Deep Learning for Computer Vision in Smart Cities

Abstract

[EN] The Digital Age has caused a rapid shift from traditional industry to an economy mainly based upon information technology. According to recent studies, 74 zettabytes (ZB) of data have been generated, captured and replicated in the world in 2021, with video accounting for 82% of internet traffic. This figure has been amplified due to the coronavirus pandemic, and it is expected to keep increasing, reaching 149 ZB by 2024. Processing this impressive amount of information is one of the main scientific challenges of our time. Against this backdrop, Machine Learning (ML) and two related paradigms have emerged: big data and deep learning. These disciplines take advantage of mathematical optimization methods, bioinspiration and modern Graphics Processing Units (GPUs) to manage large datasets efficiently and effectively. Cities from around the world have adapted the previous methods to make use of the newly available data, promoting themselves as “smart”. Apart from aiming to integrate innovative technologies in their daily operation, Smart Cities (SCs) aim to attract new residents and external investors. Some of the key motivations of the Horizon projects and NextGenerationEU funds are precisely to make cities more digital, greener, healthier and robust. Artificial Intelligence (AI) can greatly contribute to the achievement of those objectives. Several lines of action have been identified in SCs, such as: smart mobility, smart environment, smart people, smart living and smart economy. This dissertation focuses on vision applications of deep learning within the scope of SCs. Theoretical and practical research gaps are identified and suitable solutions are proposed. As a result, the state of the art has been pushed forward and new use cases have been successfully implemented. A novel solution is proposed for each of the identified lines of action. Two models have been designed and evaluated with special attention to efficiency and scalability, and a third model has been created and tested focusing on accuracy within a high-resource environment. Moreover, two novel methods have been developed: a method for automatising crucial healthcare challenges, making early diagnosis an option; and another method for automatic unbiased cadastral categorization.