Abstract

Unmanned Aerial Vehicles (UAVs), commonly known as drones, are aerial platforms that are gaining large popularity in the remote sensing field. UAVs derive from military technology, but in the last few years they are establishing as reference platforms also for civilian tasks. The main advantage of these acquisition systems lies in their simplicity of use. Indeed, a UAV can be used when and where it is needed without excessive costs. Since UAVs can fly very close to the objects under investigation they allow the acquisition of extremely high resolution (EHR) images in which the items are described with a very high level of details. The huge quantity of information contained in UAV images opens the way to develop novel applications but at the same time force us to face new challenging problems at methodological level. This thesis represents a modest but hopefully useful contribution towards making UAV images completely understood and easily processed and analyzed. In particular, the proposed methodological contributions include: i) two methods devoted to the automatic detection and counting of cars present in urban scenarios; ii) a complete processing chain which monitors the traffic and estimate the speeds of moving vehicles; iii) a methodology which detects classes of objects by exploiting a nonlinear filter which combines image gradient features at different orders and Gaussian process (GP) modeling; iv) a novel strategy to “coarsely” describe extremely high resolution images using various representation and matching strategies. Experimental results conducted on real UAV images are presented and discussed. They show the validity of the proposed methods and suggest future possible improvements. Furthermore, they confirm that despite the complexity of the considered images, the potential of UAV images is very wide.