Real-Time Detection and Mapping of Crowd Panic Emergencies

Ilias Lazarou, Anastasios L. Kesidis and Andreas Tsatsaris

Department of Surveying and Geoinformatics Engineering, University of West Attica, Athens, Greece

Keywords: Panic Detection, Biometrics, Machine Learning, Classification, Real-Time Data.

Abstract: We present a real-time system that uses machine learning and georeferenced biometric data from wearables

and smartphones to detect and map crowd panic emergencies. Our system predicts stress levels, tracks stressed

individuals, and introduces the CLOT parameter for better noise filtering and response speed. We also

introduce the DEI metric to assess panic severity. The system creates dynamic areas showing the evolving

panic situation in real-time. By integrating CLOT and DEI, emergency responders gain insights into crowd

behaviour, enabling more effective responses to panic-induced crowd movements. This system enhances

public safety by swiftly detecting, mapping, and assessing crowd panic emergencies.

1 INTRODUCTION

Crowd panic emergencies are a significant public

safety concern, particularly in densely populated

areas like cities, sports events, concerts, and festivals.

These incidents can result in injuries, fatalities, and

property damage, often triggered by perceived

threats, rumors, or stampedes. Real-time detection

and mapping of such emergencies are vital for swift

response and evacuation.

Recent advancements in machine learning and

wearable technology offer new opportunities for real-

time detection and mapping. Our system utilizes

georeferenced bio-metric data from wearables and

smartphones, providing more accurate insights into

stress levels and movement patterns. It employs a

Gaussian SVM machine learning classifier to identify

stressed individuals. We introduce the Classifier

Level of Trust (CLOT) as a parameter to balance

detection speed and noise filtering.

Once a stressed individual is detected, the system

conducts real-time spatial analysis to track their

movement and identify nearby stressed individuals. It

creates dynamic areas based on trajectories and

adjacency. The system also introduces the Domino

Effect Index (DEI) to assess the severity of the

emergency by considering factors like panic

transmission rate, panicked crowd density, and

alignment with road networks.

Incorporating DEI enhances emergency detection

and response, ensuring public safety in densely

populated areas. Emergency responders can use this

information to de-ploy resources, evacuate affected

areas, and prevent escalation. The system's

components, including the machine learning

classifier and georeferencing, are detailed in

subsequent sections, along with an evaluation of its

effectiveness and potential applications. We also

outline future research directions in this field.

2 RELATED WORK

Panic, extensively studied in psychology and human

sciences, involves intense fear resulting from real or

perceived danger. It often occurs in groups or crowds,

leading to regressive behaviors like violence, jumps,

or collective suicide. Mass panic is an abnormal

response where a group moves faster than usual due

to alarming events like stampedes, fires, fights,

robberies, or riots.

In recent literature, several studies and systems

have concentrated on panic detection through the

utilization of Closed Circuit Television (CCTV)

technology. These surveillance methods scrutinize

human behavior by analyzing both still images and

video sequences of individuals or groups. For

instance, Hao et al. (Hao, 2016) have presented an

approach based on optical flow features to identify

crowd panic behavior, while Ammar et al. (Ammar,

2021) have outlined a continuous surveillance system

for a particular public location, employing a

stationary camera and a methodology for real-time

analysis of captured images.

Lazarou, I., Kesidis, A. and Tsatsaris, A.

Real-Time Detection and Mapping of Crowd Panic Emergencies.

DOI: 10.5220/0012372200003660

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 4: VISAPP, pages

485-491

ISBN: 978-989-758-679-8; ISSN: 2184-4321

485

Another approach to panic detection systems

involves user intervention and community

engagement in reporting emergency events. While

disaster preparedness plans are crucial for community

safety, traditional methods of data acquisition and

distribution fall short, especially during time-

sensitive crises.

The Internet of Things (IoT) technology emerges

as a solution to acquire real-time da-ta and promptly

transmit it to experts for decision-making. Wearable

devices and IoT play a pivotal role in collecting

biometric data and conducting stress detection. The

wearables and IoT sector has seen exponential

growth, thanks to technological advancements in

sensors and chips. This growth allows real-time

sensor data to be combined with the capabilities of 5G

smartphones, providing essential information for

decision-making.

