ArticlePDF Available

Abstract and Figures

Microscopic agent-based simulations promise the meaningful inclusion of crowd dynamics in planning processes. However, such complex urban issues depend on a multiplicity of criteria. Thus, an isolated model cannot represent the walk of pedestrians meaningfully in planning contexts. This paper reframes crowd simulation as collaborative experimentation and embeds it directly in the design process. Beyond the simulation algorithm, this perspective draws attention to user interactions, interfaces, and visualizations as crucial simulation elements. Through a prototype, we combine an agent-based pedestrian simulation with a hybrid physical-digital interface. Based on this configuration, we explore requirements of the early design stages and accordingly discuss concepts for interaction, simulation, and visualization. The prototype blends user inputs with intuitive design interactions, adapts the simulation process to qualitative and dynamic negotiations, and presents results immediately in the discussed context. Thus, it aligns crowd simulation with contingent collaborations and reveals its potential in the early design stages.
Content may be subject to copyright.
Special Issue
International Journal of
Architectural Computing
2022, Vol. 0(0) 120
© The Author(s) 2022
Article reuse guidelines:
sagepub.com/journals-permissions
DOI: 10.1177/14780771221082258
journals.sagepub.com/home/jac
Collaborating with the crowd
Nick F ¨
orster, Ivan Bratoev, Jakob Fellner,
Gerhard Schubert and Frank Petzold
Abstract
Microscopic agent-based simulations promise the meaningful inclusion of crowd dynamics in planning
processes. However, such complex urban issues depend on a multiplicity of criteria. Thus, an isolated model
cannot represent the walk of pedestrians meaningfully in planning contexts. This paper reframes crowd
simulation as collaborative experimentation and embeds it directly in the design process. Beyond the
simulation algorithm, this perspective draws attention to user interactions, interfaces, and visualizations as
crucial simulation elements. Through a prototype, we combine an agent-based pedestrian simulation with a
hybrid physicaldigital interface. Based on this conguration, we explore requirements of the early design
stages and accordingly discuss concepts for interaction, simulation, and visualization. The prototype blends
user inputs with intuitive design interactions, adapts the simulation process to qualitative and dynamic
negotiations, and presents results immediately in the discussed context. Thus, it aligns crowd simulation with
contingent collaborations and reveals its potential in the early design stages.
Keywords
Urban design, architectural design, design decision support, pedestrian simulation, humancomputer interaction,
collaborative design, early design stages
Introduction
Planning in accordance with pedestrian movements is a many-faceted and complex issue.
1
Designing
comfortable and safe spaces for pedestrians depends on multiple criteria, perspectives, and stakeholders
ranging from ergonomic studies to building codes and subjective experience. De Certeau describes walking
the cityas complex and unforeseeable practice.
2
Hence, the planners’“strategicperspective from above
can never conceive the spontaneous movements and tactical ruses of urban crowds in their entirety. However,
this foundational aporia did not discourage urban analysts from trying precisely this. Instead of denying the
multilayered nature of urban systems, modelers embraced the idea of complexity, analyzing emergent local
TUM Department of Architecture, Technical University of Munich, Germany
Corresponding author:
Nick F¨
orster, TUM Department of Architecture, TUM School of Engineering and Design, Technical University of Munich, Arcisstraße
21, 80333, Munich, Bavaria, Germany.
Email: nick.foerster@tum.de
interactions in relation to macro-level patterns.
3
Analogously, advanced pedestrian simulations build upon
hybrid multi-scale approachesto describe crowd dynamics.
4
However, this synthesis of strategic planning and urban complexity is considered too narrow. Complex
or notoriously wicked”—problems cannot be black-boxed entirely into a simulation. Instead, they must be
resolved through the iteration of different alternatives in multi-lateral negotiations.
5
Even simulation experts
acknowledge the limitations of generalized pedestrian simulations regarding open-ended and multi-
dimensional urban problems.
6
Interdisciplinary researchers like Beaulieau et al. turn toward this second-
order complexity.
7
In a study of learning behavior, they address simulations as an open-ended socio-technical
negotiationcomparable to design processes as described by Rittel.
5
Simulation in design contexts is not to
be understood as the validation and accurate prediction of urban complexity but as the socio-technical
creation and exploration of possible futures. Following this perspective, the modeling of pedestrian
movement is not limited to an isolated and as accurate as possiblealgorithm. Instead, simulations in design
combine human collaboration and digital technologies reciprocally. Therefore, we turn toward the inter-
sections of this hybrid constellation. Besides simulation algorithms, interfaces, user interactions, and vi-
sualization techniques represent a critical part of knowledge generation and simulation practices.
8
Conventional pedestrian models separate the evaluation process clearly from the design interactions. Thus,
they frame the simulation as a second validating step following the design work. While we acknowledge the
motivation to validate nal designs reliably, this paper explores a different approach toward pedestrian
movement. We discuss the integration of pedestrian models into the early design stages focusing on col-
laboration processes and user interactions. Instead of trying to simulate pedestriansin the most reliable way,
we ask how to design with the crowd.
For this reason, the article examines the conditions of collaborative design contexts and discusses ad-
aptations to pedestrian models with respect to these requirements. By this, it extends the research from the
existing paper Designing Crowd Safety: Agent-Based Pedestrian Simulations in the Early, Collaborative
Design Stages.
9
We discuss how to adapt the simulation process to the contingencies, dynamics, and level of
decision-making of open-ended design debates. Furthermore, we analyze how collaborative interactions can
control the simulation model. Afterward, we discuss how to present simulation results so that they support the
design process. Lastly, these concepts are evaluated technically. Thus, this paper presents an extended
perspective on pedestrian simulation, including design studies, humancomputer interaction, and
visualization.
Related Work
Considering the apparent relevance of pedestrian movement for the design of buildings and public spaces,
pedestrian simulations in planning contexts receive notably little attention. While extensive research on the
behavioral and locomotive aspects of pedestrian movement exists,
10
relatively few studies discuss the
challenges of integrating them in planning processes. Nonetheless, the existing research in this domain
emphasizes the topics relevance. For instance, Aschwanden
11
shows the importance of pedestrian movement
for healthy and well-designed urban spaces and discusses how to address this problem through simulation. In
addition to this, Asriana and Indraprastha elaborate on the integration of crowd simulations in heterogeneous
urban contexts and their potential to contribute to the early stages in planning.
