ArticlePDF Available

Assessing crowd management strategies for the 2010 Love Parade disaster using computer simulations and virtual reality

The Royal Society
Journal of the Royal Society Interface
Authors:

Abstract and Figures

Dense crowds in public spaces have often caused serious security issues at large events. In this paper, we study the 2010 Love Parade disaster, for which a large amount of data (e.g. research papers, professional reports and video footage) exist. We reproduce the Love Parade disaster in a three-dimensional computer simulation calibrated with data from the actual event and using the social force model for pedestrian behaviour. Moreover, we simulate several crowd management strategies and investigate their ability to prevent the disaster. We evaluate these strategies in virtual reality (VR) by measuring the response and arousal of participants while experiencing the simulated event from a festival attendee’s perspective. Overall, we find that opening an additional exit and removing the police cordons could have significantly reduced the number of casualties. We also find that this strategy affects the physiological responses of the participants in VR.
Content may be subject to copyright.
royalsocietypublishing.org/journal/rsif
Research
Cite this article: Zhao H, Thrash T, Kapadia
M, Wolff K, Hölscher C, Helbing D, Schinazi VR.
2020 Assessing crowd management strategies
for the 2010 Love Parade disaster using
computer simulations and virtual reality.
J. R. Soc. Interface 17: 20200116.
http://dx.doi.org/10.1098/rsif.2020.0116
Received: 19 February 2020
Accepted: 13 May 2020
Subject Category:
Life SciencesEngineering interface
Subject Areas:
computational biology
Keywords:
crowd simulation, crowd disasters,
virtual reality, spatial cognition,
physiological arousal
Author for correspondence:
Hantao Zhao
e-mail: hantao.zhao@gess.ethz.ch
Electronic supplementary material is available
online at https://doi.org/10.6084/m9.figshare.
c.4994591.
Assessing crowd management strategies
for the 2010 Love Parade disaster using
computer simulations and virtual reality
Hantao Zhao1, Tyler Thrash1,4,5, Mubbasir Kapadia6, Katja Wolff2,
Christoph Hölscher1, Dirk Helbing3and Victor R. Schinazi1,7
1
Chair of Cognitive Science, ETH Zürich,
2
Interactive Geometry Lab, and
3
Computational Social Science,
ETH Zürich, Zurich, Switzerland
4
Geographic Information Visualization and Analysis, and
5
Digital Society Initiative, University of Zürich,
Zurich, Switzerland
6
Department of Computer Science, Rutgers University, Piscataway, NJ, USA
7
Department of Psychology, Bond University, Gold Coast, Queensland, Australia
HZ, 0000-0003-0398-3842; TT, 0000-0002-3011-7029; MK, 0000-0002-3501-0028;
CH, 0000-0002-7520-9146; VRS, 0000-0002-2345-2806
Dense crowds in public spaces have often caused serious security issues at large
events. In this paper, we study the 2010 Love Parade disaster, for which a large
amount of data (e.g. research papers, professional reports and video footage)
exist. We reproduce the Love Parade disaster in a three-dimensional computer
simulation calibrated with data from the actual event and using the social force
model for pedestrian behaviour. Moreover, we simulate several crowd manage-
ment strategies and investigate their ability to prevent the disaster. We evaluate
these strategies in virtual reality (VR) by measuring the response and arousal of
participants while experiencing the simulated event from a festival attendees
perspective. Overall, we find that opening an additional exit and removing
the police cordons could have significantly reduced the number of casualties.
We also find that this strategy affects the physiological responses of the partici-
pants in VR.
1. Introduction
Crowd disasters during large-scale events are a primary concern for event secur-
ity because of the related casualties and chaos. However, the investigation of
conditions that lead to crowd disasters in real environments is often infeasible
because of practical and ethical issues. By contrast, computer simulations can con-
tribute to our understanding of crowd disasters by providing a framework for the
formal analysis of an event. In addition, virtual reality (VR) experiments allow
researchers to investigate individualsresponses to simulated crowds and the
physical and social conditions surrounding the event (e.g. exiting barriers or
fences) and to precisely measure participantsphysiological reactions and spatial
behaviour. Both simulation and VR approaches may facilitate the development of
disaster prevention methods.
The goal of this paper is to employ a combination of simulation and VR
methods to gain a better understanding of crowd management and to help orga-
nizers avoid similar disasters in the future. Specifically, we investigate
interventions that might have been able to prevent the disaster at the 2010 Love
Parade music festival in Duisburg, Germany. First, we reproduce the events of
the 2010 Love Parade disaster with a simulation based on the social force
model (SFM) [1] calibrated with available data (https://loveparade2010doku.
wordpress.com/). We then test several possible crowd management strategies,
including the removal of physical obstacles and the separation of inflow and out-
flow. These strategies are evaluated with respect to crowd density, throughput,
© 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution
License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original
author and source are credited.
congestion and the number of simulated casualties. While the
results of this simulation appear to match observations from
the actual event, our approach focuses on the density of the
event within a simplified crowd behaviour framework. More
sophisticated models incorporating local behaviours are poss-
ible and can be the focus of future research. Second, we
conduct a VR experiment in order to investigate differences
in the simulated first-person experiences of the original disaster
and the best-performing crowd management strategy using a
head-mounted display. Here, individual participants are
immersed in one of two crowd scenarios as we measure their
physiological arousal and self-reported level of stress.
1.1. Computer simulations of crowd behaviour
Many researchers have attempted to address the conditions
that lead to crowd disasters using computer simulations
[25], real-world observations [68], real-world experiments
[9] and VR experiments [10]. Computer simulations of real
events have been used to study crowd disasters because of
their versatility and relatively low cost. Simulations provide
the ability to predict crowd behaviour in new and unseen
environmental conditions, allowing researchers to conduct
what ifexperiments [4]. Computer simulations typically
employ steering algorithms for individual agents and are eval-
uated with respect to the behaviour of a large number of agents
[5]. For example, the SFM describes the self-organization
of pedestrian movement using a microscopic model of ped-
estrians [1,3]. Inspired by the Newtonian law of motion, this
model has been successful in reproducing several common
crowd phenomena, such as lane formation and crowd turbu-
lence [11]. In order to evaluate the outcome of the
simulations, density [2], congestion [12] and crowding [13]
have been used as metrics to assess the level of risk.
1.2. Real-world observations and experiments
Based on observations of the Love Parade disaster, a number
of studies have begun to examine the management of the
event [7,14]. In order to prevent future disasters, Helbing &
Mukerji [7] have suggested the separation of inflow and out-
flow, the removal of obstacles (e.g. fences, police cordons) and
the provision of additional entrances/exits. Klüpfel [14] has
assessed the underlying causes and consequences of the
Love Parade disaster and has suggested that the proximate
causes of overcrowding include the late opening of the
entrance. Lian et al. [15] focus on extracting pedestrian move-
ment patterns from publicly available video footage of the
disaster. Krausz & Bauckhage [16] extend this work by
using the video footage to automatically detect the timing
of the congestion. Pretorius et al. [4] simulate several manage-
ment strategies in a model of the Love Parade disaster and
find that a one-directional flow might have prevented
injury compared with the original event.
In general, data corresponding to real-world events are
often difficult to obtain and may violate individualsprivacy.
In comparison, experiments in real environments can be
costly and difficult to organize, especially for large crowds,
and their scope is limited to situations that do not endanger
health or lives. Nonetheless, researchers have developed
crowd behaviour detection and flow computation [8] methods
that are capable of capturing aggregate behaviour without
tracking specific individuals. By applying a similar approach,
Moussaïd et al. [17] have analysed the organization of social
groups to predict walking patterns from video footage. Labora-
tory experiments have also been used to study local crowd
movement patterns at critical regions such as turning corners
[18,19] or stairs [20]. For example, Dias et al. [19] observed
crowd turning behaviour and found that higher turning
angles can reduce flow rates and velocities under normal con-
gestion. Similarly, Burghardt et al. [20] found that areas of high
density can precede a turning point at stairs, where congestion
forms. Such empirical evidence can be further used to calibrate
data-driven models for pedestrian simulations. Dias & Lovre-
glio [21] represented the floor as a continuous field in order
to better model pedestriansnavigation of a corner and vali-
dated these field representations with the observation of
walking behaviour during a laboratory experiment. In
addition, Crociani et al. [22] proposed an algorithm to repro-
duce smooth trajectories at corners and validated this
algorithm with data from laboratory experiments.
Real-world data from crowd disasters (e.g. during a Hajj
event in Mina [2]) can be used to calibrate and validate compu-
ter simulations in order to help predict future disasters.
