A simple visual compass with learned pixel weights

A mobile robot can determine its rotation with respect to a goal position by comparing the image taken at the goal with the current image. Assuming that both images are omnidirectional, the current image can be shifted horizontally to determine the minimum difference with the goal image, and hence the rotation. The size of the region around the goal wherein this strategy succeeds is reduced by the presence of nearby objects and ambiguous image features. We propose a simple pixel weighting scheme that alleviates these problems. The approach is validated on images captured by a mobile robot in a typical office environment.