A socio-technical approach to crowd management: understanding decision-making and movement patterns with virtual reality and machine learning
File(s)
Author(s)
Yang, Xiangmin
Type
Thesis
Abstract
Crowd disasters, which are events in which deaths and injuries occur due to the high density of people, happen all too often. With greater urbanisation and mobility, the frequency of such disasters increases every year. Numerous attempts to improve crowd safety by better monitoring and prediction of crowds, have been hindered by the lack of systematic and integration of human behaviour and crowd social dynamics, relying instead upon heuristics.
This thesis tackles this gap by adopting a socio-technical perspective that emphasises the role of human actors and the surrounding social environment to contribute to a more comprehensive understanding of crowd behaviour.
Given that crowd disaster databases suffer from poor definition for their characteristics and fail to adequately incorporate the effect of socio-external environments, this thesis begins by constructing a comprehensive database of 293 past crowd disasters using data from news sources. Cluster analysis identified nine distinct types of crowd disasters, with the jostle disaster having the greatest frequency (20%).
A jostle scenario is reproduced in a virtual reality (VR) experiment, given the logistical and ethical difficulties associated in undertaking a physical experiment. Over 70 individuals participated in three rounds, with over 22 simultaneous participants. This experiment not only assesses the effectiveness of VR as a tool for analysing movement patterns in crowd analysis but also identifies two distinct behaviour types: ‘patient’ and ‘impatient’.
A decision-making inference framework that integrates inverse reinforcement learning and model interpretability is proposed. Data from the VR experiments enables quantification of the impact of numerous factors, including interactions with surrounding crowds, on the individual’s decision making. This framework is also tested on data from a case study in Nanjing. Impacts from built environment features on pedestrians are quantified, and comfort zones are identified.
This thesis concludes with discussion on the implications for improving crowd management practices.
This thesis tackles this gap by adopting a socio-technical perspective that emphasises the role of human actors and the surrounding social environment to contribute to a more comprehensive understanding of crowd behaviour.
Given that crowd disaster databases suffer from poor definition for their characteristics and fail to adequately incorporate the effect of socio-external environments, this thesis begins by constructing a comprehensive database of 293 past crowd disasters using data from news sources. Cluster analysis identified nine distinct types of crowd disasters, with the jostle disaster having the greatest frequency (20%).
A jostle scenario is reproduced in a virtual reality (VR) experiment, given the logistical and ethical difficulties associated in undertaking a physical experiment. Over 70 individuals participated in three rounds, with over 22 simultaneous participants. This experiment not only assesses the effectiveness of VR as a tool for analysing movement patterns in crowd analysis but also identifies two distinct behaviour types: ‘patient’ and ‘impatient’.
A decision-making inference framework that integrates inverse reinforcement learning and model interpretability is proposed. Data from the VR experiments enables quantification of the impact of numerous factors, including interactions with surrounding crowds, on the individual’s decision making. This framework is also tested on data from a case study in Nanjing. Impacts from built environment features on pedestrians are quantified, and comfort zones are identified.
This thesis concludes with discussion on the implications for improving crowd management practices.
Version
Open Access
Date Issued
2025-07-18
Date Awarded
2026-04-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Ochieng, Washington
Majumdar, Arnab
Grass, Emilia
Publisher Department
Department of Civil and Environmental Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)