Extracting continuous crowd movements from segmented
video clips has been a major challenge for acquisition of
crowd data from cameras. Khan et al. [23] successfully used
an unsupervised clustering algorithm to generate crowd
flows from segmented video clips and then compared these
with other tracking techniques from the literature. Automatic
tracking techniques for coarse-grained data analysis can also
benefit from reflective markers carried by crowd members
[9], and some researchers have employed a more traditional
approach by manually extracting data from videos in order
to improve head-counting methods [24].
1.3. Virtual reality experiments
Compared with real-world experiments, VR studies of crowd
behaviour allow for greater experimental control [25] and
opportunities for crowd visualization [26]. This type of visual-
ization provides a first-person perspective that can guide the
organizers towards better decisions and help patrons to experi-
ence the scenario within the crowd. When conducting single-
user studies, VR researchers must consider the manner in
which the crowd is visualized [26]. Depending on the appli-
cation, these visualized crowds may need to be
representative (e.g. with human-like bodies and movements
[27]) or realistic (e.g. with human-like faces and clothing
[28]). Despite the opportunities provided by VR for experimen-
tal control, crowd visualization and multimodal assessment,
there are a few notable limitations. These limitations include,
but are not limited to, constraints on task complexity (e.g. the
number of turns along a route) [10], motion sickness [29],
lack of interaction with real social agents [30] and the lack of
real-time proprioceptive feedback (e.g. collisions between ava-
tars) [25].
1.4. Crowds and physiological arousal
Crowds may influence individual self-reported affective [31],
behavioural [32] and physiological states [33]. In terms of
behavioural effects, virtual crowds can be used positively as
a social signal for finding an unobstructed exit [27] or negati-
vely as an obstacle blocking a potential path to safety.
Realistic crowd visualization in VR also provides opportunities
to investigate changes in psychological states resulting from
dangerous scenarios such as crowd disasters. For example,
royalsocietypublishing.org/journal/rsif J. R. Soc. Interface 17: 20200116
2
the presence of a standing avatar has been found to lead to a
higher physiological arousal than the presence of a running
avatar [34], and the appearance [35] and distance [36] of virtual
avatars have been found to positively correlate with physio-
logical arousal. Physiological responses such as electrodermal
activity (EDA) [37] and heart rate variability (HRV) [38] have
been used to study stress/distress [39], user experience [40],
attention and other aspects of cognition [41]. EDA and HRV
both reflect the activity of the autonomic nervous system.
While EDA reflects sympathetic arousal, (normal) HRV results
from a balance between sympathetic and parasympathetic
activity [38]. Questionnaires can also be used to distinguish
between affective states [42].
1.5. Overview of the Love Parade crowd disaster
The Love Parade was a popular German music festival that was
first organized in 1989. The Love Parade disaster occurred in
Duisburg on 24 July 2010 [7]. The festival area was approxi-
mately 100 000 m
2
and was constrained by railway tracks to
the east and by a freeway (major road) to the west (figure 1).
A tunnel from an old freight station funnelled visitors to a
main ramp that led to the festival area. The narrowest diameter
of the tunnel was 20 m. The festival area could only be entered
and exited through this tunnel, although a side ramp was avail-
able as an additional (reserve) exit. The main ramp was 26 m
wide at its narrowest point, but with a local effective width
of only 10.59 m owing to the presence of temporary fences.
The entire festival area was surrounded by fences.
A detailed analysis of the festival area revealed potential
safety issues in the event planning, including the limited
maximum capacity and the use of the tunnel as the only
entrance and exit [7]. The use of the main ramp was tempor-
arily restricted by police cordons that were added to control
traffic at 15.50 during the festival. Around this time (between
15.30 and 16.00), congestion around the main ramp began to
form. As congestion increased, people began climbing fences,
billboards and poles in order to escape from the dense crowd.
The fences were removed at approximately 16.20 in an
attempt to relieve the congestion [7]. This, however, could
not prevent the crowd disaster, which resulted in the death
of 21 people and injury to 500 festival attendees. The primary
cause of death was suffocation.
2. Simulation and interventions
2.1. Love Parade crowd disaster simulation
To simulate the 2010 Love Parade disaster, we use publicly
available online data from video surveillance cameras
(https://loveparade2010doku.wordpress.com/), a three-
dimensional model of the festival area and computer-controlled
agents to represent the moving crowd. We then extend this
simulation by modelling various crowd management scenarios
and compare these simulations in terms of congestion and
simulated casualties. We use the Unity 3D Game Engine
(http://www.unity.com) to construct a true-to-scale three-
dimensional virtual environment of the festival area and all
potential entrances and exits (figure 2). This environment was
created based on the description from Helbing & Mukerji [7]
and publicly available maps, plans and other documents. For
the simulation, we use a simple version of the environment
without lighting and texture details. The texture and lighting
were later added for the VR experiment and modelled based
on the surveillance videos (i.e. https://loveparade2010doku.
wordpress.com/)oftherealenvironment.
We compute a triangulated representation of the walkable
areas in the environment to form a navigation mesh. The
starting point and the destinations of the crowd flows were
determined from the surveillance videos. A search graph
was constructed on the navigation mesh (nodes are the cen-
troids of triangles, and edges connect adjacent triangles).
The A* search algorithm [43] was used to perform path-find-
ing operations on the search graph to find sequences of way
points between origins and destinations of agents. The SFM
[3] was used to steer the agent along this computed path
while avoiding collisions with the environment and other
pedestrians. Details of the exact targeting mechanism can
be found in the electronic supplementary material.
The analysis of the surveillance videos also provided esti-
mates of inflow, outflow, density and individual velocities at
different times. Videos from surveillance cameras were used
instead of videos from visitorsmobile phones because they
were more stable over time and often of higher quality. Given
the availability of the data, each surveillance video was separ-
ated into segments of 20 minutes except for the first three
segments (7, 6 and 17 min, respectively). Because of technical
problems with the recording, some frames were missing, and
the effective length of the video segments were shorter than
the duration they represented by up to 4 minutes.
In order to estimate visitorsvelocities in metres per
second (V(t)) as a function of density (ρ(t)), we used the fun-
damental diagram proposed by Weidmann [44],
V(t)¼1:34 (1 e1:913(1=
r
(t))1=5:4):(2:1)
Density can also be estimated by the combined inflow and
outflow per second and metre (Q(t)) divided by velocity (V(t)),
r
(t)¼Q(t)
V(t):(2:2)
The solutions of equation (2.1) can be determined numeri-
cally using Newtons method [45]. We used the velocity of
P1a P1b
P4
P3
fences
N
S
WE
P2
camera
casualty area
police cordon
Figure 1. Illustration of the festival area, including the locations of the sur-
veillance cameras used for the simulations, the police cordons (P1P4) and
the casualty area. Adapted from an image created by Helbing & Mukerji [7].
royalsocietypublishing.org/journal/rsif J. R. Soc. Interface 17: 20200116
3
the crowd after 15.20, which remains at 0.7 m s
1
constantly
during the simulation. We computed the flow values for
the centre of each time interval based on the total width of
the main ramp (irrespective of the fences).
Based on these estimates, we generated a crowd of agents
to populate the virtual environment and to simulate the
effects of different crowd management scenarios in terms of
density, congestion, throughput and casualties. Similar to
the real event, the agents entered the main ramp from
inside the tunnel and exited from the end of the main ramp
towards the festival area. The simulations represented esti-
mates of the real event from 15.20 to 16.40.
The SFM represents systematic forces (i.e. attraction and
repulsion) exerted by targets, obstacles and other pedestrians
that influence the agentsmovements [3]. Specifically, for
agent α, these forces include an acceleration force f0
a
(v
a
), a
repulsive force f
αi
(r
α
) caused by obstacles and boundaries,
and repulsive interactions f
αβ
(r
α
,v
α
,r
β
,v
β
) between the
agents [3]. The index of obstacles is i, and βrepresents the
other agents
f¼f0
a
(v
a
)þf
a
B(r
a
)þX
b
(=
a
)
f
ab
(r
a
,v
a
,r
b
,v
b
)
þX
i
f
a
i(r
a
,ri,t):(2:3)
The acceleration force f0
a
(v
a
) is defined by the direction of
the next destination e
α
, desired speed v0
a
and the current
speed v
α
, according to
f0
a
(v
a
)¼1
ta
(v0
a
e
a
v
a
):(2:4)
Other obstacles i(i.e. fences and walls) define the repul-
sive forces. In equation (2.5), r
a
r
a
iis the distance between
an agent and the obstacle, and V
i
represents a potential
repulsive force
f
a
i(r
a
)¼r
r
a
Vi(kr
a
r
a
ik):(2:5)
Repulsive forces between the agents are defined as follows:
f
ab
¼A
a
exp (r
ab
d
ab
)
B
a

n
ab
:(2:6)
A
a
represents interaction strength and B
a
is the range of
the repulsive interactions. d
αβ
is the distance between the
centres of the mass of agent αand βand r
αβ
is the sum
of their radii r
α
and r
β
.n
αβ
represents the normalized vector
pointing from agent βto α. We used the parameters
A
α
= 0.045, B
α
= 0.2, r
α
=r
β
=1 and v0
a
¼1:3m s
1.
