Traffic Monitoring from Sensors to Systems: Optimizing Sensor Placement, Minimizing Estimation Errors, and Unifying Macroscopic Fundamental Diagrams

Traffic monitoring is the backbone of traffic management and the essential ingredient for traffic control. With the advent of big data analytics and artificial intelligence applications in traffic, it seems prudent to revisit the components needed for a successful intelligent transportation system. This thesis takes concrete steps in this direction by focusing on the main steps of traffic monitoring (data acquisition, traffic state estimation, and macroscopic traffic estimation) and addressing some gaps in the literature. Namely, this thesis addresses the lack of a solid foundation for sensor location optimization considering macroscopic traffic monitoring, the lack of an effective real-time network-wide link-level model for traffic state estimation using sparse data, and the lack of a holistic understanding of macroscopic fundamental diagrams and what predicts their shapes. This dissertation addresses these gaps by developing a unified “sensors-to-systems” workflow to improve monitoring accuracy and interpretability. First, it formulates a network sensor location problem with an explicit objective of macroscopic monitoring accuracy and proposes a PCA-based approach that ranks links by newly defined criticality measures aided by clustering. The developed model is shown to significantly reduce errors compared to a benchmark. Second, it introduces a real-time state estimation framework that tracks waves through merges and diverges and retroactively corrects estimates. Both a toy network and a large simulated network from Philadelphia verify the accuracy of the model. Third, the dissertation investigates whether MFD shape can be predicted from non-traffic features and whether a near-universal MFD form can be derived. A GNN is trained to predict the uncongested branch of the MFD, but results show strong asymmetry between flow and density predictability, indicating data and model depth limitations and a fundamentally harder density-prediction problem. Complementary statistical analysis finds that flow-related targets are relatively explainable while density at capacity is only moderately predictable. Feature importance measures highlight socioeconomic variables, urbanization proxies, environmental conditions, and detector placement as key predictors. Finally, based on these insights and parallels from ecological research, an interpretable hybrid analytical MFD is proposed that blends cubic and exponential behavior using an urbanization-dependent transition. It outperforms classical link-level FD forms across diverse networks.