Recent research shows that the field of crowd

evacuation systems, quantitative analysis, and

visualization is still evolving. Notable contributions

include Tsai's work (Tsai, 2022), which uses

wearable data to predict panic attack disorders based

on time series data, incorporating physiological

factors and air quality into a prediction model.

Kutsarova and Matskin (Kutsarova, 2021)

employ mobile crowdsensing and wearables on the

CrowdS platform, utilizing smartwatch sensors to

detect abnormal events and trigger alarms. Alsalat's

research (Alsalat, 2018) focuses on using machine

learning with wearables to classify individuals as

stressed or calm during panic situations.

Sun et al. (Sun, 2021) address crowd behavior

during emergencies, particularly in earthquake

evacuations. They conducted an evacuation drill

experiment to analyze evacuation processes,

participation ratios, and behavior characteristics.

Their study includes a computer-aided quantitative

simulation, establishing a response rule equation for

crowds in emergencies, exploring panic behavior,

exit familiarity, and the relationship between training

time and exit familiarity. The study aims to optimize

the efficiency of evacuation processes and prevent

congestion and stampede accidents.

These studies collectively contribute to our

understanding of crowd panic and emergency

response, pushing the boundaries of current research

in this field.

In a related study, Zhang et al. (Zhang, 2023)

address the challenges of urban security and

management concerning crowd gatherings in large

public spaces like shopping malls, stations, and

entertainment venues. They propose a Crowd Density

Estimation Model (CDEM-M) that utilizes deep

learning and Geographic Information System (GIS)

technology. This model surpasses the limitations of

traditional crowd density estimation methods that rely

on human head features, which can be problematic in

high-altitude scenes or when head information is

obscured. The CDEM-M provides a comprehensive

solution by integrating GIS, offering a unified map

visualization interface for accurate crowd area ex-

traction through semantic segmentation. It considers

various aspects, including crowd information

extraction, geographic mapping, number estimation,

and map visualization.

Another study by Albarakt et al. (Albarakt, 2021)

explores the role of public spaces in cities, focusing

on their political, social, economic, and sustainability

aspects. The research investigates how streets,

commercial centers, squares, and cafes either support

or restrict public engagement. It also delves into the

evolving political use of public spaces, the

contestation over space, and the competition among

various stakeholders for dominance. Using examples

from the Middle East and ArcGIS mapping, the study

examines visual and verbal narratives of protest

events in contested public spaces. The findings have

potential implications for urban planning and

management strategies related to public spaces.

In conclusion, these studies illustrate the potential

of utilizing machine learning and sensor data for real-

time detection and mapping of crowd panic

emergencies. Each paper offers a distinct approach,

utilizing various data types and machine learning

algorithms.

Our proposed system builds upon this prior

research by leveraging georeferenced biometric data

from wearable devices and smartphones, employing

a Gaussian SVM machine learning classifier for the

real-time detection and mapping of crowd panic

emergencies.

This represents a significant advancement, as it

utilizes precise data, offering a more accurate

assessment of stress levels and panic behavior

compared to traditional data sources like GPS or

video. Additionally, our system conducts real-time

spatial analysis to monitor the movement of stressed

individuals and generate dynamic areas, providing

emergency responders with accurate, up-to-date

information about the situation.

In essence, our research takes a comprehensive

and precise approach to the real-time detection and

mapping of crowd panic emergencies, enabling

emergency responders to make faster, more informed

decisions that mitigate risks and ensure public safety.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

486

3 METHODOLOGY

3.1 Workflow Process

Our crowd panic detection system aims to extract

insights from collected biometric and spatiotemporal

data to identify panic patterns in crowds, as shown in

Figure 1. The process begins with a wearable device

monitoring biometric data, while an Android

smartphone collects GPS coordinates, time, activity,

speed, and step data. This information is compiled

into encrypted UDP packets and sent to a server over

the GSM network. The server decrypts and processes

the data to identify panic patterns, handling a

significant volume of real-time data.