2.2. Simulation of intervention scenarios
After the simulation of the original crowd disaster, we simu-
lated nine alternative crowd management scenarios based on
the recommendations of Helbing & Mukerji [7]. These nine
scenarios were based on five different variations of the orig-
inal simulation. First, we varied the presence of the police
cordons. The original police cordons were simulated using
obstacles that could not be walked through. The idea was
that removing the police cordons might reduce crowd
density by removing unnecessary obstacles. Second, we
varied whether the main fences were present at the beginning
of the simulation because the fences may have been unnecess-
ary obstacles increasing crowd density. Third, we varied
whether the side ramp was open. We expected that this
might reduce the density by decreasing the number of
people attempting to move through the same channel.
Fourth, we varied whether inflow and outflow were separ-
ated or not using both ramps or just the main ramp. The
separation of inflow and outflow can avoid confrontations
of opposite flow directions. Fifth, we varied whether there
was an additional exit near the festival area because this
might remove the limitations related to the restricted width
of the ramps. In total, we devised 10 scenarios (figure 3):
the original simulation (O), the original simulation without
police cordons (OP), the removal of fences while police cor-
dons were present (F+P), the removal of fences and of police
cordons (FP), the separation of inflow and outflow in the
presence of police cordons (S+P), the separation of inflow
(a)(b)
(c)(d)
Figure 2. Four different views of the virtual environment. (a) View of the tunnel. (b) View of the overall area with both the main and side ramps. (c) View of the tunnel
and a narrow staircase from the entrance to the main ramp. (d) View of the main ramp from a similar position to surveillance camera 13.
royalsocietypublishing.org/journal/rsif J. R. Soc. Interface 17: 20200116
4
and outflow without police cordons (SP), the inclusion of
the side ramp while police cordons were present (R+P), the
opening of the side ramp and no police cordons (RP), use
of an additional exit with police cordons (E+P) and an
additional exit and no police cordons (EP).
These scenarios were each repeated 10 times (conducted
simultaneously on various computing units with the same
simulation program) and compared in terms of maximum
occupation, simulated casualties, throughput, general crowd
density near the main ramp and congestion. Consistent with
the literature [7], we assumed the possibility of casualties
when crowd density exceeded four agents per square metre
(i.e. danger zones; D
4
). Since conditions with four agents per
square metre are not necessarily deadly, we additionally deter-
mined extreme danger zones (D
6
), where the crowd density
exceeded six agents per square metre, although other research-
ers have defined danger zones differently [46]. The
throughput was considered to be the number of agents that
successfully exited the festival area over the course of the
entire simulation. General crowd density was computed as
the number of agents per square metre, recorded once every
simulated minute and averaged over 80 minutes. Congestion
was defined as the number of agents who did not move at
least 1 m in 60 s. Cronbachsαwas computed for each of
these dependent measures in order to assess its consistency
across repetitions. We also conducted 2 (with or without
police cordons) by 5 (scenario) between-group ANOVAs in
order to test for systematic differences among the scenarios
in terms of each dependent measure.
2.3. Simulation results
Overall, there were very few unrealistically high densities in
these simulations. Numerically, all of the alternative crowd
management scenarios led to less density, less congestion,
more throughput and fewer expected casualties than the simu-
lations of the original event (table 1). There was also extremely
high consistency across repetitions for each dependent
measure (Cronbachsα> 0.989). The removal of police cordons
resulted in the largest differences compared with the original
scenario in terms of the reduction in congestion and danger
zones. The removal of fences and the use of the side ramp
also reduced the number of danger zones, but the most effec-
tive scenarios were those that separated inflow and outflow or
added another exit near the festival area. These observations
were confirmed by a series of ANOVAs. The ANOVA for all
dependent measures revealed a significant main effect of
police cordons, a significant main effect of scenario and a sig-
nificant interaction (table 2). The assumption of homogeneity
was violated for each ANOVA, so we confirmed each result
O F+P
E+P
O–P F–P
E–P
police cordon crowd movement removal of fences
R+P S+P R–P S–P
Figure 3. Illustration of the 10 simulated scenarios. Green arrows represent the in- and outflow of the crowd. Blue lines represent the police cordons, and the
orange crosses represent the removal of the main fences along the main ramp. Scenario O represents the original simulation; the F scenarios represent the removal
of fences; the S scenarios represent the separation of inflow and outflow; the R scenarios represent the inclusion of the side ramp; and the E scenarios represent
opening the additional exit. Each of these intervention scenarios has two conditions, either with (+) or without () police cordons (P).
royalsocietypublishing.org/journal/rsif J. R. Soc. Interface 17: 20200116
5
using a non-parametric aligned ranks transformation ANOVA
[47]. In addition, figure 4 represents the first repetition of the O
scenario in terms of the number of danger zones over time,
and figure 5 illustrates the crowd densities for all 10 scenarios
around the accident area at 16.00.
Our simulations reproduced the crowding effect in the
original scenario at approximately the same time and location
it was observed in the video footage. Our complementary
results suggest that other crowd management strategies
may have led to fewer or no casualties by decreasing density
and congestion and increasing throughput. These results sup-
port the strategies suggested by Helbing & Mukerji [7].
Specifically, we found that the removal of physical obstacles
(i.e. fences and police cordons) and the separation of inflow
and outflow substantially reduced the expected number of
simulated casualties.
3. Virtual reality experiment
While the simulation results provide evidence for the efficacy
of these strategies for collective behaviour, they do not by
themselves reveal the mechanisms underlying individual
reactions to the crowd. In order to compare the best (EP)
and worst (O) scenarios in terms of individualsphysiological
and behavioural responses, we devised a VR experiment in
which participants experienced the simulation from a first-
person perspective (figure 6). We expected participants in
Table 1. Means and standard deviations (in parentheses) of simulation results for all 10 replications. Scenario O represents the original simulation; the F
scenarios represent the removal of fences; the S scenarios represent the separation of inow and outow; the R scenarios represent the inclusion of the side
ramp; and the E scenarios represent opening the additional exit. Each of the intervention scenarios has two conditions, with (+) or without () police
cordons (P). The measures reported here include maximum occupation (max), simulated casualties (D
4
and D
6
), throughput (TP), general crowd density (density)
near the main ramp and congestion. Cronbachsαrepresents the consistency of these measures across the 10 replications.
scenario max D
4
D
6
TP density congestion
O11.1 (1.8) 357.8 (39.2) 28.7 (6.8) 13.1 (9.7) 0.889 (0.004) 35220.5 (19.9)
OP6.6 (0.5) 192.3 (32.5) 1.4 (1.5) 9.1 (5.8) 0.899 (0.012) 35231.1 (18.4)
F+P 8.2 (0.8) 392.1 (33.8) 19.5 (5.0) 17.5 (6.1) 0.826 (0.007) 35259.9 (28.2)
FP7.6 (0.5) 338 (46) 13.8 (5.4) 20.9 (10.0) 0.832 (0.005) 3526 (37)
S+P 3 (0) 0 (0) 0 (0) 10581.1 (73.5) 0.203 (0.003) 4255.9 (14.3)
SP2.4 (0.5) 0 (0) 0 (0) 13568.8 (146.3) 0.195 (0.002) 3313.8 (43.4)
R+P 9.8 (0.9) 189.3 (31.8) 15.8 (4.6) 29.4 (8.0) 0.836 (0.013) 35200.9 (32.1)
RP6.2 (0.4) 120.2 (16.8) 0.4 (1) 31.1 (7.1) 0.856 (0.019) 35203.2 (30.8)
E+P 3 (0) 0 (0) 0 (0) 17174.2 (89.6) 0.205 (0.002) 3997.2 (13.4)
EP2.3 (0.5) 0 (0) 0 (0) 18926.3 (84.1) 0.195 (0.001) 3261.7 (14.5)
Cronbachsα0.994 0.997 0.989 0.999 0.999 0.999
Table 2. ResultsoftheANOVAsbasedonsimulationdataforeachdependent
variable (DV). Across all DVs, there are reliable effects for the presence of police
cordons and other variations across scenarios. MSE represents mean squared error.
