 ICACT20260308 Slide.01
[Big Slide]
|
Chrome  Click!! |
 |
Hello. My name is Jung-Hyun Chang from Kyonggi University.
This presentation introduces a Crowd Pressure Index framework for early detection of crush hazards in high-density crowd environments.
|
| ICACT20260308 Slide.02
[Big Slide]
|
Chrome Click!! |
 |
This research is motivated by the Itaewon crowd disaster that occurred on October 29, 2022, which highlighted the severe risks of unmanaged crowd density.
|
| ICACT20260308 Slide.03
[Big Slide]
|
Chrome Click!! |
 |
This type of accident is classified as a crowd disaster, which occurs under extremely high-density crowd conditions.
|
| ICACT20260308 Slide.04
[Big Slide]
|
Chrome Click!! |
 |
Recent studies on crowd crush risk have shown that crowd density can be used as a key indicator of dangerous conditions.When density approaches around eight persons per square meter, crowd pressure waves begin to emerge and involuntary physical contact increases.
As density exceeds ten persons per square meter, severe overcrowding occurs, leading to a critical loss of mobility.
Based on these findings, this study adopts a deep learning–based approach using the P2PNet algorithm to estimate crowd density and extract spatial key points from CCTV footage, which serve as the foundation for subsequent crowd pressure analysis.
|
| ICACT20260308 Slide.05
[Big Slide]
|
Chrome Click!! |
 |
In disaster and safety management, the HSE framework is a well-established indicator used to evaluate risk from three perspectives: Health, Safety, and Environment.
The HSE indicator provides an integrated performance metric that quantitatively and qualitatively assesses safety conditions within an organization, a system, or a physical space.
In crowd environments, safety-related HSE indicators include physical risk factors such as crowd density, velocity dispersion, and counterflow occurrence rates, as well as incident signals like collision frequency and the duration of congestion hotspots.
System-level safety factors, including CCTV coverage and the response time of warning and control systems, are also critical components of the HSE framework.
|
| ICACT20260308 Slide.06
[Big Slide]
|
Chrome Click!! |
 |
This slide introduces two fundamental concepts: crowd density and crowd pressure.
First, crowd density is defined as the number of individuals forming the crowd divided by the occupied area, measured in square meters.
As crowd density increases, physical contact between individuals becomes unavoidable, and this directly leads to the generation of crowd pressure.
Crowd pressure can be categorized into three types.
Aggregate crowd pressure refers to the total force exerted by the crowd per unit wall length.
Frontal pressure is the force transmitted in the same direction as an externally applied force acting on the crowd.
Lateral pressure represents force propagation perpendicular to the applied direction, which is a key mechanism behind pressure waves in dense crowds.
|
| ICACT20260308 Slide.07
[Big Slide]
|
Chrome Click!! |
 |
This slide explains the process and characteristics of crowd wave phenomena caused by high crowd density and crowd pressure.
The analysis is based on YouTube video footage captured by a fixed CCTV camera over a four-second interval, which allows us to observe rapid dynamic changes in the crowd.
As pedestrians continuously arrive at a key location, local congestion begins to form. When the inflow of people exceeds the outflow, density increases and movement becomes constrained.
Under these conditions, crowd pressure accumulates and is transmitted laterally through physical contact, generating wave-like motions within the crowd.
These crowd waves are characterized by synchronized body oscillations and pressure propagation, even when individuals are not intentionally moving.
This phenomenon indicates a transition from static congestion to a dynamic and unstable crowd state, which significantly increases the risk of crowd collapse.
|
| ICACT20260308 Slide.08
[Big Slide]
|
Chrome Click!! |
 |
This slide shows how crowd disasters develop step by step.
First, crowd stagnation occurs as people continue to enter a limited space.
When density exceeds about eight persons per square meter, movement becomes restricted and pressure starts to build up.
If inflow continues and density goes beyond ten persons per square meter, pressure rapidly spreads through the crowd and individuals lose mobility.
At this stage, accidents become unavoidable.
|
| ICACT20260308 Slide.09
[Big Slide]
|
Chrome Click!! |
 |
This slide shows the moment when crowd disasters become fatal.
Actual deaths do not occur during crowding itself, but during crowd collapse.
When physical support between individuals breaks down, a single fall creates empty space, pressure rushes in, and people are piled up.
Once crowd collapse begins, fatalities occur within seconds.
|
| ICACT20260308 Slide.10
[Big Slide]
|
Chrome Click!! |
 |
Before explaining our previous work, I will briefly introduce the original DBSCAN algorithm.
DBSCAN is a density-based clustering algorithm that groups data points which are located close to each other.
It defines clusters using two parameters: epsilon, which represents the neighborhood radius, and MinPts, which is the minimum number of points required within that radius.
Points that satisfy this density condition form a cluster, while points that do not belong to any cluster are classified as noise.
|
| ICACT20260308 Slide.11
[Big Slide]
|
Chrome Click!! |
 |
This figure is a simple simulation designed to illustrate how we adapt the DBSCAN algorithm for crowd collapse detection.
In this concept, areas that are not assigned to any cluster by DBSCAN are interpreted as unstable regions, which correspond to potential crowd collapse zones.
This visualization is intended to support conceptual understanding rather than represent an exact physical model.
|
| ICACT20260308 Slide.12
[Big Slide]
|
Chrome Click!! |
 |
Many studies and systems have already been introduced to improve crowd safety.
