Visual Compass Methods for Robot Navigation

A compass based purely on visual information would be a useful navigational tool for robots working in environments where magnetic and GPS compass information is not available. A visual compass processes a pair of panoramic images taken from two different positions and estimates the orientation change between them. We present in this paper several well known algorithms useful for the development of a visual compass. These methods include a search for the minimum image distance in orientation space, a refined search method, and a frequencydomain approach known as phase correlation. A comparison on efficiency and effectiveness of these algorithms and their variations is given. We find that the refined search method exhibits improved efficiency with only a marginal impact on accuracy. The 2-D phase correlation method performed almost as well for small distances, but for large distances, the accuracy decreased significantly. The 1-D phase correlation performs better with respect to both accuracy and efficiency.