Visual-Inertial Localization and Map Summarization Based on Prior Map

Abstract

Accurate localization of the robot is an important prerequisite for the autonomous mobile robot. Existing localization methods struggle with the cumulative drift errors problem. We proposes the visual-inertial localization method which based on multi prior maps and generates a summary map with a fixed map size. Specifically, for the localization problem with gravity alignment, the relative pose of the pitch and roll is known, which reduces the dimensions of the problem. For the 3D-2D data association of the map and current query image, both geometric and reprojection constrains are used. In the process of map summarization, the idea of iterative map building is proposed and a novel scoring strategy is exploited to limit the summary map to a fixed size. We evaluated our method on public and our datasets. The result indicates that our method owns higher localization accuracy and better robustness than other comparative method.