12
Chen
13
discusses the
adaptation of pedestrian models to specic behavioral patterns of urban streets. Furthermore, Wurzer
14
approaches the design of hospitals through pedestrian models and analyses how these can contribute to
traditional methods of spatial programming. A recent study by Clever et al. discusses a BIM-based integration
of pedestrian simulation in design scenarios.
15
The paper suggests a machine-learning approach to react
responsively to frequent design input.
2International Journal of Architectural Computing 0(0)
This existing research represents a valuable contribution to the topic of pedestrian simulation in design.
However, these studies predominantly adapt crowd models thematically to planning tasks (relevant scales and
criteria) and focus on the simulation model itself. This work does not cover idiosyncrasies of the early design
stages like interactivity, vague input parameters, or appropriate visualization strategies. With respect to our
topic, these aspects require further scrutiny.
At the same time, the integration of energy and environmental simulations in the early, collaborative design
stages is well explored. This eld poses similar challenges: computation heavy and effortful analyses are
adapted to tactical decision-making and dynamic negotiation processes. An extensive corpus of research
exists concerning topics like the low resolution of input parameters in early planning stages,
16
the integration
of results into the design process,
17
and the effects of simulation tools on the structure of design processes.
18
In addition to this, research associated with the presented project investigated how simulations can support
collaborative negotiations on urban heating networks,
19
how solar-potential analysis can inform interactive
and dynamic design discussions,
20
and how digital tools can be incorporated into multi-modal design
practices.
21
Project Setup
The presented project incorporates behavioral crowd simulations in the early planning phases. The project builds
upon the existing pedestrian simulation crowd:it
22
and the CDP//Collaborative Design Platform (CDP)
23
an
interactive platform dedicated to the early design stages. Experimentally we couple these modules by means of a
communication protocol.
crowd:it
The pedestrian model crowd:it resulted from Angelika Kneidls
24
research on pedestrian locomotion and
navigation. In addition to this, the tool incorporates the Optimal Steps Model,which was developed by
Michael Seitz and Prof. Gerda K ¨
oster.
22,25
These concepts describe microscopic pedestrian interactions,
biomechanical aspects,
26
as well as mesoscopic navigation behavior.
4
By this, the model simulates spatially
indiscreet crowd movements and considers local events, like cueing or congestion. Currently, the tool is
deployed to evaluate elaborated design proposals, for example, to prove a buildings safe evacuation or the
performance of a highly frequented area. Integrating the tool into open-ended design processes poses several
challenges: The simulation setup and the interpretation of results are cost-intensive and presuppose expertise.
The accuracy andrelated to thisthe computation effort must be adjusted. Furthermore, the input and
output target single-user desktop machines and should be modied to match collaborative discussions. Also,
the simulation process should be revised to match unpredictable design interactions.
CDP//Collaborative Design Platform
Conventional desktop computer interfaces are inappropriate for dynamic collaboration.
27
For this reason, we
propose controlling simulation models through intuitive multi-modal interactions. The employed Design
Platform (see Figure 1) offers a touchand object-sensitive screen on which physical models can be placed
and discussed. The planning context is directly imported from OpenStreetMap.
28
Collaborators intuitively
interact with the physical model on the table. Meanwhile, the platform recognizes these design interactions in
real-time via a 3D depth camera and uses a recreated digital model as a foundation for simulation and
visualization. The conceptual and technological base of this seamless physical-digital platform has been
developed and optimized during the last 10 years and has been published widely. A detailed description of the
F¨
orster et al. 3
Design Platform can be found in Interaction Forms for Digital Design.
23
By this, we combine qualitative
(e.g., spatial and aesthetical) and quantitative (e.g., simulation) modes of decision-making.
Communication Protocol
We introduce a software protocol that interconnects the platform as an input and output device with a server-
based simulation (see Figure 2). This modular setup increases the prototypesexibility: adaptations to the
simulation, the server, or the input devices are feasible with little effort. On the one hand, this setup allows the
connection of existing software solutions with only minor modications. On the other hand, it drastically
reduces the performance requirements for the input device. Thus, extending the servers capacity makes high
computational capacities for extensive and fast simulations available.
The protocol is implemented as a serverclient architecture following the open-source standard ZeroMQ.
29
As the client, the Design Platform was extended with a C# plugin that structures user input and visualizes
simulation output. An API was developed for the simulation model on the server-side, which allows
controlling it through the protocol.
Figure 1. The Design Platform blends analog design tools like physical models with digital simulations and visualizations.
4International Journal of Architectural Computing 0(0)
Concept and Prototype
Designers gather around the Design Platform and examine the planning site. While discussing, they move and
modify styrofoam models on the table. Intuitive design interactions with the models presuppose no specic
technical expertise and thus impose no additional hierarchies on the collaborators. These interactions happen
spontaneously, experimentally, and simultaneously. They include several hybrid modes of expression like
discussion and physical interactions. In this process, design proposals are articulated materially in several
alternatives and co-exist as different fuzzy possibilities.
30
Instead of imposing a rigid technological structure onto these practices, the simulation must adapt to these
dynamic and complex conditions. The tool follows the design interactions and responsively contributes
information to the debate (see Figure 3). Instead of evaluating nal results and returning evidence,the
simulation contributes a continuous stream of hints and suggestions. By this, designers take pedestrian
dynamics into account and juxtapose this information with other relevant design goals. Therefore, the
simulation fosters a non-deterministic engagement with the factor of crowd dynamics. This setup allows
designers to test the effects of adaptations and to improve their design in fast iterations.
To achieve this goal, we explore the described combination of tools in detail regarding the simulation
process, user interactions, and the visualization of results.
Simulation
As Petersen et al. show, advanced simulations tend toward detailed evaluation instead of constructive design
support.
18
Conventionally, they imply long design iterations and subsequent performance evaluation. By
embedding the model into the design process, it provides direct suggestions for planners.
19
However,
conventional crowd simulations are designed for long iterations and after-design use: They imply a linear
sequence of extensive input, black-boxed calculation, and nally, visualization of results. After receiving
input, the model simulates the agentsjourney step-wise from their sources to their targets. During this
Figure 2. The prototype combines the Design Platform, a hybrid collaborative interface, with a server-based simulation
by means of a communication protocol.
