Video-Based Parameter Calibration and Virtual Simulation of Pedestrian Crowd Dynamics

Abstract

Pedestrian safety and efficient crowd movement require measurable, data-driven methods that ground simulation models in real-world observations. Despite the widespread use of pedestrian simulation tools, a significant challenge remains: model parameters are often adopted from controlled studies and may not reflect site-specific conditions, while the resulting model insights are difficult to communicate to practitioners and students through traditional methods. This thesis bridges this gap through two integrated research contributions. First, we develop a video-to-simulation framework that extracts pedestrian trajectories from real-world video using deep learning-based detection and homography-based perspective correction, then calibrates the Social Force Model parameters to match observed behavior at the Virginia Tech Drillfield-an open outdoor environment with unconstrained bidirectional movement. The calibrated parameter set is validated through trajectory reconstruction and crowd-flow simulation, providing transferable reference parameters for future campus-based pedestrian research. Second, we develop a virtual reality learning platform that implements these calibrated SFM parameters in an immersive environment, enabling students to explore pedestrian dynamics interactively through parameter adjustment and real-time visualization. Beyond education, the platform serves as an extensible research foundation supporting future investigations such as evacuation modeling and exit design optimization. Together, these contributions establish a complete methodology bridging empirical observation, data-driven calibration, accessible learning, and extensible research capabilities in pedestrian crowd dynamics.