Large scale vision-based navigation without an accurate global reconstruction