DV effect FMSE p
max police cordons 163.636 0.611 <0.001
scenario 309.625 0.611 <0.001
interaction 29.495 0.611 <0.001
D
4
police cordons 116.584 714.914 <0.001
scenario 744.874 714.914 <0.001
interaction 32.211 714.914 <0.001
D
6
police cordons 189.153 12.384 <0.001
scenario 101.953 12.384 <0.001
interaction 55.153 12.384 <0.001
TP police cordons 5313.469 4230.030 <0.001
scenario 342056.024 4230.030 <0.001
interaction 2216.595 4230.030 <0.001
density police cordons 4.183 0.00008 0.044
scenario 32322.325 0.00008 <0.001
interaction 9.603 0.00008 <0.001
congestion police cordons 3759.747 735.106 <0.001
scenario 8110471.825 735.106 <0.001
interaction 1489.341 735.106 <0.001
0
10
20
time (hour.minutes)
no. danger zones
danger zone D4
danger zone D6
30
15.20 15.40 16.4016.20
16.00
Figure 4. Change in the number of danger zones (D
4
and D
6
) for the first
repetition of the O simulation from 15.20 to 16.40. The number of danger
zones increases over time until 16.15 and then suddenly decreases because
of the removal of the main fence. This graph peaked at 27 danger zones
around 16.15.
royalsocietypublishing.org/journal/rsif J. R. Soc. Interface 17: 20200116
6
the O condition to be more physiologically aroused than
participants in the EP condition with respect to EDA and
HRV. We also expected participants in the O condition to
report higher levels of distress and produce more head
movements than participants in the EP condition.
Participants were recruited via the University Registration
Center for Participants (www.uast.uzh.ch). A total of 58
participants (27 females; mean age = 27, range = 1939) partici-
pated in the study. Two participants were excluded from the
study because of equipment failure. All participants received
CHF15 for approximately 45 min of participation. Before
each experimental session, participants were given an infor-
mation sheet and asked to complete an informed consent
form. After this consent procedure, the experimenter helped
the participant attach the three electrocardiogram electrodes.
Next, participants completed the demographics questionnaire,
video game experience questionnaire and the pre-test ques-
tions from the short stress state questionnaire (SSSQ). The
experimenter then attached two electrodes to the participants
fingers to collect EDA data and placed the head-mounted
display (HMD) on the participants head. After adjusting the
HMD, participants completed a training procedure in which
they were asked to look left and right and then look towards
a button that was shown at a specific location on the display.
After training, a 7 minute nature video was presented to par-
ticipants in order to record a baseline measure of their
physiological activity. During the testing phase, the partici-
pants viewed four identical replays of one simulated scenario
from a first-person perspective. Each replay was 2 minutes
long, and participants had small breaks between replays. The
video sequences did not contain any distressful content.
Participants were randomly assigned to one of two groups
(O or EP), each of which represented a scenario from the
simulations above. Specifically, we compared replays of the O
scenario with replays of the EP scenario. These simulated
scenarios were thus the only between-subject independent
variable. Our dependent variables included responses to the
video game experience questionnaire, the three subscales of
the SSSQ, EDA, HRV, and the magnitude of head movements
derived from the gyroscope in the HMD. EDA, HRV and head
movement data were aggregated across trials, and all of these
measures were compared between the O and EP scenarios
using two-tailed, independent-samples t-tests. For more details
regarding the experiment methods, please see the electronic
supplementary material.
Results of the VR experiment showed that, for the SSSQ,
there were no significant differences between the O and
EP scenarios in terms of distress (O = 0.66 ± 0.71, EP=
0.54 ± 1.31), t(56) = 0.444, s.e. = 0.277, p= 0.659; engagement
(O = 0.30 ± 0.82, EP=0.08 ± 0.97), t(56) = 0.93, s.e. =
0.236, p= 0.357; or worry (O = 0.26 ± 0.76, EP=0.15 ±
0.96), t(56) = 0.489, s.e. = 0.227, p= 0.627. For EDA, we
found a significant difference between the O and EP scenarios
in terms of non-specific skin conductance responses (nSCRs)
(O = 3.57 ± 13.91, EP = 11.24 ± 12.09), t(56) = 2.241, s.e. =
3.422, p= 0.029, but not in terms of the sum of the amplitude
of these peaks (AmpSum) (O = 2.93 ± 10.57, EP = 4.24 ±
5.76), t(56) = 0.586, s.e. = 2.237, p= 0.560 (figure 7). This
suggests that the frequency (but not the magnitude) of skin
O–PO
F+P F–P
10
0
5
S+P S–P
R+P R–P
E+P E–P
Figure 5. Density maps of the general ramp area at 16.00 for the first rep-
etition of each scenario. The blue lines represent walls and fences. The red
dots represent the numbers of pedestrians in each grid cell. The colour bar on
the right reflects the number of people in each cell.
(a)
(b)
Figure 6. Screenshots from (a) the O (original) scenario and (b)EP scenario
(additional exit without police cordons) in the virtual reality environment.
royalsocietypublishing.org/journal/rsif J. R. Soc. Interface 17: 20200116
7
conductance responses was higher in the EP scenario than in
the O scenario. For the HRV data, there was no significant
difference between the O and EP scenarios in terms of
Log(HF) (O = 0.27 ± 0.79, EP=0.12 ± 0.50), t(53) = 0.894,
s.e. = 0.176, p= 0.375. In addition, there was not a significant
difference between the O and EP scenarios in terms of head
movements (O = 1150.22 ± 635.17, EP = 1267.46 ± 527.22),
t(56) = 0.765, s.e. = 5.837, p= 0.448. During the course of the
whole experiment, no participants reported motion sickness
or interrupted the procedure because of discomfort.
The results reveal that, when viewing from a first-person
perspective, participants had higher nSCRs in the effective
intervention group (EP) than in viewing the original (O)
simulation. The simulation of the original crowd disaster
may have been less stressful than in reality because, owing
to ethical and technical reasons, we could not simulate and
present the participants with animations of crowd members
falling and stepping on each other. In addition, our design
cannot be used to disentangle the effects of crowd movement
and visual exposure to the virtual environment. This limit-
ation makes it difficult to interpret the observed effect of
management strategies on individual physiological arousal.
In order to explain this effect, future research could systema-
tically vary crowd movement and exposure to the
environment. However, this approach would require the
experimenter to control the viewing direction of the partici-
pants during the replays. Future research can also address
the possibility of an interaction between motion sickness
[48] and stress resulting from differences between the
crowd scenarios, although we did not observe any motion
sickness in the present study.
4. Discussion
In this paper, we present the results of computer simulations
and a VR experiment that investigated the effects of possible
interventions during the 2010 Love Parade disaster on
simulated casualties and physiological arousal in VR. We
simulated the original event along with several other scenarios,
including the removal of the main fences and/or police cor-
dons, the opening of a side ramp for entering or exiting, the
separation of inflow and outflow using the main and side
ramps, and the opening of an additional exit near the festival
area. These simulations revealed that, compared with the orig-
inal scenario, all of these interventions led to less congestion,
more throughput, and fewer or no simulated casualties. Our
simulations provide a mechanism to assess previous disasters
and may support event managers in devising strategies to
avoid future crowd disasters. Specifically, we demonstrate
that crowd simulations based on the SFM and rendered in
Unity can be used to determine possible causes of disasters.
The application of the SFM was effective for recreating the
physical properties and dynamics of the observed crowd.
Our introduction of the danger zone metrics allowed us to
easily assess the risk level of the simulated event. Our simu-
lation of various management strategies demonstrated that
alternative organizational decisions regarding crowd control
during the event could have helped to prevent the disaster.
In addition, rendering these simulations in Unity may help
event managers visualize the effects of specific interventions
on the crowd. Our VR experiment revealed that the best
(EP) and worst (O) scenarios significantly differed in terms
of the frequency of skin conductance responses.
With respect to the simulations, we found that the most
effective strategy for reducing simulated casualties was a com-
bination of removing police cordons and opening an additional
exit from the festival area. Most previous research specifically
focused on the detection of crowd movement patterns during
the Love Parade disaster [8,15,16]. We extended these
approaches by simulating several interventions suggested by
previous assessments of the organization and operation of
this event [7,14]. Our results support the potential of these
interventions for saving lives. However, the scope of our find-
ing is limited in terms of using density as a proxy to estimate
critical crowded conditions. The simulation of more realistic
crowd behaviours such as falling and stepping upon others
might help one to establish a greater degree of accuracy in
this respect in the future, but there are ethical issues to be
considered.
