Self-organizing property of nonlinear mapping for change detection

We had proposed a nonlinear mapping method for the purpose of change detection based on two images photographed at different points of time. The change detection method involves three fundamental research factors, i.e. (1) characteristics to find the corresponding regions between two images, (2) generation of mapping vector field, and (3) characteristics to detect changed area. We, so far, have been focused on the third factor of the change detection and the shifting vectors, i.e. the non-parametric expression of deformation between two images, were generated based on ad hoc parameters through trial and error. In this paper, we shift the research target from evaluating change detection performance to statistical analysis of shifting vector generation which corresponds to the second fundamental research factor. At first, the conventional method, which involves dependency between the size of local regions and the size of reference points, is redefined so as to permit the combination of arbitrary size of local region and reference points distribution. We examine statistical characteristics, i.e. matching error, divergence and estimation error, of shifting vector estimation for the purpose of the separation and reduction of misleading vectors among estimated vector field.