F¨
orster et al. 5
process, the simulation stays passive and gives no feedback. Changing input parameters requires aborting and
restarting the calculation.
Even on a highly performant server, this process endures several minutes for extensive simulations. In
heated debates, these calculations are likely to be interrupted by design interactions. Therefore, we propose an
adapted continuoussimulation process that immediately reacts to changed user input and constantly
supplies feedback.
After design interactions, the communication protocol exibly starts a new simulation and aborts the
running one. Thus, the model allows interactions during runtime and reacts directly. This approach was
preferred over changing input objects live during the simulation. It is important to clear the simulation history
after changed input to keep averaged simulation results conclusive (like the mean densityof an area or the
agentstrajectories). Thus, the model stays interactive but still provides accurate results.
Furthermore, the protocol must return intermediary feedback as soon as it becomes available. For this
reason, we implement an observer-class through the simulation-API. After each simulation step, this class
transmits all agent positions to the plugin on the client side. Thus, design interactions entail direct feedback.
When a variant rests on the table and is discussed in more detail, the simulation proceeds and contributes more
information.
Figure 3. The project targets the integration of pedestrian simulation in the hybrid digital-physical design interactions of
the Design Platform. The physical objects serve as sources and targets through which agents enter the simulation and
leave the simulation space.
6International Journal of Architectural Computing 0(0)
Thus, we have found an interactive way to control the simulation and provide constant feedback.
Nonetheless, the simulation still restarts after interactions: pausing, clearing the simulation area of all agents,
and restarting them at their sources. We propose two measures to compensate for this issue.
Conventional pedestrian models generate a specied number of agents in source areas and simulate how
they move toward their targets where they ultimately leave the simulation (see Figure 4). For this reason, the
agents initially evaluateareas close to their sources and later on the target areas. The simulation treats the
design space anisotropically, which is misleading in design debates.
Figure 4. Whereas the standard simulation model simulates a progressivemovement of agents with a xed start and
endpoint, the proposed model aims at an instantaneous and continuousdistribution.
F¨
orster et al. 7
Whereas the described linear simulation process is accurate concerning absolute evacuation or entrance
times, early design discussions rather prot from spatial feedback on problematic areas and qualitative
movement patterns. Thus, we need to achieve a continuous distribution of agents in the relevant design areas.
For this reason, we adjust the generation of agents: Initially, agents spawn not only in the source areas but also
in the eld between sources and targets. When a simulation is aborted, agents are re-spawned in their former
position and keep moving toward their respective targets (if unmodied). The protocol treats each simulation
as one topicon which it communicates messages with simulation results. Instead of stopping a simulation
instantaneously after user interaction, the old simulation keeps running for a short time while the new one
initializes. This strategy allows blending the simulation results of new and old simulations, enabling a more
continuous user experience.
Interactions
Earlier studies on digital tools in collaborative design processes suggest connecting simulations with multi-
modal design interactions like sketching and modeling.
21
Instead of effortful settings, the simulation must be
controlled with the intuitive geometric interactions of the Design Platform. Though, how to control an
intricate simulation algorithm through styrofoam blocks? For this purpose, we conduct two translations: On
the one hand, we categorize settings as front and back end parameters. On the other, we investigate how to
mediate design interactions to control these settings. We must consider an appropriate level of detail in the
early design stages regarding inputs and feedback: Overly specic input parameters are not available in the
early design stages
16
or require a complicated setup process. Too detailed results are unnecessary for early
decision-making or may even distract the users.
Front and Back End Parameters. The protocol communicates three types of input data to the simulation server:
·Geometric input (the simulation objects);
·Agent settings (agent properties and their way paths); and
·Foundational simulation settings (e.g., accuracy).
Regarding the relatively low precision and large scale of early design stages, agent settings like size and
speed can be assumed as a standard distribution regardless of the specic context. All design-specic setup
derives from intuitive geometric interactions and the choice of a benchmark scenario. Geometric interactions
inuence where pedestrian agents spawn and go, whereas the benchmark scenario denes how they move
through the geometry.
Geometric Interactions. During the design process, stakeholders test spatial compositions, examine conse-
quences, and respectively adapt the design setting. Thus, geometric operations represent one of the primary
activities during the design process. For this reason, they are the most naturalway to structure user input
during the design process.
The Design Platform allows importing urban contexts directly from OpenStreetMap. Physical objects on
the table represent elements that are subject to the design process. They serve as a haptic interface. The
platforms 3D scanner recognizes each object with its RGB-D sensor and assigns it a unique ID. By this, the
platform interprets changes in the design model automatically as adapted user input.
Hence, designers work with the physical design and the simulation model simultaneously (see Figure 5).
Each objects height denes its signicance for the pedestrian simulationits attraction.The styrofoam
color species the element type: entrances/exits are red, minor attractions (e.g., info-points) yellow, major
attractions (e.g., festival stages) blue. As soon as the platform recognizes a simulation object, another option
8International Journal of Architectural Computing 0(0)
for interaction appears: A bold contour surrounds the object-perimeter on the table. The selection of these
lines sets the respective edge as a barrier. By this, a simulation object can be given a specic direction. Other
barriers can be created via hand gestures: A long touch on a free area sets a marker that tracks the ngers
movement. After releasing the gesture, this path is translated into a barrier.
Reduction to Benchmark Scenarios. While the geometric setup denes the examined space, other settings
congure the agentsmovement between entrances/exits, major and minor attractions. Conventionally, this
setup entails manually assigning agents to sources, several intermediate targets, and main targets. However,
such extensive parameters often are unavailable or insecure in the early design stages.
16
In a study on energy
evaluation tools, Attia et al.
17
suggest that un-intuitive user input often overcharges designers as well as
laypersons. Therefore, he argues that the early design stages instead prot from benchmarking and the
possibility to compare alternativesthan evaluating absolute values.
17
Hence, we simplify and condense the
complicated simulation setup to few crucial benchmark scenarios.This allows the examination of a design
variant from few crucial perspectives. This strategy is also followed by Roetzel,
16
who reduces the complex
calibration of occupant behavior to a best and worst-case scenario in the early design stages. Instead of these
extremes, we test design variants against foundational movement patterns: entrance, exit, and stay. Thus, we
reduce the effortful conguration of way paths to three foundational scenarios. Instead of accurate
Figure 5. Intuitive design interactions with physical objects serve as simulation input. The objectscolors and sizes
represent their signicant in the planning context (e.g., red = entrance/exit and yellow = minor attraction).
