Landmark selection for scene matching with knowledge of color histogram

Scene matching is used in the vision based automated navigation error correction technique in the absence of global positioning systems for unmanned aerial vehicles. When knowledge of landmarks in the scene is known a priori, the scene matching can be carried out in a more accurate and efficient way by considering a landmarks-only matching process. In this paper, we present two landmark selection algorithms where knowledge of landmarks in an aerial image is represented by color histograms which can be computed in advance. In landmark selection, one method treats the landmark selection as a population sampling problem and searches the population of a given landmark over the image via a Kullback Leibler type divergence measure. The other method computes the probability that an image point originates from a landmark and this probability is approximately calculated via the color histogram of the landmark. The performance of the two proposed algorithms is compared in a landmark detection scenario along with the selection results from a SUN saliency model trained using the landmark data as well. Experimental results show that the proposed algorithms are simple but effective for the landmark selection task.