Mining visual experience for fast cross-view UAV localization

A novel visual image retrieval technique for fast cross-view UAV localization is presented in this paper. Our first contribution is to address the computational complexity of raw image matching, which can be time/space intractable due to the high dimensionality of raw image data. We propose to exploit raw image matching, not for the direct matching between query and database images, but for mining an available visual experience to find discriminative visual landmarks. The mined library images are then compared between query and database images using a naive Bayes nearest neighbor (NBNN) distance metric that has proven to be successful in cross domain (i.e., cross-view) image comparison. We developed a practical localization system consisting of a pipeline of two stages: (1) image retrieval using the NBNN distance metric, and (2) post verification of image matches using CNN feature. Experimental results show that our proposed framework tends to produce stable localization results despite the fact that our approach is significantly space/time efficient.