Construction of Feature Landmark Database Using Omnidirectional Videos and GPS Positions

Abstract

This paper describes a method for constructing feature landmark database using omnidirectional videos and GPS positions acquired in outdoor environments. The feature landmark database is used to estimate camera positions and postures for various applications such as augmented reality systems and self-localization of robots and automobiles. We have already proposed a camera position and posture estimation method using landmark database that stores 3D positions of sparse feature points with their view-dependent image templates. For large environments, the cost for construction of landmark database is high because conventional 3-D reconstruction methods requires measuring some absolute positions of feature points manually to suppress accumulative estimation errors in structure-from-motion process. To achieve automatic construction of landmark database for large outdoor environments, we newly propose a method that constructs database without manual specification of features using omnidirectional videos and GPS positions.