Consistent with Pretorius et al. [4], we found that the
separation of inflow and outflow (resulting in one primary
direction of motion) increases throughput and reduces con-
gestion in the simulated crowd. However, unlike in our
study, Pretorius et al. [4] did not find any benefit in removing
the police cordons. Owing to several differences between
the implementations and analyses of the two studies, we
cannot identify the exact reason for this difference in results.
In our study, the effect of the removal of police cordons was
extremely consistent across measures and most scenarios.
Indeed, the removal of police cordons in the present simu-
lations decreased the maximum occupation, decreased the
number of danger zones, increased throughput, decreased
density and decreased congestion. Such an effect is also con-
sistent with recent official expertise [49,50]. Importantly, our
study extended that by Pretorius et al. [4] by including a
measure of simulated casualties based on danger zones. As
decreased throughput and increased congestion do not
necessarily always result in more danger zones, event plan-
ners may adopt metrics such as danger zones in order to
gain a better understanding of the potential benefits of par-
ticular interventions in the future.
With respect to the VR experiment, we found that view-
ing a simulation of an effective intervention from a first-
person perspective led to higher nSCRs than viewing the
original simulation. One possible explanation of our results
is that participants in the EP group moved further along
the ramp and were exposed to more variation in the virtual
environment than participants in the O group. This
additional visual exposure may have increased arousal by
inducing engagement or curiosity. The skin conductance
0
5
10
nSCR
0
2
4
AmpSum (mus)
O E–PO E–P
(a)(b)
Figure 7. Difference between O and EP scenarios in terms of (a) nSCR and
(b) AmpSum. The error bars represent the standard error of the difference
between the two groups. Although both trends are in the same direction,
we only found a significant difference in terms of nSCR (p= 0.029).
royalsocietypublishing.org/journal/rsif J. R. Soc. Interface 17: 20200116
8
results by themselves cannot disentangle these possibilities.
Previous research in VR has found that the idle behaviour
of a single avatar [34], the presence of a group of avatars
[35], and smaller distances between avatars and the observer
[36] increased physiological arousal. Furthermore, Llobera
et al. [36] also found that more moving avatars led to
higher physiological arousal, but the number of avatars in
their study (maximum four agents) was much lower than
in the present study (up to 2000 agents in the original
group). Our two VR scenarios were similar with regards to
avatar presence and distances, and the main difference
between scenarios is the level of congestion. However, con-
gestion was negatively related to avatar motion. Thus, our
finding that idle avatars lead to lower physiological arousal
may be inconsistent with Fox et al. [34]. The VR experiment
was also limited in terms of the presence of the simulated
crowd in that it did not reproduce crowd turbulence from
the original disaster. Hence, the absence of crowd turbulence
is likely to have caused participants to experience less stress
from the first-person perspective than they would have
experienced during the disaster. Future research can focus
on explaining the exact mechanisms underlying the signifi-
cant difference between scenarios by systematically varying
crowd parameters (e.g. density, congestion) and environment
parameters (e.g. spaciousness, visual noise).
Since the study was limited to a simplified crowd behav-
iour model, it was not possible to completely reproduce the
complex phenomenon and spontaneous crowd movement
patterns of the Love Parade crowd disaster. One limitation
of our simulations is that rendering simulations in Unity is
time intensive and prohibits a large number of replications.
As a result, parameters in the SFM were difficult to optimize
with respect to the original video data. Nonetheless, conduct-
ing 10 repetitions of each scenario allowed us to assess the
consistency of each measure and the statistical significance
of differences across scenarios. In future studies, one way to
overcome this limitation would be to conduct the crowd
simulation on a lighter platform and then only render
the simulated crowd with optimized trajectories in Unity.
Another possible limitation is the manual extraction of
crowd data from video footage. In the future, computer
vision technology may be used to extract more precise esti-
mates of in- and outflows. Lastly, the simulated crowd is
notably less intelligent than members of the real crowd and
so does not necessarily reflect the spontaneous behaviours
of people reacting to a dangerous situation. Many aspects
of the crowd behaviour (e.g. realistic turning behaviour at
the corners, crowd turbulence) could have augmented this
simplified model and improved the veracity of the simu-
lations. Future research can propose and attempt to validate
these more sophisticated models of pedestrian dynamics for
large crowds and environments [21,22,51].
Despite these limitations, we demonstrate a novel meth-
odology for the research of crowd disasters and their
prevention. To our knowledge, this is the first study to com-
bine simulations and experimentation in VR of a crowd
disaster of this complexity and size. We expect that follow-
up work will further increase the sophistication and precision
of this approach and thereby underline its huge potential and
value. In conclusion, the coordination of crowd simulations
and VR technologies can help event managers to assess
potential dangers more realistically and to make more effec-
tive decisions about crowd management strategy in advance.
Ethics. All of these methods were approved by the Ethics Commission
of ETH Zürich (EK2015-N-37).
Data accessibility. The experimental data are available in the electronic
supplementary material.
Authorscontributions. H.Z., T.T., M.K., K.W., C.H., D.H. and V.R.S.
designed the research. H.Z., K.W., M.K. and T.T. performed the simu-
lation. H.Z. performed the VR experiments. H.Z., T.T. and V.R.S.
analysed the data and wrote the paper. All authors reviewed the
manuscript.
Competing interests. We declare that we have no competing interests.
Funding. The study was funded by Prof. Markus Gross, Computer
Graphics Laboratory, ETH Zürich. M.K. was funded in part by NSF
IIS-1703883 and NSF S&AS-1723869.
Acknowledgements. We thank Yesol Park for helpful assistance during
the experiment and Mehdi Moussaid for insightful discussions.
Reference
1. Helbing D, Molnar P. 1995 Social force model for
pedestrian dynamics. Phys. Rev. E 51, 4282. (doi:10.
1103/PhysRevE.51.4282)
2. HelbingD,JohanssonA,Al-AbideenHZ.2007
Dynamics of crowd disasters: an empirical study. Phys.
Rev. E 75, 046109. (doi:10.1103/PhysRevE.75.046109)
3. Helbing D, Buzna L, Johansson A, Werner T. 2005
Self-organized pedestrian crowd dynamics:
experiments, simulations, and design solutions.
Transp. Sci. 39,124. (doi:10.1287/trsc.1040.0108)
4. Pretorius M, Gwynne S, Galea ER. 2015 Large crowd
modelling: an analysis of the Duisburg Love Parade
disaster. Fire Mater. 39, 301322. (doi:10.1002/
fam.2214)
5. Kapadia M, Singh S, Hewlett W, Faloutsos P. 2009
Egocentric affordance fields in pedestrian steering.
In Proc. of the 2009 Symp. on Interactive 3D
graphics and games, Boston, MA, February 2009,
pp. 215223. New York, NY: ACM.
6. Johansson A, Helbing D, Shukla PK. 2007
Specification of the social force pedestrian model by
evolutionary adjustment to video tracking data. Adv.
Complex Syst. 10, 271288. (doi:10.1142/
S0219525907001355)
7. Helbing D, Mukerji P. 2012 Crowd disasters
as systemic failures: analysis of the Love Parade
disaster. EPJ Data Sci. 1, 1. (doi:10.1140/epjds7)
8. Krausz B, Bauckhage C. 2012 Loveparade 2010:
automatic video analysis of a crowd disaster.
Comput. Vis. Image. Underst. 116, 307319.
(doi:10.1016/j.cviu.2011.08.006)
9. Motsch S, Moussaid M, Guillot EG, Moreau M, Pettre
J, Theraulaz G, Appert-Rolland C, Degond P. 2017
Forecasting crowd dynamics through coarse-grained
data analysis. (https://www.biorxiv.org/content/10.
1101/175760v2)
10. Suma E, Finkelstein S, Reid M, Babu S, Ulinski A,
Hodges LF. 2010 Evaluation of the cognitive effects
of travel technique in complex real and virtual
environments. IEEE Trans. Vis. Comput. Graph 16,
690702. (doi:10.1109/TVCG.2009.93)
11. Helbing D, Farkas I, Vicsek T. 2000 Simulating
dynamical features of escape panic. Nature 407,
487490. (doi:10.1038/35035023)
12. Feliciani C, Nishinari K. 2018 Measurement of
congestion and intrinsic risk in pedestrian crowds.
Transp. Res. Part C: Emerg. Technol. 91, 124155.
(doi:10.1016/j.trc.2018.03.027)
13. Bandini S, Crociani L, Gorrini A, Nishinari K,
Vizzari G. 2020 Unveiling the hidden dimension
of pedestrian crowds: introducing personal space
and crowding into simulations. Fundam.
