Traffic state prediction in a connected automated driving environment

Accurate traffic state prediction is critical to implementing an effective traffic management strategy. Unfortunately, the dynamic nature of traffic flow and the limitations of conventional data sources (e.g., loop detectors and stationary sensors) have historically made it difficult to predict traffic state accurately. Furthermore, the driving environment evolves due to interactions among individual vehicles and their surrounding environment. Accordingly, capturing the interaction among the vehicles can potentially lead to more accurate traffic state prediction. Such interactions, however, cannot be captured via conventional sensors. The introduction of connected and automated vehicles can potentially address the limitations of conventional sensors by providing accurate vehicle trajectory data. Such data can be translated into the time-space diagram, which can capture the interactions among vehicles. Utilizing such data, this study proposes: (1) traffic state prediction methodologies based on convolutional neural networks (CNN) that can directly utilize the time-space diagram in the prediction process. Convolutional layers allow capturing the features embedded in the time-space diagram and accounting for the disturbances such as shockwaves in the traffic stream; and (2) a combined microscopic and macroscopic traffic state prediction methodology to directly utilize the interactions among vehicles in the traffic state prediction process. These interactions can be defined as a combination of lateral and longitudinal maneuvers of vehicles in response to their driving environment. This approach predicts the future trajectory of individual vehicles in the traffic stream and converts that microscopic-level prediction to macroscopic-level prediction. One of the challenges with predicting the trajectory of vehicles is that more than one maneuver is feasible for every vehicle in many driving scenarios. Consequently, the accuracy of prediction at the individual level decays with the increase in the prediction horizon due to the uncertainty in the drivers’ choice of maneuvers and the possibility of various configurations and outcomes. Therefore, this dissertation adopts a probabilistic approach to predict the location of individual vehicles based on different maneuvers. The key step in this combined microscopic and macroscopic traffic state prediction approach is to convert such probabilistic trajectory predictions to aggregated traffic state predictions (i.e., flow, space-mean speed, and density). Note that the traffic state prediction methodologies proposed in this dissertation are data-driven approaches, which require accurate and comprehensive training datasets. Accordingly, this dissertation utilizes both simulation-based and real-world vehicle trajectory datasets to train and evaluate the proposed models.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Civil Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Khajeh-Hosseini, Mohammadreza

Contributors dc:contributor

• Talebpour, Alireza
• Ouyang, Yanfeng
• Golparvar Fard, Mani
• Gupta, Saurabh

• Traffic Prediction
• Fundamental Diagram
• Time-Space Diagram
• Vehicle Trajectory

Statement dc:rights

• Copyright 2022 Mohammadreza Khajeh-Hosseini

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/115760

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F115760