Correlation and alternatives for scene matching

A common approach to matching two scenes is the use of an area cross-correlation measure. This measure is proportional to the root-mean-square difference between the corresponding pixel intensities for the two scenes. In practice, it is often found that the magnitude of this measure at the correct match position may be rather low, and that the maximum value appears at an incorrect match position. This problem of false-fixing results from differences between the reference and sensed scene in their respective range, viewing angle, illumination, contrast level, and imaging sensor type. In this paper we present a comparison and trade-offs between correlation and alternative matching techniques that do not suffer from the same weaknesses as those of correlation. A specific example of a matching scheme is also discussed that uses a scene model based on local structural features from the scene. The results of such a matching scheme on a set of two representative images are presented.