3.2 Stress Profile Index (SPI)

Classification

The proposed method characterizes an individual as

calm or in a panic state using a classifier that takes

various biometric and geospatial data from wearable

devices as input, as described in (Lazarou, 2022). To

select the most suitable machine learning classifier,

several classifiers were assessed with a dataset

comprising of 27 subjects. This dataset includes

biometric information such as heart rate, heart rate

variability, spatiotemporal data including location

coordinates, activity type, subject speed, step count,

and descriptive data like gender, age, weight, and a

unique identification code for each subject.

The biometric and spatiotemporal attributes in the

dataset are categorized into four groups, with values

informed by relevant studies: i) biometric data from

wearables, including heart rate and heart rate

variability; ii) spatiotemporal data from smartphones,

which includes location coordinates, activity type,

subject velocity, and step count; iii) descriptive data

from wearables, encompassing subject gender, age,

and weight; and iv) the unique ID assigned to each

subject from smartphones. Additionally, a feature

called "heart rate moving average deviation"

(HRMAD) is introduced to detect sudden panic

conditions based on heart rate values.

Machine learning models are trained on this

dataset to differentiate between panic states and

normal behavior. Various classifiers, including

decision trees, logistic regression, Gaussian and

kernel naïve Bayes, Gaussian SVM, SVM kernel, and

boosted trees, were examined. The Gaussian SVM

classifier yielded the highest accuracy, especially

when using the HRMAD60 feature. Consequently,

the Stress Profile Index (SPI) is introduced as a binary

index, indicating a Calm state (value 0) or a Stressed

state (value 1) based on the classifier's output

(Lazarou, 2022)

3.3 Real-Time Analysis of Spatial

Patterns

The purpose of real-time spatial analysis in

monitoring panic conditions is supported by a data

model as in (Lazarou, 2023), represented in Figure 3.

This model processes streaming data containing

spatiotemporal and biometric information collected

from wearable devices and smartphones. As stated in

the previous section, a Gaussian SVM machine

learning classifier is utilized to distinguish between

normal behavior and panic conditions, assigning the

SPI values of 0 and 1, respectively. The resulting

categorization labels the data as either Points of No

Interest or Panic Points.

Real-time spatial analysis for panic monitoring

relies on a data model, as illustrated in Figure 3 and

detailed in (Lazarou, 2023). This model processes

streaming data, combining spatiotemporal and

biometric information from wearables and

smartphones. A Gaussian SVM classifier discerns

normal behavior from panic, assigning SPI values of

0 and 1, respectively. The data is categorized as

"Points of No Interest" (SPI 0), marking the end of a

sequence of "Panic Points" (SPI 1) representing

highly stressed profiles. If isolated Panic Points are

followed by a Point of No Interest, no further action

is taken. However, consecutive Panic Points form a

"Panic Trajectory" with an associated "Panic

Trajectory Origin."

A Panic Trajectory is a continuous sequence of

Panic Points linked to a subject, ending with one or

more Points of No Interest. The initiation of a Panic

Trajectory can depend on a single or multiple panic

points, with the required initiation and termination

points determined by the "Classifier Level of Trust"

(CLOT).

In Section 4, we delve into various start and end-

point scenarios for Panic Trajectories, exploring

variations where two or more Points of No Interest

are needed to conclude a Panic Trajectory. We also

examine scenarios requiring two or more Panic Points

for initiation. Once a Panic Trajectory begins, the first

point becomes the "Panic Trajectory Origin." We use

the DBSCAN algorithm to identify spatiotemporal

correlations among these origins. This algorithm

works within a 100-meter radius and a 10-second

timeframe, aiding our understanding of panic

behavior patterns.

Meeting specific conditions triggers the creation

of "Crowd Panic Areas," comprising the "Origin

Real-Time Detection and Mapping of Crowd Panic Emergencies

487

Crowd Panic Area" and the "Current Location Crowd

Panic Area." The Origin CPA traces the origin of

correlated Panic Trajectory starting points, while the

Current Lo-cation CPA relies on the most recent

points of ongoing correlated Panic Trajectories.

Additionally, the "Domino Effect Index,"

introduced later, assesses the severity of panic-

induced crowd behavior during emergencies.

3.4 Classifier Level of Trust (CLOT)

CLOT, a numerical parameter from 0 to 10, indicates

the system's confidence in the classifier's output.

