In-Flight Initialization of Global Visual-Inertial Estimators using Geospatial Data

In this work, we propose a solution that leverages geospatial data to initialize the monocular visual-inertial navigation system. For Visual-Inertial Navigation Systems (VINS) operating on UAVs, the ability to perform initialization and relocalization in mid-air is essential. However, degenerate motion can cause VINS to lose scale, making traditional initialization algorithms less reliable. To address this issue, we fuse geographic information in the initialization process, and utilize a learning-based feature matching algorithm to associate the information with inertial states. The proposed approach demonstrates adaptability to the degenerate motions of UAVs and significantly surpasses the estimation accuracy of conventional VINS initialization algorithms. Compared to methods that assist initialization by using a laser-range-finder (LRF), the proposed method solely relies on low-cost satellite imagery and elevation information. We evaluate the proposed approach on a large-scale UAV dataset, and compare with existing methods. The results demonstrate the superior effectiveness of the proposed method.