Intention-informed urban trajectory prediction across scales

Abstract

Urban trajectory prediction is crucial for managing the complexities of densely populated, interconnected urban areas, forecasting movements of pedestrians and vehicles. Essential for traffic management, urban sustainability, and smart city initiatives, it supports autonomous vehicle integration and influences commercial strategies. Despite years of model development, gaps in intention-informed predictions and cross-scale analysis persist, highlighting the need for methodologies that consider both physical movements and underlying intentions, offering deeper insights for urban planning and management. This dissertation provides a detailed study on intention-informed urban trajectory prediction across three scales: micro, network, and city. It proposes distinct models tailored for scale-specific trajectory predictions. The dissertation addresses three research questions: 1) How can intentions behind urban trajectories at each scale be effectively extracted and encoded to enhance prediction accuracy? 2) What are the characteristics and differences of trajectories at each scale, and how can we predict future trajectories using historical data and estimated intentions? 3) What are the implications of incorporating intention information into urban trajectory prediction models in terms of accuracy, scalability, and applicability across different scales? We first present a study on micro-scale trajectory prediction for road agents, such as pedestrians and vehicles, also known as trajectory prediction in Free Space. A deep learning-based micro-scale trajectory prediction framework with intention-guided contrastive clustering is proposed. This model effectively handles the fuzziness and abstraction of human intentions and incorporates the learned intention to enhance the final trajectory prediction. For network-scale trajectories on road networks, informed by cognitive science that humans navigate towards their destination based on the direction, we consider direction information as the underlying human intention in network-scale trajectory prediction. We propose knowledge graph-based methods to effectively leverage road network structure and direction information, querying future routes for more accurate predictions with low latency, suitable for large-scale navigational applications. To understand human movements on a larger scale, we introduce a hierarchical graph attention recurrent network for city-scale human mobility trajectory prediction. Drawing on travel behavior theories and empirical evidence suggesting human mobility patterns are largely activity-driven, we design a hierarchical structure in the model to efficiently utilize activity features for more accurate human mobility predictions. Lastly, we provide comprehensive comparisons and analyses across scales, focusing on data formats, intention learning, and predictive methodologies. This dissertation supports future research in trajectory prediction and mining. Additionally, we introduce an exploratory example of leveraging large language models for open-domain urban itinerary planning, marking the beginning of further research into urban trajectory predictions and generation. This underscores the vast potential for innovative approaches in urban general intelligence and computing.