Lower CLOT values prioritize fast detection with less

noise reduction, while higher values filter out more

noise, reducing false positives but slowing detection.

In essence, adjusting CLOT balances detection

speed and noise filtering, enabling performance

testing under different settings and noise levels.

Figure 1: Example of CLOT = 3.

In Figure 1, two examples highlight the influence

of a CLOT value set at 3. On the top, a subject initially

exhibits calmness with two Points of No Interest.

Then, a sequence of Panic Points unfolds, triggering

the system to mark the third successive Panic Point as

the Point of Trust (POT), initiating a Panic

Trajectory. If the sequence continues uninterrupted,

the trajectory extends. Points of No Interest

eventually appear, and the system assesses if at least

three consecutive Points of No Interest are present to

end the Panic Trajectory. In the bottom example,

another subject remains composed, and the

subsequent Panic Points don't surpass the CLOT

threshold of 3. As a result, the system classifies them

as noise, leading to no trajectory formation.

3.5 Domino Effect Index (DEI)

The DEI assesses panic severity by considering

factors such as panic transmission rate, panicked

population density, new panic origins, convex hull

area change rates, and alignment with the road

network. It's rated from 0 to 5, with higher values

indicating more severe panic. This scale has five

levels, with DEI scale 1 being the lowest severity, and

DEI scale 5 indicating the highest severity. By

incorporating various factors contributing to the

domino effect, DEI offers a dependable evaluation of

crowd panic, aiding decision-makers in shaping

effective emergency response strategies.

Methodologically, DEI is determined by a

combination of weighted and normalized factors

influencing panic propagation, detailed in Table 2

below:

Table 1: DEI contributing factors.

Factor Description

Rate of panic

transmission (𝑓



)

The rate at which panic spreads

among the crowd

Number of new panic

origins within the

panic origin convex

hull (𝑓



)

The distribution of new panic

origins within the area where panic

first emerged

Density of panicked

people (𝑓



)

The concentration of panicked

individuals within the current

location convex hull

Area change rate of

the panic origin

convex hull (𝑓



)

The rate at which the area of the

panic origin convex hull changes

over time

Area change rate of

the current location

convex hull (𝑓



)

The rate at which the area of the

current location convex hull

changes over time

Number of aligned

clusters (𝑓



)

The count of panic clusters aligned

with the road network, which

might indicate the crowd's

tendency to use streets for escape

Each factor is normalized between 0 and 1, and

then multiplied by a weight that reflects its

importance in contributing to the domino effect. The

DEI is then calculated as the sum of these weighted

factors:

DEI =

∑

𝑤



𝑓



for i =1…6

where 𝑓



and 𝑤



denote the 𝑖 -th factor and the

corresponding weight, respectively.

To normalize the contributing factors for DEI,

each factor is scaled between 0 and 1, ensuring fair

comparisons and combining different numerical

values. This process involves three steps. First,

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

488

determining the factor's minimum and maximum

values to set its range. Second, scaling the current

factor value at a time step to a normalized value

within 0 to 1 by subtracting the minimum and

dividing by the range between the maximum and

minimum values.

normalized_value = (current_value - min_value) /

(max_value - min_value)

To compute DEI, normalization ensures that

various factors, regardless of their original scales, are

equitably assessed for their collective influence on the

domino effect's severity. Normalized values are then

weighted by user-defined weights and summed to

determine the final DEI value. This quantifies the

potential panic propagation extent in a crowd and aids

in intervention prioritization. The DEI value is

classified into five intervals (0-0.2, 0.2-0.4, 0.4-0.6,

0.6-0.8, 0.8-1), with each interval corresponding to

DEI scales from 1 to 5, as shown in Table 3.

Table 2: DEI scales.

DEI Scale DEI value

1 0-0.2

2 0.2-0.4

3 0.4-0.6

4 0.6-0.8

5 0.8-1

DBSCAN clustering is employed to identify

panicked individual clusters based on their alignment

with the road network. DBSCAN, a widely used

density-based clustering algorithm, identifies dense

regions in datasets. Each cluster is enclosed by a

minimum area bounding rectangle (MABR), and the

axis ratio is calculated. If the axis ratio is below a

certain threshold (e.g., 0.5), it is deemed an aligned

cluster. This information is valuable, suggesting that

panic transmission is more likely when a significant

portion of a panicked crowd flees through the streets.

