Observability-Weighted Visual-Inertial NavigationSystem

Abstract

In recent years, autonomous driving has continuously developed alongside daily technological advancements. Simultaneous Localization and Mapping (SLAM) is one of the significant techniques applied in this field. However, the adoption of autonomous driving techniques for self-driving cars or drones is not yet widespread. The main reason is that the accuracy and robustness of localization systems still need improvement to meet the requirements for autonomous driving and extended applications. This work examines the relationship between the observability and uncertainty of the estimation system and identifies the features that should be prioritized by analyzing their effect on system observability. Based on this analysis, the work aims to determine which features are not significant and can be temporarily excluded from the current estimation in a multi-sensor system. Additionally, this research weighs feature points by their individual observability, considering the influence of each observed feature point to improve estimation accuracy. The study introduces methods to consider system observability in estimation and presents simulation results. Ultimately, applying this method to multiple datasets demonstrates better estimation results compared to other methods.