Informaticae 171,1938. (doi:10.3233/FI-2020-
1870)
14. Klüpfel H. 2014 The love parade disaster. In
Pedestrian and evacuation dynamics 2012,
pp. 13851394. New York, NY: Springer.
royalsocietypublishing.org/journal/rsif J. R. Soc. Interface 17: 20200116
9
15. Lian L, Song W, Richard YKK, Ma J, Telesca L. 2017
Long-range dependence and time-clustering
behavior in pedestrian movement patterns in
stampedes: the Love Parade case-study. Physica A
469, 265274. (doi:10.1016/j.physa.2016.11.048)
16. Krausz B, Bauckhage C. 2011 Analyzing pedestrian
behavior in crowds for automatic detection of
congestions. In Proc. 2011 IEEE Int. Conf. on Computer
Vision Workshops (ICCV Workshops), Barcelona, Spain,
November 2011, pp. 144149. Piscataway, NJ: IEEE.
17. Moussaïd M, Perozo N, Garnier S, Helbing D,
Theraulaz G. 2010 The walking behaviour of
pedestrian social groups and its impact on crowd
dynamics. PLoS ONE 5, e10047. (doi:10.1371/
journal.pone.0010047)
18. Zhang J, Klingsch W, Schadschneider A, Seyfried A.
2011 Transitions in pedestrian fundamental
diagrams of straight corridors and T-junctions.
J. Stat. Mech: Theory Exp. 2011, P06004.
19. Dias C, Sarvi M, Shiwakoti N, Ejtemai O, Burd M.
2014 Examining the impact of different turning
angles on the collective egress of crowds. J. Transp.
Saf. Secur. 6, 167181. (doi:10.1080/19439962.
2013.831964)
20. Burghardt S, Seyfried A, Klingsch W. 2013
Performance of stairsfundamental diagram and
topographical measurements. Transp. Res. Part C:
Emerg. Technol. 37, 268278. (doi:10.1016/j.trc.
2013.05.002)
21. Dias C, Lovreglio R. 2018 Calibrating cellular
automaton models for pedestrians walking through
corners. Phys. Lett. A 382, 12551261. (doi:10.
1016/j.physleta.2018.03.022)
22. Crociani L, Shimura K, Vizzari G, Bandini S. 2018
Simulating pedestrian dynamics in corners and
bends: a floor field approach. In Proc. Int. Conf. on
Cellular Automata, Como, Italy, September 2018,
pp. 460469. Cham, Switzerland: Springer
International Publishing.
23. Khan SD, Bandini S, Basalamah S, Vizzari G. 2016
Analyzing crowd behavior in naturalistic conditions:
identifying sources and sinks and characterizing
main flows. Neurocomputing 177, 543563. (doi:10.
1016/j.neucom.2015.11.049)
24. Al-Ahmadi HM, Alhalabi WS, Malkawi RH, Reza I.
2018 Statistical analysis of the crowd dynamics in
Al-Masjid Al-Nabawi in the city of Medina, Saudi
Arabia. Int. J. Crowd Sci. 2,6473.
25. Moussaïd M, Schinazi VR, Kapadia M, Thrash T.
2018 Virtual sensing and virtual reality: how
new technologies can boost research on crowd
dynamics. Front. Robotics AI 5, 82. (doi:10.3389/
frobt.2018.00082)
26. Kallmann M, Kapadia M. 2014 Navigation meshes
and real-time dynamic planning for virtual worlds.
In Proc. ACM SIGGRAPH 2014 Courses, Vancouver,
Canada, August 2014, p. 3. New York, NY: ACM.
27. Moussaïd M, Kapadia M, Thrash T, Sumner RW,
Gross M, Helbing D, Hölscher C. 2016 Crowd
behaviour during high-stress evacuations
in an immersive virtual environment.
J. R. Soc. Interface 13, 20160414. (doi:10.1098/
rsif.2016.0414)
28. Sharma S, Otunba S, Han J. 2011 Crowd simulation
in emergency aircraft evacuation using virtual
reality. In Proc. 2011 16th Int. Conf. on Computer
Games (CGAMES), Louisville, KY, 2730 July 2011,
pp. 1217. Piscataway, NJ: IEEE.
29. Kim HK, Park J, Choi Y, Choe M. 2018 Virtual reality
sickness questionnaire (VRSQ): motion sickness
measurement index in a virtual reality environment.
Appl. Ergon. 69,6673. (doi:10.1016/j.apergo.2017.
12.016)
30. Shiban Y, Diemer J, Brandl S, Zack R, Mühlberger A,
Wüst S. 2016 Trier social stress test in vivo and in
virtual reality: dissociation of response domains.
Int. J. Psychophysiol. 110,4755. (doi:10.1016/j.
ijpsycho.2016.10.008)
31. Epstein YM, Woolfolk RL, Lehrer PM. 1981
Physiological, cognitive, and nonverbal responses to
repeated exposure to crowding. J. Appl. Soc. Psychol.
11,113. (doi:10.1111/j.1559-1816.1981.tb00818.x)
32. Langer EJ, Saegert S. 1977 Crowding and cognitive
control. J. Pers. Soc. Psychol. 35, 175. (doi:10.1037/
0022-3514.35.3.175)
33. Evans GW. 1979 Behavioral and physiological
consequences of crowding in humans. J. Appl. Soc.
Psychol. 9,2746. (doi:10.1111/j.1559-1816.1979.
tb00793.x)
34. Fox J, Bailenson JN, Ricciardi T. 2012 Physiological
responses to virtual selves and virtual others.
J. CyberTherapy Rehabil. 5,6973.
35. Christou C, Herakleous K, Tzanavari A, Poullis C.
2015 Psychophysiological responses to virtual
crowds: implications for wearable computing. In
Proc. 2015 Int. Conf. on Affective Computing and
Intelligent Interaction (ACII), Xian, China, September
2015, pp. 3541. Piscataway, NJ: IEEE.
36. Llobera J, Spanlang B, Ruffini G, Slater M. 2010
Proxemics with multiple dynamic characters in an
immersive virtual environment. ACM Trans. Appl.
Percept. (TAP) 8,3.
37. Boucsein W. 2012 Electrodermal activity. New York,
NY: Springer Science & Business Media.
38. Acharya UR, Joseph KP, Kannathal N, Lim CM, Suri
JS. 2006 Heart rate variability: a review. Med. Biol.
Eng. Comput. 44, 10311051. (doi:10.1007/s11517-
006-0119-0)
39. Wiederhold BK, Jang DP, Kim SI, Wiederhold MD.
2002 Physiological monitoring as an objective tool
in virtual reality therapy. CyberPsychol. Behav. 5,
7782. (doi:10.1089/109493102753685908)
40. Egan D, Brennan S, Barrett J, Qiao Y, Timmerer C,
Murray N. 2016 An evaluation of heart rate and
electrodermal activity as an objective QoE
evaluation method for immersive virtual reality
environments. In Proc. 2016 Eighth Int. Conf. on
Quality of Multimedia Experience (QoMEX), Lisbon,
Portugal, June 2016, pp. 16. Piscataway, NJ: IEEE.
41. Figner B, Murphy RO. 2011 Using skin conductance
in judgment and decision making research. A
handbook of process tracing methods for decision
research (eds M Schulte-Mecklenbeck, A
Kuehberger, JG Johnson), pp. 163184. New York,
NY: Psychology Press.
42. Helton WS. 2004 Validation of a short stress state
questionnaire. In Proc. of the Human Factors and
Ergonomics Society Annual Meeting, New Orleans,
LA, September 2004, vol. 48, pp. 12381242. Santa
Monica, CA: Human Factors and Ergonomics Society.
43. Hart PE, Nilsson NJ, Raphael B. 1968 A formal basis
for the heuristic determination of minimum cost
paths. IEEE Trans. Syst. Sci. Cybern. 4, 100107.
(doi:10.1109/TSSC.1968.300136)
44. Weidmann U. 1992 Transporttechnik der Fussgänger.
Zurich, Switzerland: Institut für Verkehrsplanung,
Transporttechnik, Strassen- und Eisenbahnbau (IVT),
ETH Zürich.
45. Kelley CT. 2003 Solving nonlinear equations with
Newtons method, vol. 1. Berlin, Germany: SIAM.
46. Fruin J. 1981 Crowd disastersa systems evaluation
of causes and countermeasures. National Bureau of
Standards Information Report, pp. 813261.
Gaithersburg, MD: National Bureau of Standards.
47. Wobbrock JO, Findlater L, Gergle D, Higgins JJ. 2011
The aligned rank transform for nonparametric
factorial analyses using only anova procedures. In
Proc. of the SIGCHI Conf. on Human Factors in
Computing Systems, Vancouver, Canada, May 2011,
pp. 143146. New York, NY: ACM.