F¨
orster et al. 9
quantitative results like evacuation time or absolute building performance, the simulation contributes spatial
information about pedestrian movement, comfort, and safety.
Based on this scenario selection, the plugin of the Design Platform interprets recognized design objects as
specic simulation objectsagent sources, intermediate targets, main targets, and barriers: During an en-
trance scenario, agents spawn at an entrance/exit, move toward the minor attractions, and nish at the main
attraction. An (emergency) exit simulation interprets main attractions as starting points and the entrance/exit
as the targetall other objects become obstacles. During a stay scenario, agents start at the entrance/exit,
move toward minor and major attractions and nally return to the entrance/exit.
From this classication, we generate way paths automatically: A recursive algorithm creates several
branching paths from each source. Based on a list of unvisited targets, their attraction(the physical objects
height), their distance, and a minimal amount of targets to visit, the algorithm semi-randomly weights
different targets and composes a list of targets for each agent. As soon as the Design Platform recognizes a
source and a target, it automatically generates way paths and spawn intervals. These intervals are created for
an extended interval of 8 hso that a complete simulation run would depict 1 day, which is divided into 48
spawn intervals of 10 min.
Visualization
In a study on design simulation in the early design stages, Attia et al. describe several requirements to embed
simulation results in the early design stages. They suggest that analytical results must be presented intuitively
and directly in the design context.
17
The simulation must provide responsive feedback to be informative
during the decision-making process.
17
Furthermore, conventional simulation results are hard to interpret for
laypersons and even for designers. Therefore, we need to scrutinize visualization methods that are intuitively
understandable in the planning context. Instead of accurate results, they must indicate design problems and
provide spatial suggestions. The Design Platform allows the presentation of simulation results on the table:
They directly augment the design model (see Figure 6). Hence, results are well visible for all collaborators.
Furthermore, the relationship between input adaptations and simulation output is comprehensible.
In the chapter Simulation,we have shown how to achieve continuous stream of results. For each
simulation step, an XML le containing all agent positions is provided. Since this format is not intuitively
readable, we explore visualization methods to support the planning process. Furthermore, we discuss how to
generate preliminary ad hoc results as responsive feedback.
Heat Cloud. Absolute numbers and precise graphs are not directly related to the design context and cannot be
interpreted easily as planning suggestions.
17
Furthermore, they imply a sense of denitiveness, which is
inappropriate for the contingent exploration of design variants. Thus, we need to engage critically with the
topic of visualizations to make pedestrian simulations relevant in collaborative situations. Simulation results
need to be selected and translated regarding the key factors of our benchmark scenarios. Instead of absolute
numbers, Roetzel
16
suggests representing complex processes by ballpark guresto make results com-
parable to other quantitative and qualitative criteria. Following this perspective, designers should receive
spatial information that is commensurable with the design process.
The visualized indicator must represent different design targets relevant in our benchmark scenarios:
pedestrian comfort, safety, velocity. Thus, we introduce the Level of Serviceindicator, which is a
foundational concept in trafc planning
31,32
and relates density with velocity.
33
Therefore, it concerns
pedestrian movement, comfort, and safety and is relevant in all three benchmark scenarios. Five Levels of
Service are differentiated, ranging from a free ow of pedestrians to potentially dangerous congestion.
34
This information must be presented spatially and directly in the design model. Crowd movement patterns
become visible and spatial issues like bottle-necks or building performance can be discussed. Classical
10 International Journal of Architectural Computing 0(0)
visualizations for crowd ows are animated agents,”“movement paths,and heatmaps (see Figure 7).
Following the moving agents gives an immediate insight into the simulation. However, it is hard to compare
relevant features such as the velocity and density at different locations. Movement pathsdisplay each
agents trajectory as a curve and reveal general movement patterns. Other information like density and
dynamic factors stays incomprehensible. Lastly, heatmaps represent average values in a discreet areal grid.
Thus, they can depict various information types such as occupancy, signicant congestion, or frustration.
35
In the design process, heatmaps demonstrate potential problems and make different areas easily comparable.
Nonetheless, the areal discretization is not ideal for the representation of uid crowd movement. Furthermore,
covering the entire simulation space with a heatmap may allege a false sense of completeness when only
preliminary results are presented.
The reviewed visualization methods reveal different relevant aspects of crowd movement. However,
switching and comparing them is complicated and confusing. For this reason, we combine these approaches
to a Heat Cloudvisualization, which presents the Level of Service indicator, as well as spatial crowd
dynamics (see Figure 7).
Figure 6. The visualization of simulation results directly augments the design model. Clouds of congestion events
agglomerate between sources and targets.
F¨
orster et al. 11
The protocol communicates each simulation step to the plugin on the client side. For each agent, the
number of agents in a radius of 1 m is counted. This result signies the local pedestrian density and the agents
respective Level of Service. When the agent is congested, its position is added to the Heat Cloud. Instead of
generating a heatmap over the entire design space, the Heat Cloud stores one point for every congestion event.
They appear as tiny spheres in a hue representing the congestion level: Low densities (1 person) are yellow,
while red represents grave congestion (6 persons). This visualization method not only demonstrates
problematic areas but also reveals the crowd dynamics leading to these situations (see Figure 8).
As soon as more than one simulation step is calculated, each sphere is rendered with transparency equal to
1/n, where nis the number of simulation steps already received plus 1. The transparency range is dened from
[0, 1]. This approach allows increasing the resolution of simulation results gradually. After a simulation
restart, initial events indicate problematic spots and soon conglomerate to ner-grained patterns.
While the Level of Service concept categorizes congestion events on an absolute scalemaking it easier to
understand the gravity of each eventthe eventsrelevance is relative to the simulated time. By this, we avoid
a false sense of exhaustiveness.
Figure 7. Visualization approach: Classical display methods for crowd simulations are combined. The Heat Cloud
correlates information on agentsmovement, their trajectories, and local pedestrian densities.
12 International Journal of Architectural Computing 0(0)
Figure 8. Mockup of the Heat Cloud: The visualization method gives insights into movement patterns as well as
pedestrian densities in the design context.
