Date of Award

Summer 8-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Civil & Environmental Engineering

Program/Concentration

Civil and Environmental Engineering

Committee Director

Kun Xie

Committee Member

Mecit Cetin

Committee Member

Hong Yang

Committee Member

Qing Tang

Abstract

Traffic crashes between vehicles and pedestrians arise from complex, split-second interactions in which both parties make rapid evasive decisions. Four fundamental gaps persist in existing research. Surrogate safety measures assume linear trajectories, failing to capture the curved movements of turning vehicles and crossing pedestrians at intersections. Single-agent modeling treats one party as a fixed obstacle, ignoring the joint decision-making that governs near-miss outcomes. The effect of vehicle type on pedestrian avoidance behavior—whether pedestrians respond differently to automated vehicles (AVs) than to human-driven vehicles (HDVs)—remains poorly understood. Finally, automated driving system development is hampered by a severe scarcity of large-scale, behaviorally realistic safety-critical datasets. This dissertation addresses these four gaps through a progressive, interconnected research program grounded in deep reinforcement learning.

Chapter 3 establishes the measurement foundation. A curvilinear time-to-collision (CurvTTC) metric accommodates the actual curved trajectories of road users at intersections. ANOVA and Tukey HSD tests confirm that CurvTTC detects safety-critical conflicts significantly more effectively than standard TTC and 2DTTC. CurvTTC then serves as the conflict-screening criterion throughout the remaining chapters. Chapter 4 advances from single-agent to multi-agent modeling. The MA-SST-DDPG framework integrates a state-space module and Transformer within multi-agent deep reinforcement learning to jointly model vehicle–pedestrian interactive decision-making. The framework achieves the lowest trajectory reconstruction errors among all baselines. Chapter 5 extends the framework to a comparative context. Using the Argoverse 2 dataset, the SMamba-DDPG framework—incorporating a smoothing module that stabilizes learning under abrupt state transitions—reconstructs and compares pedestrian avoidance behaviors in Pedestrian–AV and Pedestrian–HDV interactions. Counterfactual analysis shows pedestrians reduce mean speed by 14.7% and yield more frequently toward AVs (53.30% vs. 51.53% for HDVs), with consistently lower conflict rates in Pedestrian–AV interactions. Chapter 6 closes the data scarcity gap. A three-stage hybrid learning framework combining real-world pre-training, CARLA-based online learning, and large-scale scenario generation produces the VPSCI dataset—198,157 safety-critical episodes across eight intersection scenarios. A Turing test with 51 participants finds no statistically significant difference in realism between generated and real-world interactions (p = 0.92). The four chapters form a coherent pipeline: CurvTTC identifies events → MA-SST-DDPG models interactive behavior → SMamba-DDPG reveals vehicle-type effects → the three-stage framework scales these behaviors into a large, publicly released dataset.

The four studies demonstrate that deep reinforcement learning can capture realistic, interactive crash avoidance behaviors in safety-critical vehicle–pedestrian scenarios. CurvTTC provides a geometrically accurate conflict measure suited to intersection dynamics. MA-SST-DDPG reproduces speed-dependent negotiation patterns observed in real near-miss data. SMamba-DDPG reveals systematic behavioral differences between Pedestrian–AV and Pedestrian–HDV interactions with direct implications for mixed-traffic safety modeling. The VPSCI dataset supplies a large-scale, behaviorally validated resource that the field has lacked. Together, these contributions advance safety-aware traffic simulation, surrogate safety measurement, and data-driven autonomous driving system design.

The developed frameworks translate directly into engineering and policy tools. The learned avoidance policies serve as pedestrian intent models at unsignalized crossings, enabling earlier AV yield decisions. For emergency braking systems, the interaction models supply speed- and geometry-aware conflict thresholds that reduce false alarms and late activations. As drop-in controllers for SUMO, VISSIM, or CARLA, the trained agents enable higher-fidelity simulation for intersection design and AV policy validation. For traffic agencies, the framework generates safety-critical scenarios for a given intersection and ranks countermeasures by predicted conflict reduction—proactively, before crashes occur.

Future work will extend the frameworks to T-junctions, roundabouts, and unsignalized crossings, and expand multi-agent modeling to include surrounding vehicles and diverse pedestrian populations. More sophisticated reward formulations and higher-fidelity physics integration will further improve behavioral realism. Embedding the learned policies into closed-loop ADAS pipelines represents the path toward real-world autonomous vehicle validation.

Rights

In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/ This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).

DOI

10.25777/fgtt-sa94

ISBN

9798193214298

ORCID

0009-0002-4852-5597

Download

Share

COinS