These include simulation-based approaches such as the Social Force Model and pedestrian flow models, as well as proxemics-based theories focusing on interpersonal distance.
In addition, several commercial and AI-based systems use CCTV data for crowd monitoring and risk analysis.
However, most of these approaches focus on density or flow, rather than directly capturing crowd collapse dynamics.
|
| ICACT20260308 Slide.13
[Big Slide]
|
Chrome Click!! |
 |
Many studies and systems have already been introduced to improve crowd safety.
These include simulation-based approaches such as the Social Force Model and pedestrian flow models, as well as proxemics-based theories focusing on interpersonal distance.
In addition, several commercial and AI-based systems use CCTV data for crowd monitoring and risk analysis.
However, most of these approaches focus on density or flow, rather than directly capturing crowd collapse dynamics.
|
| ICACT20260308 Slide.14
[Big Slide]
|
Chrome Click!! |
 |
To calibrate the model, we applied our deep learning framework to real disaster cases, including the 2021 Meron and 2022 Seoul Halloween events.
Using P2PNet, head-level spatial coordinates were extracted from CCTV footage, providing key parameters such as inter-personal distance, cluster compactness, and local density gradients.
By comparing these cases, we identified a critical transition threshold at approximately nine to ten persons per square meter, just before physical crowd collapse.
These results confirm that the proposed CPI can function as an early warning indicator for dangerous crowd conditions.
|
| ICACT20260308 Slide.15
[Big Slide]
|
Chrome Click!! |
 |
We validated the proposed model through three complementary evaluations.
First, cross-event comparative analysis was conducted by comparing CPI trends across different disaster cases to confirm consistency under varying crowd conditions.
Second, sensitivity analysis was performed by adjusting the weight parameters alpha and beta to test the robustness of CPI under different density and motion patterns.
Finally, temporal stability was evaluated by tracking CPI frame by frame, ensuring stable predictions even during abrupt density changes.
|
| ICACT20260308 Slide.16
[Big Slide]
|
Chrome Click!! |
 |
Our validation results show clear and consistent density-dependent crowd behavior patterns.
Under normal flow conditions, with density below four persons per square meter, individuals moved freely with minimal physical interaction.
As density increased to between five and eight persons per square meter, mobility became restricted and body contact became unavoidable.
In the high-risk zone, between eight and ten persons per square meter, lateral pressure waves emerged, indicating growing crowd instability.
At densities above twelve persons per square meter, voluntary movement completely ceased, resulting in full immobilization.
These results are consistent with established crowd dynamics models and confirm that density alone can act as an early proxy for elevated risk.
|
| ICACT20260308 Slide.17
[Big Slide]
|
Chrome Click!! |
 |
Our second result focuses on inter-personal spacing as a micro-level indicator of crowd compression.
We measured the average distance between each individual and their nearest neighbors, which captures physical compression independent of overall crowd density.
In the Meron 2021 case, spacing decreased to approximately 0.12 to 0.15 meters shortly before collapse, while in the Itaewon 2022 case, it dropped further to 0.08 to 0.10 meters, indicating extreme physical restriction.
Frame-by-frame analysis showed that spacing declined more sharply than density as collapse approached, suggesting that inter-personal distance is a more sensitive early indicator of crush onset.
|
| ICACT20260308 Slide.18
[Big Slide]
|
Chrome Click!! |
 |
Our third result evaluates the interpretability and validity of the Crowd Pressure Index.
The CPI combines crowd density and inter-personal spacing using empirically tuned weights, where the inverse of spacing captures the nonlinear increase in pressure as distance decreases.
Based on CPI values, crowd states can be clearly classified into stable, compression, crowd-wave, and collapse-imminent phases.
In the Meron 2021 case, CPI reached approximately eighty, indicating an active crowd-wave regime.
In contrast, the Itaewon 2022 case exceeded one hundred, corresponding to a collapse-critical state.
These results closely match the physical conditions observed in real footage, validating CPI as a quantitative measure of crushing pressure.
|
| ICACT20260308 Slide.19
[Big Slide]
|
Chrome Click!! |
 |
Frame-level analysis revealed a clear four-stage CPI progression from stable flow to collapse.
Most importantly, CPI increased rapidly within five to ten seconds before collapse, demonstrating strong potential as an early warning signal.
The system operates in real time, with low response latency and stable processing performance, confirming its practical applicability for crowd safety monitoring.
|
| ICACT20260308 Slide.20
[Big Slide]
|
Chrome Click!! |
 |
In this study, we proposed a CPI-based deep learning framework for early detection of crowd collapse using CCTV video data.
By integrating crowd density and interpersonal spacing, the model successfully captured critical transitions leading to crowd collapse.
Experimental results from real disaster cases confirmed that CPI increases sharply several seconds before physical collapse, enabling early warning.
This approach moves crowd safety analysis from post-event interpretation to proactive, real-time risk prevention.
|
| ICACT20260308 Slide.21
[Big Slide]
|
Chrome Click!! |
 |
In conclusion, this study demonstrates that crowd collapse is not a sudden event, but a process that can be quantitatively observed and predicted.
By leveraging deep learning and spatiotemporal crowd indicators, we believe this work contributes to safer public spaces and proactive disaster prevention.
Thank you very much for your attention.
I would be happy to answer any questions.
|
