Multi-image matching for 3D mapping in vision-based navigation applications

In this paper, we introduce multi-image matching using invariant features into the 3D mapping process for vision-based navigation applications. Our previously developed vision-based positioning system is modified accordingly, and experiments prove that the 3D map has been enhanced under a better imaging geometry. Consequently, the localization accuracy is much improved. Final positioning accuracy is around 0.2 m. The test also reveals some limitations: the system cannot perform well in areas where features are not rich or not included in the view; mismatches affect the positioning accuracy.