Hierarchical Vision-Based Localization in Large-Scale GNSS-Denied Environments

This paper presents a hierarchical framework for vision-based absolute localization in large-scale GNSS-denied environments. Initially, a coarse position is estimated using visual observations within a Monte Carlo Localization framework. Each query image is encoded as a global image descriptor and matched against a database of georeferenced images. These reference images resemble what an aerial vehicle would see with a downward-facing camera at that location. Image similarities between the query and reference images are then used as likelihood to update the particle weights in the Monte Carlo localization scheme. We demonstrate that a coarse position estimate can reduce the position uncertainty from multiple kilometers to less than 150 meters. Within this reduced search space, we then match local image features against a satellite image or orthoimage and a corresponding digital surface model to obtain six-degree-of-freedom poses in the global frame. These poses are used to track the absolute position and attitude, which are then fused with inertial measurements using an Extended Kalman Filter. We evaluate our hierarchical localization framework using a challenging real-world aerial dataset that spans more than 38 kilometers in distance at an altitude of 300 meters above ground and covers various types of landscapes, such as urban and vegetated environments. Our evaluations show that our framework can localize over 65% of query images with an error threshold of 5 meters, which allows - combined with pose tracking - for continuous global pose estimation over our test scenario.