Figure 9. Based on the results of the initial calculation stepstargets, navigation, and intermediate movement steps
preliminary feedback can be presented. User interactions entail responsive feedback, which becomes more accurate
step by step.
F¨
orster et al. 13
Display of Preliminary Results. The Heat Cloud is rened more and more with each simulation step. Depending
on the size of the scenario, however, the simulation does not deliver responsive feedback. However, early
decision-making processes entail experimental and frequent interactions. This open-ended and contingent
solution nding must be supported with instantaneous feedback. By this, simulation results become a hint for
tentative design adaptations and support agile problem-solving.
18
Studies concerning solar models in the early design stages propose using multiple calculation processes of
varying accuracyand therefore computation time.
20
Thus, the rst results can be visualized ad hoc and are
replaced with accurate feedback later. Besides the step-wise calculation of agentsmovement, the pedestrian
simulation entails two additional processes: The generation of way paths (see chapter Interactions) and the
calculation of navigation elds.
These way paths are calculated on the client side and are available instantaneously and independently from
the simulation process. These polylines connecting the agentstargets can be subdivided with respect to the
agentsstep length and indicate virtualagent movement (see Figure 9: step 1). This gives the rst im-
pression of global movements and density. Before the rst simulation step, the model creates the navigation
eld which directs agents toward their targets. Based on this eld, the rst estimation of pedestrian movement
can be rened (see Figure 9: step 2). As soon as the server returns the actual simulation steps, these virtual
positions, which do not take into account interactions with other agents or complex locomotion, are updated
(see Figure 9: step 3). By this, the model responds immediately with initial results and renes the output
successively. Thus, it answers design interactions directly with rst hints instead of interrupting the col-
laboration with passive waiting time.
Evaluation
We have described concepts to integrate pedestrian simulations as collaborative tools using design inter-
actions as simulation input and reacting with direct and instantaneous feedback. We delegated computation
heavy calculations to an external server, adapted the simulation process to dynamic design situations, and
examined appropriate and responsive visualization methods. Consequently, we conduct a technical evaluation
of this prototype, focusing on its performance in relation to design interactions.
Regarding the meaningful and non-interruptive usage of simulations in collaborative processes, we
scrutinized the following questions: When is an interaction answered with relevant feedback? Where in the
proposed serverclient conguration emerge delays? Can these delays be compensated through the concepts
elaborated in the earlier chapters?
We divide the visualization and interaction device (front-end) from the computational core of the sim-
ulation (back end). For evaluation, we used a computer with the hardware in Table 1 as a server and created a
client on a computer with the hardware specied in Table 2. This client logs each message received through
the protocol with the time point of its arrival. We deployed a study area of 660*400 m for this evaluation and a
navigation accuracy of 0.25 m. The scenario ran multiple times, each time with an increasing number of
Table 1. Server hardware.
Component Model Specication
CPU AMD Ryzen 5 3600 6x 3.6 GHz/4.2 GHz Turbo
RAM Corsair DDR4 3000 MHz Vengeance LPX 64 GB (4 × 16 GB)
GPU MSI GeForce RTX3070 Ventus 2X OC 8 GB GDDR6
SSD Crucial MX500 1000 GB (560 MB/s510 MB/s)
Mainboard MSI B550-A PRO
14 International Journal of Architectural Computing 0(0)
agents spawning at each interval (50, 100, 250, and 500) to measure the dependency of computation time on
the increase of agents and the prototypes behavior in different scenarios.
Interactions with the physical objects of the Design Platform entail a change in the simulation settings.
These are communicated via the protocol, which initializes a new simulation on the server-side. As Figure 10
shows, this initiation causes a signicant overhead for the server performance. These preprocessing steps
include creating the navigation eld as well as the allocation of memory needed for the simulations overhead.
In total, these steps took between 2500 ms and 12,000 ms for the described test scenarios (see Table 3).
This interruption of several seconds after each interaction represents a problematic break in the user in-
teractions. Therefore, initial feedback is given using the agentstargets (as described in the Visualization chapter).
Because this information is generated on the client side, it is instantaneously available as feedback. This
displayed tendency bridges the gap until simulation results become available.
After the preprocessing phase, the server simulates the agentsmovement round-wise. In the scenario with
500 agents per spawn interval, one timestep was computed in 10.52 ms on average with a deviation of
11.18 ms. Consequently, a full-day simulation of 8 h requires around 600 s to complete. In comparison, a
simulation with only 250 pedestrians spawning in regular intervals would require 256 s since it has an average
computation time of 4.45 ms per timestep. We observed some additional delays during the rst few steps due
to the lazy initialization of simulation parts. Figure 11 indicates that the computation speed of each step
linearly correlates with the number of agents in the simulation: The computation time for one step is given as
f(x)=x*0.02 where xis the number of currently active agents in the simulation. This limitation stems from the
simulation software used, which does not fully exploit multi-core machines and largely depends on the java
virtual machine conguration setup.
Several minutes of computation time is an unacceptable delay in the design process. Therefore, we have
introduced the visualization of intermediary simulation results. After each simulated timestep, the protocol
returns the agentsmovement steps to the client. There it is presented as intermediary feedback, which extends
as the simulation continues. In the test scenario with 500 agents, the server simulates 1 minute of agents
movement in approximately 1 s computation time. With an agent speed of 1.4 m/s, after 1 s of computation
time, the agents moved about 80 m in the simulation space. Since the agents are initially distributed between
the sources and targets (as described in the Simulation chapter), results for the whole simulated area become
available fast. This feedback could be enhanced through visualizations based on the navigation graph, which
is available after the preprocessing phase (as discussed in the Visualization chapter). However, this visu-
alization method was not part of the technical evaluation.
Finally, we measured the protocol performance on the client side. Figure 12 demonstrates that the server
requires a similar time as the simulation to receive the results for each step. Consequently, the protocol itself
introduced no detrimental overhead in the communication process. The most signicant disruption of the
simulation occurred due to small instabilities in the network connection, which appear as the random peaks in
the graph.
Table 2. Client hardware.