48. Kim HG, Cheon EJ, Bai DS, Lee YH, Koo BH. 2018
Stress and heart rate variability: a meta-analysis
and review of the literature. Psychiatry Invest. 15,
235. (doi:10.30773/pi.2017.08.17)
49. Hans B. 2010 Gutachten belegt Verantwortlichkeit
der Polizei. See https://www.spiegel.de/panorama/
love-parade-gutachten-belegt-verantwortlichkeit-
der-polizei-a-712771.html.
50. Kurz P. 2018 Die Rolle der Polizei bei der
Loveparade. See https://www.wz.de/panorama/die-
rolle-der-polizei-bei-der-loveparade-aid-32910255.
51. Dias C, Abdullah M, Sarvi M, Lovreglio R,
Alhajyaseen W. 2019 Modeling and simulation of
pedestrian movement planning around corners.
Sustainability 11, 5501. (doi:10.3390/su11195501)
royalsocietypublishing.org/journal/rsif J. R. Soc. Interface 17: 20200116
10
... In the [20], they proposed using the simulation for traffic movement and crowd flow by utilizing platforms such as Deepint.net to estimate and simulate pedestrian movement and congestion times. While in the [21] utilized virtual reality to simulate stampede incidents and applied them to various events. The experiments highlighted the necessity of emergency exits as reporting, searching for missing things, searching for points of interest, etc. ...
... Helpful tools, but insufficient without their use within a unified framework and activated algorithms [20,21] Using simulation, virtual and augmented reality to train crowds Inadequate, did not care about the health aspect users in each sector to prevent congestion that may lead to catastrophic stampedes or violations of distancing measures. The existing numbers in each sector will be precisely regulated through wristbands and gates without the need for machine learning algorithms or deep learning to calculate the numbers, which require high computational power. ...
Article
Full-text available
Many countries witness the organization of various events, whether sports, religious, artistic, cultural, or even political. These events gather large crowds in specific locations. A minor incident within these crowds can escalate into a disaster if not dealt with immediately and effectively. Therefore, crowd management and ensuring its safety present significant challenges to date. Preserving the safety and health of participants within crowds is one of the most sensitive issues, which becomes even more complex during pandemics. Unfortunately, there are still no effective solutions to guarantee the safety and health of crowds. The world has witnessed disasters in recent years that have led to the deaths of thousands in crowds. This research presents a classification of all challenges facing crowds. Then, it proposes a comprehensive framework to address them and enhance the health and safety of participants. The framework relies on integrating various solutions and technologies in the form of stages, tasks, and services. The research also proposes a crowd-monitoring algorithm and smart scheduling for the sanitization process, in addition to early warning and emergency procedures in case of any threat. The algorithm is based on image processing and text analysis. The algorithm has been tested on a dataset for a set of crowd images and others for tweets about various incidents, achieving an accuracy exceeding 98%. The framework will enhance the level of safety and health in crowds and pave the way for many future supporting services which can be provided via a general application for smartphones.
... It guarantees security by easily regulating the variables that might lead to the emergence of catastrophes as a result. It has been shown that allowing exits at public conveniences is more effective than enforcing police cordons blocking entrance to (or escape from) the event areas [3]. It is well established that humans copy one another. ...
Article
Full-text available
Under the recent COVID-19 pandemic circumstances, crowd management has received a lot of attention, with the goal of ensuring that ill people do not become a source of viral transmission. The World Health Organization (WHO) recommends maintaining a safe distance and wearing a mask in crowds as part of standard operating procedures (SOP), which are currently the most effective. Methods of COVID-19 prevention. Face detection and social distance detection models have been built using a variety of methodologies and tactics. In this research a deep learning model is provided to identify those people who do not have masks and those people who do not maintain social distance to keep themselves safe and keep the virus contained. Individuals who break the SOP are also counted. To enhance the model’s detection capabilities, the proposed framework employs the Single Shot MultiBox Detector (SSD) for feature extraction. To further integrate the extracted features, Spatial Pyramid Pooling (SPP) is utilized, enabling the model to handle features of varying spatial dimensions effectively. As a framework for the classifier, the MobilenetV2 architecture develops the model very light, fast, and computationally efficient, enabling it to be used in real-time face mask detection and safe social distance detection after embedding in the devices, which is the main goal of this study. The method used in this study achieves a 97% accuracy rate and decreases the loss to 0.10%.
... The crowd management and seat allocation system for closed venues can further be used for different applications such as conference hall seating arrangements, aeroplane seating arrangements, stadiums, bus seating arrangements, etc. The maximum number occupying any event can be hard and thus cause an exceeding inflow and outflow of people in the venue (Ferleger 2022;Vouriot, Burridge, Noakes, and Linden 2021;Zhao et al. 2020). Furthermore, it gets very hard for security guards to keep track of the number of people entering and leaving any venue to ensure that the recommended occupancies are not exceeded. ...
... Crowd dynamics can also impact individuals' emotional, behavioral, and physiological states. In 2020, a study conducted by [20] investigated the potential effects of various interventions in the tragic 2010 Love Parade incident, which resulted in 21 fatalities and numerous injuries. Using computer simulations and virtual reality (VR) experiments, the study examined simulated casual-ties and physiological arousal. ...
Article
Full-text available
This study aims to create and examine a multimodal dataset to enhance crowd management during the Hajj seasons. Sixty-four participants were engaged in Hajj rituals such as Tawaf, Saai, prayer, and Doaa providing location and peripheral physiological data, collected and annotated using a custom-made smartphone application. The collected data was leveraged to conduct a comprehensive analysis, specifically focusing on the classification of the type of Hajj activity, level of fatigue, and emotional states based on peripheral physiological signals. Three deep learning classification models were developed and validated using feedforward neural networks. The models achieved satisfactory accuracy scores in classifying the type of Hajj activity (41.71%), level of fatigue (85.27%), and emotional states (82.47%). While presenting a straightforward use case, this research chiefly provides decision makers and the scientific community with a statistically significant open data set aside with a deep learning architecture capable of characterizing crowd behavior for the purpose of automating crowd management and monitoring.
... Arias et al. [3] reconstructed the Love Parade disaster in VR to evaluate various disaster prevention strategies. Zhao et al. [45] used VR to test different fire evacuation systems for a building under construction, highlighting the cost-effectiveness of VR in such studies. ...
Conference Paper
Full-text available
An increasing number of virtual reality applications require environments that emulate real-world conditions. These environments often involve dynamic virtual humans showing realistic behaviors. Understanding user perception and navigation among these virtual agents is key for designing realistic and effective environments featuring groups of virtual humans. While collision risk significantly influences human locomotion in the real world, this risk is largely absent in virtual settings. This paper studies the impact of the expected collision feedback on user perception and interaction with virtual crowds. We examine the effectiveness of commonly used collision feedback techniques (auditory cues and tactile vibrations) as well as inducing participants to expect that a physical bump with a real person might occur, as if some virtual humans actually correspond to real persons embodied into them and sharing the same physical space. Our results indicate that the expected collision feedback significantly influences both participant behavior—encompassing global navigation and local movements—and subjective perceptions of presence and copresence. Specifically, the introduction of a perceived risk of actual collision was found to significantly impact global navigation strategies and increase the sense of presence. Auditory cues had a similar effect on global navigation and additionally enhanced the sense of copresence. In contrast, vibrotactile feedback was primarily effective in influencing local movements.
Article
Detecting and quantifying abnormal crowd motion emerging from complex interactions of individuals is paramount to ensure the safety of crowds. Crowd-level abnormal behaviors (CABs), e.g., counter flow and crowd turbulence, are proven to be the crucial causes of many crowd disasters. Unlike individual-level anomaly, CABs usually do not exhibit salient difference from the normal behaviors when observed locally and the scale of CABs could vary from one scenario to another. It is also challenging to quantify the risk level of these CABs from video surveillance. In this paper, we present an improved version of our crowd motion learning framework for CABs detection, multi-scale motion consistency network (MSMC-Net) with a dual-attention fusion process to accommodate both the spatio-temporal and scale variations of different CABs. In addition, we propose an assessment method to quantify the risk level of detected CABs based on the anomaly score generated from our MSMC-Net. The risk quantification is performed in an online and accumulated manner and it can reflect the risk level of CABs consistent with other offline assessment metrics (e.g., crowd pressure), but without the extraction of detailed crowd data (e.g., pedestrian trajectories). For empirical study, we evaluate our method on large-scale crowd event datasets, including UMN, Hajj and Love Parade. Experimental results show that MSMC-Net could improve the AUC performance by 7.9%, 12.2% and 29.5% on three datasets respectively, compared to the best results of the state-of-the-art methods.
