Adeep cnn-based Framework for enhanced aerial imagery registration with applications to uav geo-localization /
Ahmed Nassar
- 2017
- 102 p. ill. 21 cm.
Abstract: Using the momentous advancement of deep learning in traditional computer vision problems such as perception has been greatly benecial. Along with the availability of high resolution Earth Observation imagery, extraction of information such as roads, buildings, vegetation, etc., using deep learning is a major topic for remote sensing applications. We transfer these methods to UAVs or drones equipped with an on- board camera, and GPU. In this thesis, we present a method for geo-localizing UAVs using only it's onboard camera. A pipeline has been implemented that makes use of the availability of satellite imagery, and traditional computer vision (SIFT), along with renowned deep learning methods (semantic segmentation) to be able to locate the aerial image captured from the UAV inside of the satellite imagery. A novel solution of matching segmented shapes is used to register the segmented images, and increase the accuracy of the system overall. Additionally, this method enables the UAV to be self-aware of its surrounding and the context its in. Consequently, this allows UAVs to be able to navigate without the use of GPS, which could be benecial in environments where GPS is denied or weak, or be an additional enhancement module to GPS. Our framework is evaluated on two dierent datasets generated from dierent regions, to show the viability of the method in dierent situations. These datasets, and the process of generating them is explained in the literature. Furthermore, we believe our results show that the framework oers a promising approach for outdoor navigation for UAVs using vision only.