The DEI metric and its scale are valuable for assessing

panic severity in real-world scenarios like evacuations,

natural disasters, or terrorist attacks. By quantifying the

domino effect and categorizing it into five severity

levels, emergency planners and responders can better

understand crowd behavior and develop more effective

response strategies to mitigate risks associated with

panic-induced crowd movements.

4 EXPERIMENTAL SETUP AND

RESULTS

We conducted a proof of concept in Syntagma

Square, Athens, testing three unique crowd panic

scenarios: ESCAPE, SHRINK, REPULSION. These

scenarios were designed to examine different crowd

panic behaviors and DEI dynamics.

In these scenarios, the crowd responds to aversive

events by dispersing (ESCAPE), contracting towards

the center (SHRINK), or reacting to repulsive forces

(REPULSION). Weight variations in each scenario

were applied to analyze the DEI factors' impact on

crowd behavior, contributing to a better

understanding of panic propagation.

Regarding the ESCAPE scenario that will be

presented in this paper, approximately 30 individuals

from diverse backgrounds gather in a controlled

environment, initially in a calm state, engaging in

various activities. At a predetermined moment, an

unpleasant event is deliberately introduced, causing a

sudden onset of stress and panic among some

participants. This event triggers physiological

symptoms like increased heart rate and rapid

breathing. As panic spreads, individuals' emotions

influence each other, resulting in a chain reaction of

stress and anxiety. This phenomenon is known as

emotional contagion, where emotions transfer

between people through nonverbal cues and social

interactions. Those initially calm also become

stressed as they observe the panic. As the situation

unfolds, panic continues to propagate, with

individuals instinctively seeking escape in various

directions. This amplifies the scale and magnitude of

the event.

In Figures 2, 3, and 4, the maps illustrate the

progression of the phenomenon over time. Panic

points are depicted as red dots, calm points as green

dots, and recovered points as blue dots. Panic

trajectories are represented by red lines, while the

origins of these trajectories are marked by green flags.

The shaded orange region denotes the Origin CPA

(Common Panic Area), and the hollow red region

indicates the Current Location CPA.

Figure 2: Initial expansion.

Real-Time Detection and Mapping of Crowd Panic Emergencies

489

Figure 3: Panic starts to spread widely.

Figure 4: After some time it still expands but some subjects

tend to recover (blue dots).

Figures 5, 6, 7, and 8, depict real-time counts of

individuals categorized as stressed, calm, and

recovered. These visualizations facilitate the effective

monitoring of emotional distribution within the

group. Real-time charts graphically represent

emotional trends for each category, helping identify

influencing factors and individual transitions between

emotional states.

Furthermore, the Panic Transmission and

Recovery Rate Charts offer insights into the speed of

panic propagation and recovery rate, providing

valuable information about the effectiveness of

interventions and the overall resilience of the group.

Additionally, the DEI Current Value offers real-time

insights into the collective emotional state, reflecting

stress and anxiety levels. The DEI Progress Diagram

tracks the evolution of the emotional state over time,

providing valuable information about its progression

throughout the scenario.

Figure 5: Transmission rate and panicked population.

The count of recovered individuals demonstrates

a progressive increase after a certain period, as

evidenced by the recovery rate. Simultaneously, the

number of calm individuals exhibits a noticeable

decline, gradually approaching zero.

Figure 6: Recovery rate and recovered population.

Figure 7: Calm population.

Ultimately, the comprehensive evaluation of the

DEI reveals that, in this particular scenario, the

phenomenon only marginally surpasses the threshold

of 0.40, resulting in a DEI scale of 2.

Figure 8: Evolution of DEI.

In Figure 9, it is evident that the population of

panicked individuals exhibits considerable

fluctuations over time, indicating the arbitrary nature

of the phenomenon's expansion and its variable

impact on different individuals. During the initial

minutes, the transmission rate remains predominantly

low, as the panic has yet to propagate to a wider

population. However, in subsequent stages, the

transmission rate reaches higher values, signifying

the widespread dissemination of panic.

