Visual compass

We present in this paper a method to incrementally estimate the heading of a robot using panoramic images grabbed while the robot moves in its environment. We present results that show a performance similar to a magnetic compass, if not better in some cases. In particular, the heading measured by the visual compass does not drift as normal odometry does, provided the robot turns roughly by the same amount in both directions. We discuss the performance of the system as well as its limitations. We also briefly give examples of applications where the visual compass could be used and constitute an improvement.