Pedestrian Movement Simulation through Augmented Potential Field Model

Abstract

This thesis examines the use of a potential field model for simulating pedestrian dynamics in complex environments. The study first reviewed the different types of pedestrian dynamics models, highlighting their strengths and weaknesses. From this study, a research gap emerged regarding hybrid pedestrian movement models. As a base for such a model, a microscopic pedestrian dynamics model has been developed combining potential fields and gradient descent optimization as the drivers for trajectory selection. In this formulation, agents follow trajectories along the gradient of the potential field, naturally balancing goal seeking behavior with obstacle and inter-agent avoidance. The potential field approach was also discussed as a foundation for hybrid models, in which microscopic and macroscopic modeling strategies are combined to exploit the advantages of both. To evaluate and calibrate the model, real-world trajectory data from a bidirectional corridor experiment and a bottleneck experiment were used. A surrogate model was constructed to accelerate the optimization process, given the high computational cost of the original simulation model. The surrogate model enabled systematic parameter calibration and sensitivity analysis, focusing on three key parameters: the goal potential function weight (KG), the wall potential function weight (KW), and the obstacle potential function weight (KO). The results demonstrated that the optimized model is capable of reproducing key crowd phenomena observed in the empirical datasets. A sensitivity analysis further showed the relative importance of the potential function weights across different key performance indicators. Moreover, predictive uncertainty analysis confirmed that the model exhibited relatively high confidence around the optimum and avoided regions of overfitting. Despite these contributions, the research was constrained by the computational cost of the simulation model. The reliance on a surrogate model limited the optimization to a small subset of parameters, assuming that other model parameters were already sufficiently calibrated. This assumption likely introduced some biases, such as underestimation of obstacle repulsion, leading to overly frequent close inter-agent encounters. In conclusion, this thesis has demonstrated that potential field models, when combined with real-world data and surrogate-based optimization, provide a valid and powerful framework for simulating pedestrian dynamics in complex environments. Their ability to model pedestrian trajectories through potential functions and gradient descent makes them conceptually simple yet effective, while their extensibility offers a pathway toward hybrid models. Nevertheless, computational burden and limited parameter coverage remain key challenges, highlighting the need for more efficient implementations and broader parameter optimization in future research.