Article
Full-text available
Experiments as Code (ExaC) is a concept for reproducible, auditable, debuggable, reusable, & scalable experiments. Experiments are a crucial tool to understand Human-Building Interactions (HBI) and build a coherent theory around it. However, a common concern for experiments is their auditability and reproducibility. Experiments are usually designed, provisioned, managed, and analyzed by diverse teams of specialists (e.g., researchers, technicians, engineers) and may require many resources (e.g., cloud infrastructure, specialized equipment). Although researchers strive to document experiments accurately, this process is often lacking. Consequently, it is difficult to reproduce these experiments. Moreover, when it is necessary to create a similar experiment, the “wheel is very often reinvented”. It appears easier to start from scratch than trying to reuse existing work. Thus valuable embedded best practices and previous experiences are lost. In behavioral studies, such as in HBI, this has contributed to the reproducibility crisis. To tackle these challenges, we propose the ExaC paradigm, which not only documents the whole experiment, but additionally provides the automation code to provision, deploy, manage, and analyze the experiment. To this end, we define the ExaC concept, provide a taxonomy for the components of a practical implementation, and provide a proof of concept with an HBI desktop VR experiment that demonstrates the benefits of its “as code” representation, that is, reproducibility, auditability, debuggability, reusability, & scalability.
Article
Full-text available
Owing to the complexity of behavioral dynamics and mechanisms associated with turning maneuvers, capturing pedestrian movements around corners in a mathematical model is a challenging task. In this study, minimum jerk and one-thirds power law concepts, which have been initially applied in neurosciences and brain research domains, were utilized in combination to model pedestrian movement planning around bends. Simulation outputs explained that the proposed model can realistically represent the behavioral characteristics of pedestrians walking through bends. Comparison of modeled trajectories with empirical data demonstrated that the accuracy of the model can further be improved by using appropriate parameters in one-thirds power law equation. Sensitivity analysis explained that although the paths are not sensitive to the boundary conditions, speed and acceleration profiles could remarkably be varied depending on boundary conditions. Further, the applicability of the proposed model to estimate trajectories of pedestrians negotiating bends under different entry, intermediate and exit conditions was also identified. Proposed model can be applied in microscopic simulation platforms, virtual reality and driving simulator applications to provide realistic and accurate maneuvers around corners.
Chapter
Full-text available
It is suggested that the motion of pedestrians can be described as if they would be subject to "social forces. " These "forces" are not directly exerted by the pedestrians' personal environment, but they are a measure for the internal motivations of the individuals to perform certain actions (movements). The corresponding force concept is discussed in more detail and can also be applied to the description of other behaviors. In the presented model of pedestrian behavior several force terms are essential: first, a term describing the acceleration towards the desired velocity of motion; second, terms reQecting that a pedestrian keeps a certain distance from other pedestrians and borders; and third, a term modeling attractive effects. The resulting equations of motion are nonlinearly coupled Langevin equations. Computer simulations of crowds of interacting pedestrians show that the social force model is capable of describing the self-organization of several observed collective efFects of pedestrian behavior very realistically.
Article
Full-text available
Understanding and predicting the collective behaviour of crowds is essential to improve the efficiency of pedestrian flows in urban areas and minimize the risks of accidents at mass events. We advocate for the development of crowd traffic management systems, whereby observations of crowds can be coupled to fast and reliable models to produce rapid predictions of the crowd movement and eventually help crowd managers choose between tailored optimization strategies. Here, we propose a Bi-directional Macroscopic (BM) model as the core of such a system. Its key input is the fundamental diagram for bi-directional flows, i.e. the relation between the pedestrian fluxes and densities. We design and run a laboratory experiments involving a total of 119 participants walking in opposite directions in a circular corridor and show that the model is able to accurately capture the experimental data in a typical crowd forecasting situation. Finally, we propose a simple segregation strategy for enhancing the traffic efficiency, and use the BM model to determine the conditions under which this strategy would be beneficial. The BM model, therefore, could serve as a building block to develop on the fly prediction of crowd movements and help deploying real-time crowd optimization strategies.
Article
Full-text available
The collective behavior of human crowds often exhibits surprisingly regular patterns of movement. These patterns stem from social interactions between pedestrians such as when individuals imitate others, follow their neighbors, avoid collisions with other pedestrians, or push each other. While some of these patterns are beneficial and promote efficient collective motion, others can seriously disrupt the flow, ultimately leading to deadly crowd disasters. Understanding the dynamics of crowd movements can help urban planners manage crowd safety in dense urban areas and develop an understanding of dynamic social systems. However, the study of crowd behavior has been hindered by technical and methodological challenges. Laboratory experiments involving large crowds can be difficult to organize, and quantitative field data collected from surveillance cameras are difficult to evaluate. Nevertheless, crowd research has undergone important developments in the past few years that have led to numerous research opportunities. For example, the development of crowd monitoring based on the virtual signals emitted by pedestrians' smartphones has changed the way researchers collect and analyze live field data. In addition, the use of virtual reality, and multi-user platforms in particular, have paved the way for new types of experiments. In this review, we describe these methodological developments in detail and discuss how these novel technologies can be used to deepen our understanding of crowd behavior.
Article
Full-text available
Purpose The purpose of this study is to analyze the crowd dynamics of the visitors at Al-Masjid al-Nabawi during the most oversaturated period to characterize the most critical conditions and suggest technical solutions to accommodate visitors and provide them safe passage. Design/methodology/approach In this study, the time of entrance from the Al-Salam Gate to the tomb and from the tomb to the exit from the Al-Baqi’ Gate has been collected in the most oversaturated period. To be precise and to model the worst case, important crowd measures of effectiveness data are collected in the two holiest times considered by Muslims, during the holy month of Ramadan and the month of Dhul-Hijjah and during the busiest hours of the day to consider safety factors while proposing future solutions. The conventional manual head-counting method has been adopted to determine the crowd density and to carry out actual counting of the visitors from the recorded videos and photos captured by the legitimate authority. Findings The analyses revealed that the crowd dynamics in the month of Ramadan (peak) are statistically different from those for other times (off peak). In general, the crowd dynamics at all times on days other than Ramadan are almost identical. Originality/value The results of crowd characterization from this study are expected to help optimize crowd management in the Masjid at the most critical location and time. The data collected in this study could be used for future research to simulate similar crowd scenes or for even different crowd management scenarios in case of emergencies such as fire hazards or evacuation process. Keywords: Al-Masjid Al-Nabawi, Crowd dynamic, Visitor flow, Visitors density, Medina
Preprint
Understanding and predicting the collective behaviour of crowds is essential to improve the efficiency of pedestrian flows in urban areas and minimize the risks of accidents at mass events. We advocate for the development of a & “crowd forecasting system„whereby real-time observations of crowds are coupled to fast and reliable models to produce rapid predictions of the crowd movement and eventually help crowd managers choose between tailored optimization strategies. Here, we propose a Bi-directional Macroscopic (BM) model as the core of such a system. Its key input is the fundamental diagram for bi-directional flows, i.e. the relation between the pedestrian fluxes and densities. We design and run a laboratory experiments involving a total of 119 participants walking in opposite directions in a circular corridor and show that the model is able to accurately capture the experimental data in a typical crowd forecasting situation. Finally, we propose a simple segregation strategy for enhancing the traffic efficiency, and use the BM model to determine the conditions under which this strategy would be beneficial. The BM model, therefore, could serve as a building block to develop on the fly prediction of crowd movements and help deploying real-time crowd optimization strategies.
Article
In this study, we present a method to quantify the amount of congestion and the intrinsic risk in pedestrian crowds. Levels of congestion are estimated based on the velocity vector field obtained from the analysis of video recordings of moving crowds. By using data collected during supervised experiments, we show that the so-called “congestion level” allows to define a threshold for congestion under safe conditions and to measure the smoothness of pedestrian flows. The proposed approach has been compared with alternative quantities such as density, flow or the “crowd pressure” showing a more universal and consistent description of crowd motion. Later, the “crowd danger” of different pedestrian streams has been computed confirming that multidirectional motion is more dangerous than unidirectional one for equal levels of density. From a more practical perspective, the congestion level allowed to get a complete picture of the region in front of bottlenecks and to identify the formation of organized structures also under constant density and flow conditions. In addition, since only velocities are used in the computational process of the congestion level, it is more suitable for applications involving computer vision and emerging technologies, since density is usually difficult to obtain in very crowded situations. The congestion level and the crowd danger may help in the design of pedestrian facilities by simplifying interpretation of results from simulation and efficiently identify hotspots or design flaws. Finally, crowd control may benefit from the methods presented by potentially allowing a clear identification of dangerous locations during mass events.