An Artificial Intelligent methodology to estimate the population density of urban areas to compute risk maps for Unmanned Aircraft Systems

As the adoption and usage of Unmanned Aircraft Systems (UAS) is growing over time, the safety of the surrounding environment when they are used autonomously over urban areas has been questioned. Urban areas are characterized by high population density therefore a potential UAS crash may involve people injuries or even people death, as a consequence, it is essential to preserve human safety in mission planning. This thesis is based on a previous work done by the research group, in which a path planning algorithm for UAS has been developed to determine a safe flight mission. The proposed methodology uses a risk-based map that quantifies the risk of fly over specific areas. In order to compute a safe path for the UAS several input parameters are required, but, unfortunately it’s not always easy to obtain those specific data, such as the population density and distribution that is one of the most important parameter for the risk assessment. The main contribution of this thesis is the use of an Artificial Intelligent methodology to estimate the population density of urban areas using aerial images. First of all, the Bing Maps REST Services has been used to easy download aerial images. Then, a Convolutional Neural Network has been developed in order to estimate the population density and distribution and to define the risk-based map for the UAS flights. Due to the big amount of computational effort needed to train an enough depth CNN, the pre-trained CNN VGG16 has been chosen as starting point thanks to its specificity for large-scale image recognition and its capability to classify common objects over 1000 categories. As a starting point, the pre-trained CNN has been readapted to recognize and classify 224 x 224 RGB images taken from Bing Maps. Then, more and more complexity has been added with the goal of classifying images into multiple ranges of population density value in order to distinguish which zones are with an higher population density, or rather, an higher risk. As last step, to get even more accurate results, it has been decided to develop another training approach which would avoid errors due to misleading images. Therefore, the Neural Network has been trained through the numerical regression to predict the exact value of the population density of aerial images. This value has been used to compute risk-maps and to plan safe paths for UAS.