Component Model Specication
CPU Intel Xeon E5-2623 4x 3.0 GHz/3.5 GHz Turbo
RAM 16 GB (4 × 4 GB)
GPU NVIDIA GeForce GTX 1060 6 GB GDDR6
SSD Samsung SM871 2.5 256 GB (443 MB/s413 MB/s)
Mainboard Dell 0KJCC5-A00
F¨
orster et al. 15
Figure 10. Server computation time per simulation step for four test scenarios: The rst computation-intensive steps
are part of the initialization, while the following, faster steps represent the agentsmovement.
Table 3. Comparison of simulation time for preprocessing steps and the average simulation step between the different
test scenarios.
Agent count Preprocessing phase, ms Average step, ms Simulated/calculation time proportion
50 2500 0.8 625
100 5000 1.32 379
250 8500 4.21 119
500 12 000 10.30 49
Figure 11. Dependency Comparison between active pedestrian agents and computation time per simulation step: The
partial graph for the fourth scenario (500 agents per spawn interval) indicates a direct proportionality.
16 International Journal of Architectural Computing 0(0)
In conclusion, the initialization step causes a delay of a few seconds, which can be compensated by
visualizing preliminary results based on the agentstargets. The following rounds of agentsmovement
compute fast but still endure several minutes. We can return intermediary results as instantaneous feedback
through the suggested adaptations to the simulation and the proposed visualization methods. The protocol
itself causes no problematic delays. In total, this allows for real-time integration of the pedestrian simulation
into dynamic planning processes.
In addition to this performance-focused analysis, the proposed methods should be evaluated in design
processes in a future step. This could be achieved in collaboration with our research partner, who also consults
in planning processes. Two projects could be comparedone conducted with external simulations and one
employing the presented integrated methods. Besides a qualitative evaluation of the tool in collaborative
situations, we could compare these planning processes and assess their progress regarding pedestrian
movement and related criteria.
Conclusion
Based on the understanding of simulation as a collaborative process of generating and exploring possible
design variants, this paper investigated an alternative perspective on crowd models. Acknowledging the
inherent complexity of urban practicessuch as the walk of pedestrianswe emphasize the necessity to
include different perspectives in the reciprocal process of design and simulation. Following this sensibility,
we focus on the requirements of collaborative design: intuitive interfaces, the dynamics of the design ow,
and adequate presentation of inputs and results. Building upon the existing Design Platform and the
simulation crowd:it we have shown how to embed pedestrian simulation in the early design stages. By
connecting these elements with a communication protocol, we created an experimental setup to test and
investigate the challenges and potentials of microscopic crowd simulations in collaborative design processes.
The presented research makes information on crowd movement available in the early design stages. It
empowers planners to integrate this knowledge into their foundational planning decisions and enhance their
design regarding pedestrian comfort and safety. Thus, these criteria can be brought into accordance with other
Figure 12. Client side processing time of received simulation steps, including data transfer from the server. Compared
to the server computation time, no signicant delay is noticeable.
F¨
orster et al. 17
relevant aspects. Including this perspective in the initial design decisions prevents effortful changes in later
planning phases.
The resulting concepts and prototypes form a foundation for further exploring the integration of crowd
simulation in design processes. On the one hand, this approach could be extended by an in-depth study of the
user perspectiveconsidering collaboration and participatory design in more detail. By this, the study could
specically focus on the role of the adapted simulation in concrete design discussions. On the other hand, the
projects modular setup allows experimentation with different simulation approaches and interfaces. For
instance, studies like the mentioned one of Clever et al.
15
suggest improvements to the simulation core. Such
methods could be integrated into the presented prototype to test their potential combined with specic
interfaces and user interactions. Thus, the presented protocol becomes an infrastructure to explore crowd
simulations and their integration in collaborative design scenarios.
Declaration of conicting interests
The author(s) declared no potential conicts of interest with respect to the research, authorship, and/or publication of this
article.
Funding
The author(s) disclosed receipt of the following nancial support for the research, authorship, and/or publication of this
article: This project was funded by the research initiative Zukunft Bauof the German Federal Ofce for Building and
Regional Planning (Aktenzeichen: SWD-10.08.18.7-18.12) and accu:rate GmbH, Munich, Germany.
ORCID iDs
Nick F¨
orster https://orcid.org/0000-0002-4274-8127
References
1. Gonzalez-Urango H, Inturri G, Le Pira M, et al. Planning for pedestrians with a participatory multicriteria approach.
J Urban Plann Develop 2020; 146(3 05020007), doi:10.1061/(ASCE)UP.1943-5444.0000585.
2. de Certeau M. The Practice of Everyday Life. 3rd ed. Berkely: University of California Press, 2011.
3. Batty M. The New Science of Cities. Camebridge: The MIT Press, 2013.
4. Kneidl A, Hartmann D and Borrmann A. A hybrid multi-scale approach for simulation of pedestrian dynamics.
Transportation Res C: Emerging Tech 2013; 37: 223237, doi:10.1016/j.trc.2013.03.005.
5. Rittel HWJ and Webber MM. Dilemmas in a general theory of planning. Pol Sciences 1973; 4(2): 155169.
6. Aschwanden G. Health and Placean analysis of the built environments impact on walking behavior and health.
PhD Thesis: ETH Zurich 2014.
7. Beaulieu A, Ratto M and Scharnhorst A. Learning in a landscapesimulation-building as reexive intervention.
Mind Soc 2013; 12(1): 91112.
8. Calvillo N. Digital Visualizations for Thinking with the Environment. In: J Vertesi and D Ribes (eds) Digital STS.
New Jersey: Princeton University Press, 2019, pp. 6176.
9. F¨
orster N, Bratoev I, Fellner J, et al. Designing Crowd SafetyAgent-Based Pedestrian Simulations in the Early,
Collaborative Design Stages. In: Globa A, van Ameijde J, Fingrut A, Kim N, and Lo TTS (eds) PROJECTIONS -
Proceedings of the 26th CAADRIA Conference. Hong Kong: the Chinese University of Hong Kong and Online, 2021,
2, pp. 729738.
10. Weidmann U, Kirsch U and Schreckenberg M. Pedestrian and evacuation dynamics 2012. Springer International
Publishing, 2014.
11. Aschwanden G. Agent-based social pedestrian simulation for the validation of urban planning recommendations.
16th Iberoamerican Congress of Digital Graphics (SIGRADI) 2012; 332336, doi:20.500.11850/58147.
