepancy between the linearization points of the prior state in the historical SWF and the current one. In this case, the covariance matrix of the estimation errors is larger than that predicted by the SWF, leading to inaccurate state updates and loss of accuracy. The original system achieves delayed marginalization by increasing the capacity of the SWF, thereby reducing the discrepancy of the linearization points and, in turn, minimizing the disruption to the consistency of the SWF. In this paper, we adopt a more efficient strategy by aligning the null space matrix of the prior with that of the current SWF, thereby avoiding the destruction of the consistency. We evaluate our method on the EuRoC, TUM-VI, and 4Seasons datasets, comparing it with state-of-the-art methods. Experimental results show that our approach achieves high localization accuracy even with smaller SWF, making it more suitable for non-terrestrial networks (NTN)-based deployments with resource constraints. The code will be published on GitHub.2.