Figure 9: Final state where the event has spread

significantly, and multiple subjects are now in the recovery

phase.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

490

5 CONCLUSIONS AND FUTURE

WORK

In our experiments, we closely monitored participants

to understand panic behavior in groups. We used a

digital map to visualize how panic evolves,

identifying clusters of stressed individuals and

support networks. The Domino Effect Index (DEI) is

a vital tool for assessing emergency severity. It

considers panic speed, density, and road alignment.

The Classifier Level of Trust (CLOT) balances noise

filtering and quick detection. Our research can shape

interventions for managing panic in real-life

situations, reducing negative consequences. In

conclusion, our real-time spatial analysis, using

wearables and smartphones, advances crowd panic

monitoring. serves as a valuable index for prioritizing

interventions in scenarios characterized by

concurrent multiple events. Empirical validation of

this approach has been substantiated through rigorous

experimental investigations. Future work will delve

into bio-algorithms and mathematical models to

better understand panic spread, refining our approach

in crowd safety and security.

REFERENCES

Hao, Y.; Xu, Z.; Wang, J.; Liu, Y.; Fan, J. An Approach to

Detect Crowd Panic Behavior using Flow-based

Feature. In Proceedings of the 22nd International

Conference on Automation and Computing, Colchester,

UK, 7–8 September 2016; ISBN 9781862181328.

https://doi.org/10.1109/iconac.2016.7604963.

Lazarou, I.; Kesidis, A. L.; Hloupis, G.; Tsatsaris, A. Panic

Detection Using Machine Learning and Real-Time

Biometric and Spatiotemporal Data. ISPRS Int. J. Geo-

Inf. 2022, 11(11), 552.

Lazarou, I.; Kesidis, A.; Tsatsaris, A. Real-Time

Monitoring of Crowd Panic Based on Biometric and

Spatiotemporal Data. In Proceedings of the 18th

International Joint Conference on Computer Vision,

Imaging and Computer Graphics Theory and

Applications - Volume 5: VISAPP, SciTePress, 2023;

pp. 1021-1027. DOI: 10.5220/0011789900003417

Ammar, H.; Cherif, A. DeepROD: A deep learning

approach for real-time and online detection of a panic

behavior in human crowds. Mach. Vis. Appl. 2021, 32,

57. https://doi.org/10.1007/s00138-021-01182-w.

Tsai, C. H.; Chen, P. C.; Liu, D. S.; Kuo, Y. Y.; Hsieh, T.

T.; Chiang, D. L.; ... & Wu, C. T. Panic Attack

Prediction Using Wearable Devices and Machine

Learning: Development and Cohort Study. JMIR Med.

Inform. 2022, 10(2), e33063.

Kutsarova, V.; Matskin, M. Combining Mobile

Crowdsensing and Wearable Devices for Managing

Alarming Situations. In 2021 IEEE 45th Annual

Computers, Software, and Applications Conference

(COMPSAC), pp. 538-543. IEEE, 2021.

Alsalat, G. Y.; El-Ramly, M.; Fahmy, A. A.; Karim, S.

Detection of Mass Panic using Internet of Things and

Machine Learning. Int. J. Adv. Comput. Sci. Appl.

2018, 9(5).

Sun, H.; Hu, L.; Shou, W.; Wang, J. Self-Organized Crowd

Dynamics: Research on Earthquake Emergency

Response Patterns of Drill-Trained Individuals Based

on GIS and Multi-Agent Systems Methodology.

Sensors 2021, 21, no. 4: 1353. https://doi.org/

10.3390/s21041353

Zhang, X.; Sun, Y.; Li, Q.; Li, X.; Shi, X. Crowd Density

Estimation and Mapping Method Based on

Surveillance Video and GIS. ISPRS Int. J. Geo-Inf.

2023, 12, no. 2: 56. https://doi.org/10.3390/

ijgi12020056

Albarakt, R.; Selim, G.; Iaaly, A. Reshaping Riyadh Alsolh

Square: Mapping the Narratives of Protesting Crowds

in Beirut. Urban Reg. Plan. 2021, 6, no. 4, 126-133. doi:

10.11648/j.urp.20210604.13.

Real-Time Detection and Mapping of Crowd Panic Emergencies

491