Deep Learning Methods in Transportation and Urban Planning

QC 20260511

Abstract

Rapid urban development and the overwhelming amount of data provided by connected sensors, mobile devices, and open-source datasets challenge traditional transport-analysis methodologies. These methodologies largely depend on rigid mathematical models, static assumptions, and sparse measurements. This thesis demonstrates how deep learning (DL) has the potential to build a systematically better data collection and inference capability across three critical domains of the LUTI cycle: traffic management, population synthesis, and workplace location choice. Through five research papers, this thesis demonstrates that DL methods can complement or outperform traditional approaches by extracting more comprehensive data, building better predictive models, and providing actionable implications for planners. The analysis is separated into two main themes: data acquisition and analytical inference.

The data acquisition theme proposes methods for transforming overlooked or missing data into valuable input for transport models. Paper 1 uses street-view video from vehicle-oriented cameras to integrate YOLOv5 object detection, StrongSORT tracking, photogrammetry, and geodesy, automating the creation of link-wide time-space diagrams. The analysis of the method indicates that ordinary vehicles can be used to construct a low-cost virtual sensor network with far more spatial coverage than fixed detectors or GPS floating-car data. Paper 3 deals with missing-attribute surveys by including a binary mask in a Wasserstein GAN (WGAN) training setup. The masked WGAN learns directly from incomplete microsamples and creates synthetic populations that closely match marginal and joint distributions similarly to models trained on complete data, thus recovering survey entries that would otherwise be wasted.

The analytical inference theme associates the novel data sources with DL-enabled and hybrid models. Paper 2 adapts the Cell Transmission Model (CTM) and a two-stage Genetic Algorithm to perform fundamental-diagram parameter calibration and boundary condition assignment from partial camera trajectories. It recovers free-flow speed and critical density across three simulated arterial flows, and estimates unobserved densities, suggesting that sparse vision data can promote robust, link-level traffic state estimation. Paper 4 uses a Conditional Tabular GAN (CT-GAN) to construct target-year synthetic populations out of marginals directly, alongside a hybrid CT-GAN and Fitness-Based Combinatorial Optimisation pipeline that fine-tunes the marginal fit. During experimentation, CT-GAN demonstrates better performance for conditioned attributes over a baseline IPF-style model, achieving higher convergence and fidelity across the whole sequence. Paper 5 presents a custom Deep Neural Network (DNN) with zone blocks that ingest occupation type, socio-economic variables, and multimodal accessibility to predict workplace choice for over 1,000 job options. The DNN achieves superior performance metrics and reproduces observed distance distributions more accurately than a two-level nested logit model, without needing to manually construct a utility specification.

Overall, this thesis makes four main contributions to the literature: (1) Identifying vehicle-mounted cameras and incomplete surveys as valid, high-resolution sources of data when combined with DL pipelines; (2) Extending traffic state estimation to use partial trajectories extracted from video sequences to extract useful traffic states by means of GA-calibrated CTMs; (3) Improving synthetic population generation, providing evidence that GANs can satisfy future marginal constraints and learn from sparse, masked data; and (4) Generalising deep choice modelling to thousand-alternative examples, showing that DNNs can equal or exceed traditional discrete-choice models in accuracy and behavioural realism.

This work also examines drawbacks such as sensor calibration mistakes, modelling assumptions, and the generalisation and scalability issues of DL models, presenting solutions to these problems. The thesis also discusses policy implications of the research and provides recommendations for the use of DL, with lessons that are transferable to sectors such as energy, public health, and spatial systems.

In summary, this thesis advances the state of the art by embedding modern deep learning approaches into the core of urban transport modelling, paving the way for more data-rich, flexible, and socially-driven transport-planning tools in an increasingly complex society. Finally, it demonstrates that by combining domain knowledge with scalable DL architectures, the potential of current urban data can be unlocked to support more efficient, sustainable, and equitable mobility systems.

Link to DiVA