A deep learning approach to estimate population density of urban areas to compute risk maps for UASs

Unmanned aerial vehicles (UAVs) have gained interest in recent years in many different fields. UAVs are used for remote sensing, surveillance, precision agriculture and goods delivery, to name a few. The growing interest in their use leads to safety concerns which can be faced using a systematic approach to define safe operations. The approach proposed in this thesis consists in the definition of risk maps that quantify the risk connected with a flight operation over a certain area. Risk maps can be used for risk-informed decision making: by National Aviation Authorities to evaluate the risk of a specific flight mission, or to compute safe routes using a risk-aware path planning algorithm. Specifically, the risk is measured as the probability of having a fatality due to a drone crash. This risk value is computed combining several layers consisting of data essential for the risk assessment, such as population density, sheltering factor, no-fly zones and obstacles. The main goal of this thesis is to estimate the population density layer using aerial imagery and a proprietary database of the population density as ground truth. Knowing how people are distributed in a given area is essential to understand how dangerous the impact of a vehicle on the ground could be. Generally speaking, sparsely populated areas will be considered safer than highly populated ones to fly over. In this thesis, a deep learning-based approach for the creation of population density maps is adopted. Some of the most promising results in this field have been obtained in the last years using convolutional neural networks (CNNs). A CNN is a machine learning model in which particular layers are employed, with the goal of learning features from databases composed of images. The pre-trained VGG16 network has been employed as feature extractor, on the top of which dense layers were added to reduce the output dimension. Techniques like fine tuning and test-time augmentation have been employed to increase the precision of the maps obtained. The deep learning model was implemented using Tensorflow and Keras, which are libraries that provide ready-to-use tools to train and evaluate machine learning models. The model was trained in an Amazon Web Service (AWS) EC2 instance with a dedicated GPU that is designed for deep learning applications. The CNN was trained on the area of Turin (city centre and surrounding area), and satellite imagery were downloaded from Bing Maps. After the evaluation of the network, an analysis has been performed looking at the most and less precise estimates. This process is useful to understand the limits of the model used and the problems of the database (for example, limited resolution and perspective of the images). In the final part of the thesis, the same configuration of CNN is used to directly estimate of the risk obtaining promising results, but with a less accuracy than the population density estimation. The thesis is carried out within the activities of the Amazon Research Award “From Shortest to Safest Path Navigation: An AI-Powered Framework for Risk-Aware Autonomous Navigation of UASs” granted to Prof. Rizzo.