18 International Journal of Architectural Computing 0(0)
12. Nova A and Aswin I. Making sense of agent-based simulation: Developing design strategy for pedestrian-centric
urban space. Proceedings of the 21st International Conference of the Association for Computer-Aided Architectural
Design Research in Asia (CAADRIA) 2016; 343352.
13. Chen CH. The design of an interactive scenario-based agent simulator for supporting the early stages of urban design.
In Proceedings of the 13th International Conference on Computer Aided Architectural Design Research (CAADRIA),
2008: 4248.
14. Wurzer G. In-process agent simulation for early stages of hospital planning. Math Computer Model Dynamical Syst
2013; 19(4): 331343.
15. Clever J, Abualdenien J and Borrmann A. Deep learning approach for predicting pedestrian dynamics for trans-
portation hubs in early design phases. 28th International Workshop on Intelligent Computing in Engineering. Berlin:
Germany, 2021, pp. 112.
16. Roetzel A. Occupant behaviour simulation for cellular ofces in early design stagesArchitectural and modelling
considerations. Building Simulation 2015; 8: 211224, doi:10.1007/s12273-014-0203-6.
17. Attia S, Gratia E, De Herde A, et al. Simulation-based decision support tool for early stages of zero-energy building
design. Energy and Buildings 2012; 49: 215, doi:10.1016/j.enbuild.2012.01.028.
18. Petersen S and Svendsen S. Method and simulation program informed decisions in the early stages of building
design. Energy and Buildings 2010; 42(7): 11131119, doi:10.1016/j.enbuild.2010.02.002.
19. Bonnet C, Bratoev I, Chokhachian A, et al. Designing and Evaluating District Heating Networks with Simulation
Based Urban Planning. In: Burlando M, Canepa M, Magliocco A, et al. (eds) International Conference on Urban
Comfort and Environmental Quality. University of Genoa, Italy: Genova University Press, 2017.
20. Bratoev I, Schubert G, Chokhachian A, et al. Interactive Solar Potential Simulation for Early Stage Urban Planning.
BAUSIM 2018; 407411.
21. Schubert G, Artinger E, Petzold F, et al. Tangible tools for architectural designseamless integration into the
architectural workow. Proceedings of the 31st Annual Conference of the Association for Computer Aided Design in
Architecture (ACADIA) 2011; 252259, In press.
22. accu:rate. Our simulation model: What is behind it? https://www.accu-rate.de/en/software-crowd-it-en/ (2021 ac-
cessed 12 August 2021).
23. Schubert G. Interaction Forms for Digital Design. Munich: TUM.University Press, 2021.
24. Kneidl A. Methoden zur Abbildung menschlichen Navigationsverhaltens bei der Modellierung von Fußg ¨
an-
gerstr¨
omen. PhD Thesis: Technical University of Munich, 2013.
25. Seitz MJ. Simulating pedestrian dynamics. PhD Thesis: Technical University of Munich, 2016.
26. Seitz MJ and K¨
oster G. Natural discretization of pedestrian movement in continuous space. Phys Rev E 2012;
86(2 046108), doi:10.1103/PhysRevE.86.046108.
27. Kunze A, Burkhard R and Gebhardt S. Visualization and Decision Support Tools in Urban Planning. In: Arisona SM,
Aschwanden G, Halatsch J, et al. (eds) Digital Urban Modeling and Simulation. Communications in Computer and
Information Science. Berlin: Springer, 2012, 242, pp. 279298.
28. Petzold F, Schubert G, Mühlhaus M, et al. GIS vs. BIM: semantische Modelle in frühen Planungs- und En-
twurfsphasen. In: Kolbe T, Bill R, and Donaubauer A (eds) Geoinformationssysteme. Berlin: Wichmann, H, 2015,
pp. 3546.
29. The ZeroMQ authors. Zeromq Documentation. https://zeromq.org/ (2021 accessed 12 August 2021).
30. Yaneva A. Scaling Up and Down. Soc Stud Sci 2005; 35(6): 867894.
31. Oeding D. Verkehrsbelastung und Dimensionierung von Gehwegen und anderen Anlagen des Fussg¨
angerkehrs.
Technische Hochschule Braunschweig Inst für Stadtbauwesen, 1963.
32. Fruin JJ. Designing for pedestrians: A level of service concept. Polytechnic Inst Brooklyn, 1970.
33. Holl S. Methoden für die Bemessung der Leistungsf ¨
ahigkeit multidirektional genutzter Fußverkehrsanlagen.
Forschungszentrum Jülich GmbH, 2016.
F¨
orster et al. 19
34. Bitzer TF. Mikroskopische Modellierung und Simulation des Fußg¨angerverkehrs. PhD Thesis: University of
Stuttgart, 2010.
35. Kretz T. Pedestrian Trafc Simulation and Experiments. PhD Thesis: University of Duisburg-Essen, 2007.
20 International Journal of Architectural Computing 0(0)
ResearchGate has not been able to resolve any citations for this publication.
Book
Full-text available
In the current day-to-day life of the architect, computers are used largely in the later, executive phases of planning – the early, creative phases remain mostly unaffected by this technology. The architect still designs using working models and hand-drawn sketches. However, digital calculations, analyses and simulations are increasingly used to check and verify architectural ideas. Yet these applications are completely detached from the activity of designing. Due to inadequate interfaces and inadequate software concepts, the workflow between physical models, analog sketches and digital tools is characterized by media disruptions. The aim of this work is to bridge the current discrepancy between established working methods and digital design support tools. This work is based on an analytical examination of the design process. On this basis, it is necessary to define the interaction methods necessary for designing and the basic requirements for design tools. Furthermore, the topic requires a consideration of the framework conditions of both actors: human and computer. This work focuses on defining an application concept for a computer-aided design system. Based on the resulting requirements, a corresponding system structure and solution approach is described. In addition, a prototypical implementation of relevant sub-areas is carried out. The goal here is not to replace the architect’s established working methods with digital methods. Rather, both worlds must be connected in such a way that their strengths merge with each other. The core idea is therefore to create a seamless coupling of established tools and digital design support tools such as analyses and simulations. In this way, design decision support is made possible at an early stage of the design process. The computer assists the architect without disturbing or overwhelming them in their creative work. This work is the English translation of the 2014 dissertation with the original title “Interaktionsformen für das digitale Entwerfen - Konzeption und Umsetzung einer rechnergestützten Entwurfsplattform für die städtebaulichen Phasen in der Architektur”
Conference Paper
Full-text available
A seamless integration of model analysis and simulations into the design process is a key for supporting the different decisions, including deciding upon the position, dimensions, and materiality of building elements. Such design options are explored from the early design phases, where a decision is taken based on their performance. A crucial analysis that is necessary for the different types of buildings, especially transportation hubs, is pedestrian flow dynamics, as it evaluates the occupants' comfort and ability to evacuating the building in case of emergency. Currently, analysing pedestrians' flow is decoupled from the BIM-authoring tools, requires multiple manual steps, and is time consuming. Hence, this paper proposes a framework that leverages the latest advancements of Deep Learning (DL) for replacing pedestrian dynamics simulations by an DL model providing intermediate feedback. In more detail, a representation of the building model, including simulation parameters, is proposed as input and a Convolutional Neural Network (CNN) architecture is developed and trained to predict pedestrians' flow density heatmaps and tracing maps.
