Comparison of image similarity functions and sampling algorithms in vision-based particle filter for UAV localization

Abstract

This paper analyses parts of a computer vision based particle filter for Unmanned Air Vehicle (abbr. UAV) localization. Localization is done by matching camera image from downward looking camera on a UAV to a previously known orthophoto map. Few image matching functions are compared, to select the best fit matching coefficient for the case. Normalized correlation-coefficient with min-max 110 normalization was used to calculate the most fit probability density function. Few sampling techniques are reviewed, implemented and compared to achieve UAV localization in a GPS denied environment. Kueller-Leiblach distance (abbr. KLD) sampling technique has shown the best localization success rate (96 %) with lowest computational requirement (about 1,7 times faster than other sampling algorithms).