This project focuses on development of an algorithm for Template Matching on aerial imagery by implementing classical Computer Vision based techniques and deep-learning based techniques. This template matching technique is applied to extract GPS coordinates of the localized template image from the main image. This project primarily focuses on the problem of Vision-based Localization of Unmanned Aerial Vehicles.
This algorithm is implemented using SIFT[1] descriptors. The GPS coordinates are also extracted using GDAL functions.
This algorithm is currently under development. It utilizes Neighbourhood Consensus Networks[2] presented in NeurIPS 2018 to extract point to point correspondances between two images.
[1] David G Lowe, “Distinctive Image Features from Scale-Invariant Keypoints,” International Journal of Computer Vision, vol.50, No. 2, 2004, pp.91-110
[2] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelovic, Akihiko Torii, Tomas Pajdla, Josef Sivic, “Neighbourhood Consensus Networks,” NeurIPS 2018