Conference Paper
Full-text available
Contemporary agent-based pedestrian simulations offer great potential to evaluate architectural and urban design proposals in terms of medical risks, crowd safety, and visitor comfort. Nevertheless, due to their relative computational heaviness and complicated input-parameters, pedestrian simulations are not employed during the design process commonly. Simulation results significantly impact planning decisions, especially when they are already available in the early design phases. This paper analyzes the requirements of pedestrian simulations for early planning stages, such as seamless integration into iterative and collaborative design processes, interactivity, and appropriate visualization of results. For this purpose, we combine two existing projects: a high-accuracy pedestrian simulation and the CDP//Collaborative Design Platform. To adapt the simulation method to the requirements of early planning stages, we investigate interactions that blend intuitively with the design process and enable multiple users to interact simultaneously. We simplify simulations’ input parameters to match the level of detail of the early design phases. The simulation model is adapted to facilitate continuous and spontaneous interactions. Furthermore, we develop visualization techniques to support initial design negotiations and present strategies for compensating computation time and giving constant feedback to a dynamic design process.
Conference Paper
Full-text available
The integration of energy and indoor comfort aspects into the early stage of urban planning is essential to conceive energy efficient urban structures and to avoid expensive compensation measures in the later building design. This could end up with increased insulation levels, complex air conditioning systems as well as increased energy demand during the building usage or inefficient energy supply concepts. Besides indoor comfort and the energy demand for heating, cooling, ventilation and daylight, energy supply options based on district heating (or cooling), have to be considered in an early stage of design as they may constitute highly efficient alternatives with respect to building specific heating units. Incidentally, this sub-project is part of the project titled Collaborative Design Platform (CDP) which is a computer based real time tool for urban development in early stages of design decision making. Referring to a subproject intending to integrate different energy aspects into the CDP, the paper concentrates on district heating networks (DHN). The objective of this study is to evaluate the possibilities of implementing a new plug-in in the early stage of urban development in order to design and evaluate the adequacy of district heating networks.
Article
The design of accessible walking routes needs to take into account the different stakeholders' preferences and factors affecting walking. It is a complex issue which policy-makers should deal with to foster sustainable mobility. A participatory multicriteria decision analysis approach is presented to help the planning and designing of pedestrian routes, based on a sound analysis of factors affecting walking behavior and the attributes of the roads, and a stakeholder-driven evaluation of the same. A group of different stakeholders has been involved to select the criteria for designing pedestrian routes in the city center of Cartagena (Colombia). Some of them have been selected based on the results of a social network analysis (SNA) to be involved as key stakeholders for the evaluation of the selected criteria through an analytic network process (ANP). An index to measure the importance of each criterion in designing pedestrian routes has been obtained. Results provide valuable inputs to understand how to redesign and reconfigure streets for pedestrians in a city so as to improve walkability and foster a shift toward active and sustainable transport modes.
Book
The 6th International Conference on Pedestrian and Evacuation Dynamics conference (PED2012) showcased research on human locomotion. This book presents the proceedings of PED2012. Humans have walked for eons; our drive to settle the globe began with a walk out of Africa. However, much remains to discover. As the world moves toward sustainability while racing to assess and accommodate climate change, research must provide insight on the physical requirements of walking, the dynamics of pedestrians on the move and more. We must understand, predict and simulate pedestrian behaviour, to avoid dangerous situations, to plan for emergencies, and not least, to make walking more attractive and enjoyable. PED2012 offered 70 presentations and keynotes and 70 poster presentations covering new and improved mathematical models, describing new insights on pedestrian behaviour in normal and emergency cases and presenting research based on sensors and advanced observation methods. These papers offer a starting point for innovative new research, building a strong foundation for the next conference and for future research.
Thesis
Social interactions and collective dynamics are ubiquitous in our lives and are increasingly being studied in computational science and engineering. Pedestrian dynamics focuses on the interactions and movements of humans on foot in a great variety of contexts, including public transportation systems and mass events. Computer simulation can serve both for studying pedestrian behaviour and as part of an information system -- the former focusing on scientific scrutiny and the latter on engineering. Due to the complexity of human behaviour, building a scientifically credible computer simulation is a great challenge for mathematical modelling and software engineering. Known approaches mostly focus on selected phenomena and do not incorporate findings from other disciplines, such as crowd psychology and biomechanics. Perhaps because of this, they lack a plausible representation of the decision making and physical process. At first, I present a software framework that facilitates the development of new simulation approaches for pedestrian dynamics. I then discuss known approaches, and introduce the optimal steps model, which represents decision making through utility optimisation and locomotion as a discrete stepping process. The approach overcomes the limitations of the grid in cellular automata and is the first model of pedestrian dynamics representing the natural stepping process. However, according to findings in behavioural sciences, humans make decisions with heuristic reasoning, not mathematical optimisation. Therefore, I use a representation in separate layers for physics, psychology, and social behaviour in the second part of this work and propose dedicated simulation models drawing on findings from the respective fields. This approach constitutes a shift towards psychological and physical models that not only predict pedestrian behaviour but also serve as an explanation of the underlying processes. Specifically, I demonstrate how simple heuristics can be used to reproduce crowd phenomena and allow for new insights into pedestrian behaviour. The new concept facilitates interdisciplinary exchange and provides a basis for research in many directions, including psychology and biology.