İnsansız hava araçları ile alınan görünür ve kızılötesi görüntülerin hizalanması

Aerial images acquired by satellites, unmanned aerial vehicles (UAVs) or by other means such as balloons, are used in many applications in today's world including entertainment, agriculture, delivery, surveillance, rescue, and disaster management applications. Recent advances in technology and storage hardware allowed cameras to record videos as opposed to individual images in aerial vehicles in multiple modalities. Consequently, there is a recent trend in the relevant research community to leverage such information, when the goal is aerial image or video analysis. As a particular example, timely and accurate segmentation and association of objects of interest in such aerial video frames has been an important research field, recently. When the data is acquired by different cameras focusing on infrared and visible spectrum individually, a particular and an immediate problem is aligning the images (or videos) captured by such different cameras. Image alignment (or registration) is, essentially, the task where an image taken by a different camera is mapped onto the coordinate system of the other camera so that the same objects appear at the same coordinate on both images. In this thesis, we mainly focus on image alignment problem. We develop and propose a deep learning-based solution and compare its performance to traditional and other recent deep learning-based solutions from the relevant literature on multiple datasets. In addition to the image alignment, as additional tasks where such alignment can later be used as input, and as a part of future work, we provide a study on video instance segmentation algorithms that summarizes recent relevant advances and compares their performances. Main contributions in this thesis include: (i) proposing a technique to align images acquired by different cameras on the same edge device, (ii) comparing and reporting the performance of such alignment algorithms on multiple datasets focusing on aerial datasets, (iii) a comprehensive study summarizing the performance evaluations of recently proposed state-of-the-art VIS models on multiple datasets, (iv) presenting an approach on replacing the trainable layers in VIS applications with the non-trainable Fourier based layers.