Urban environments are becoming increasingly diverse, dynamic, and complex ecosystems. This transformation is driven by global urbanisation, rising population densities, and technological advancements. Understanding urban environments is becoming both essential and challenging due to their rapid evolution and inherent dynamism. Traditional methods often struggle to meet such demand. Therefore, there is a growing need to leverage new forms of urban data that can capture contextual and behavioural insights at a higher frequency to facilitate timely and responsive decision-making within urban environments. In recent decades, the widespread use of smart devices, data-sharing platforms, and improvements in computational hardware have facilitated the continuous generation of new forms of urban data. Advances in artificial intelligence and deep learning models, along with such data, have opened up significant opportunities for developing cost-effective and scalable applications that enhance our understanding of urban environments from new perspectives. This thesis makes three complementary contributions to the field of human-centric urban analytics. First, it presents an innovative approach for building height estimation by leveraging ubiquitous mobile signals. The proposed approach provides a cost-effective, globally accessible, and efficient solution for creating 3D maps, which requires no dedicated equipment beyond consumer-level mobile phones. Second, the thesis contributes humancentric research to support humans in cities. It demonstrates the effectiveness of UltraWideband (UWB) signals for fine-grained human activity recognition, further emphasising its cost-effective, non-intrusive, and reliable potentials. Additionally, the thesis analyses urban pedestrian disorientation in complex city environments based on survey data from the Greater London Area. The research identifies and quantifies factors leading to disorientation, employing expert-led Analytical Hierarchy Process (AHP) and data-driven regression methods. Both studies offer insights for designing more inclusive and navigable urban spaces. Third, this thesis addresses an often-overlooked issue in the development of deep learning models for positioning: temporal bias in visual urban datasets. Using cross-view geo-localisation (CVGL) as a case study, this thesis evaluates the performance of two state-of-the-art CVGL models on an original benchmark dataset and a custom dataset spatially aligned with the benchmark. Our findings reveal significant degradation in model performance due to temporal variations between two datasets. Semantic segmentation and SHapley Additive exPlanations (SHAP) explainability framework are used to further illustrate how temporal visual changes affect model reliability. In summary, this thesis presents an effective framework for urban analytics that prioritises human-centric approaches. It covers the collection and creation of custom datasets in new formats, as well as the development of novel sensing methods at various scales. The research offers valuable insights aimed at improving positioning and navigation services, while also expanding humans’ understanding of urban environments. Importantly, it identifies hidden biases in urban data applications. Overall, this work demonstrates how emerging data sources and analytical techniques can improve our ability to model, understand, and design smarter and more inclusive urban environments.
Yu Wang (Thu,) studied this question.