Abstract
The escalating complexity of modern urban infrastructure and the growing frequency of emergency evacuations underscore the critical need for intelligent, adaptive evacuation guidance systems. This thesis advances the field of building evacuation management by addressing the limitations of conventional simulation-based direction-setting methods, particularly their computational inefficiency and oversimplified assumptions of full evacuee obedience. Through an integrated framework combining advanced simulator, graph-theoretic optimization, and machine learning technique, this work establishes novel methodologies to enhance the practicality and performance of Dynamic Evacuation Signage (DES) systems.To reduce computational time, this thesis introduces a modified Cell Transmission Model (CTM) for rapid evacuation prediction. This mesoscopic model discretizes the evacuation space into a network of cells, capturing the fundamental characteristics of pedestrian movement while considering pedestrian preferences at an aggregate level. This approach maintains computational efficiency while providing accurate predictions. The model is validated through experiments, demonstrating its ability to accurately predict evacuation times and crowd dynamics. Building on the improved CTM, a coupled CTM-Directed Rooted Forest (DRF) method is proposed for fast direction setting with minimal simulation iterations. This method encodes evacuation plans into a DRF structure, allowing for effective optimization through node reconnections that better utilize simulation outputs. Results show that the DRF method significantly reduces evacuation time by balancing the load across multiple exits. The effectiveness and efficiency of the proposed method are validated through numerical tests in a university canteen scenario, demonstrating its capability to handle large-scale scenarios.
Furthermore, this thesis considers the reality of partial obedience among evacuees and enhances the DRF-based heuristic optimization to couple with microscopic simulators that can reproduce partial obedience, thereby accounting for non-ideal conditions. The extended DRF-based method is validated through tests in the university canteen scenario with varying numbers and distributions of evacuees. Results indicate that the method can optimize direction-setting plans with minimal computational time, even under different crowd densities and non-ideal obedience levels. To further ensure the stability of the simulation-based optimization process, a machine learning-based pedestrian movement prediction model is developed using trajectory data collected from real-world unguided scenarios. During this process, fundamental issues related to feature collection perspectives are explored to ensure high accuracy and interpretability. The integration of this model with microscopic simulations not only better captures pedestrian dynamics under partial obedience but also mitigates simulation anomalies caused by partial obedience.
In summary, the proposed methods significantly improve computational efficiency, accuracy, and adaptability, making them more suitable for applying DES in large buildings. The presented framework not only overcomes existing constraints in evacuation planning but also supplies a flexible direction-setting approach. This approach allows for the integration of cutting-edge crowd psychology models and IoT-enabled dynamic sensing, thereby laying the foundation for intelligent building safety systems within smart cities.
| Date of Award | 21 Jul 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Wai Ming LEE (Supervisor) & Siu Ming LO (Supervisor) |