Vision-based robot homing in dynamic environments

This paper presents a vision-based robot homing method capable to deal with a series of dynamic changes in the environment. First we build a robust and stable representation of the goal location using scale invariant features (SIFT) as visual landmarks of the scene, followed by a matching and a voting scheme. We end up with a description that contains the most repetitive features which best represent the target location. The vision-based homing algorithm used assumes that the images have the same orientation, therefore in the second stage, using the position of the SIFT features we recover the orientation misalignment between the current view and the goal view. Finally, the home vector between these two positions is calculated using the SIFT matches as a correspondence field. Experiments in static and dynamic environment show the suitability of using local invariant features for robot homing in outdoor environments.