Sheltering factor recognition via machine learning for fully autonomous UAV urban path planning

The environment of the work done in this thesis are the smart cities. A smart city is an urban area that collects data in order to efficiently manage infrastructures and resources. The proposed architecture involves the use of three different systems: the autonomous vehicles, the non-autonomous vehicles and the cloud system. The autonomous vehicles have the purpose of acquiring information, the non-autonomous vehicles have the purpose of acquiring information and support the autonomous vehicles, the cloud system processes the monitoring data and execute the path algorithm. The path planner algorithm calculate the path optimizing a risk function, evaluated considering the cells of a risk map. In order to generate the risk map, four different layers are required: occupation layer, population density layer, sheltering factor layer and signal power layer. The aim of this thesis is to improve the performance of the path planning algorithms considering two aspects: considering the autonomous vehicle crash distance after a failure and automating the generation of the sheltering factor layer. The crash distance was considered differentiating fixed wing UAVs from rotary wing UAVs. For fixed wing, starting from the equilibrium equations of a body, a simple equation has been found. The equation is able to evaluate the crash distance knowing the aerodynamic efficiency and initial altitude of the UAV. The effect of the implementation of the crash distance on the risk map generation are analyzed, in particular it was found that the risk shows a more uniform distribution with larger values of crash distance. Vice versa, for smaller crash distances, the risk distribution in the risk map is less uniform, with areas characterized by a much higher risk value than others. For the rotary wing, the ballistic equations have been numerically integrated. The obtained trajectories have been compared, considering different values of mass, drag coefficient, altitude, initial velocity and reference surface. The results are presented in two forms, one specific and one more generic. In the specific case, knowing a priori all the characteristics of the UAV considered (mass, drag coefficient, reference surface and flying altitude), it is possible to obtain an equation that evaluate the crash distance as a function of the initial speed. In the second generic case, a table has been generated. This table contains numerous combinations of mass, drag coefficient, reference surface, flight altitude and speed associating each of them the corresponding crash distance. The automation of the generation of the sheltering factor layer was performed by training neural networks in order to solve the problem of image classification. A database was created considering open source satellite images. Two different networks have been trained. The first was made from scratch, in order to be very light and fast. This network considers only one channel of the input images and is composed of a small number of layers. The second network is an already existing neural network, in this case transfer learning has been performed. This network is more complex and the training much slower. In the first case, the training was not successful enough to be used for generating the sheltering factor layer. In fact, the trained network accuracy is relatively small and presents overfitting. Necessary improvements in order to increase this network performance are described, including the implementation of a variable learning rate and the consideration of all three image channels instead of only one. In the second case, the network gave satisfactory results. Accuracy is relatively high with no overfitting observed. The improvements that could increase the performance of this network mainly concern the database. In particular, using a better database with higher image resolution and quality is suggested as future developments.