Improving Crowd Behaviour for Games and Virtual Worlds
Saskia Groenewegen
Universiteit Utrecht
Padualaan 14
3584CH Utrecht, The Netherlands
saskia@cs.uu.nl
Figure 1: Interactive crowd in the game Assassin's
Creed 2 [1].
ABSTRACT
Interactive crowds are gaining importance in games and vir-
tual worlds, but simulating realistic crowds is a complex
task. This paper discusses how to improve the believability
of real-time crowd behaviour by enhancing path-planning
methods and introducing more behavioural diversity.
Categories and Subject Descriptors
I.6.5 [Computing Methodologies]: Simulation and Mod-
elling|Model Development ; I.3.7 [Computer Graphics]:
Three-Dimensional Graphics and Realism|Virtual Reality ;
K.8.0 [Personal Computing]: General|Games
Keywords
Crowds. Real-time. Path planning. Behaviour.
1. INTRODUCTION
When wandering around in the real world, a person is
often surrounded by other people which inhabit the same
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
FDG 2010, June 19-21, Monterey, CA, USA
Copyright 2010 ACM 978-1-60558-937-4/10/06 ...$10.00.
Figure 2: Pedestrian crowd simulation at a street
crossing [2].
area. In a virtual world, the absence of others thus creates
an unrealistic eect and can destroy the feeling of immer-
sion. Filling the world with crowds is an important step
towards more realism, in computer games and other virtual
environments.
Populating virtual environments with realistic behaving crowds
is a challenging and diverse subject, where the need to com-
pute large crowds in real time has to be balanced against
the wish to simulate complex human behaviour.
In the computer game community, games that utilize crowds
are just starting to emerge, making this eld a very inter-
esting area to work in. At the forefront of this development
is Ubisoft's Assassin's Creed 2 [1], shown on Figure 1. The
game uses crowds of up to 120 characters at a time to serve
as obstacles to the player, but also allows a limited interac-
tion with the crowd: the player can follow groups of people
to "blend in" on a crowded street and avoid being seen. But
this is "only the beginning", as the game's creative director
puts it: "we are learning how to do a crowd, how to play
with crowds, it is something you don't do in other games,
or you don't do that much" [3].
In the scientic community, much research on crowd sim-
ulation has been published in recent years. Crowds are in
particular used for simulating emergencies, but the popu-
lation of virtual environments to enhance credibility is also
a topic. One of the biggest problems in real-time simula-
tions remains the believable behaviour of individuals within
a crowd. Many crowds are only realistic if one does not
watch them too long or too closely, as techniques which save
computation costs also limit the diversity in look and be-
256

haviour of virtual characters.
Common problems with crowd behaviour, in games as well
as in research, are characters that follow unrealistic paths,
abruptly change their direction, oscillate back and forth,
walk in circles because they have no memory, walk into ob-
stacles, do not react to changes in the environment, or all
walk by themselves instead of forming groups. There is also
a notable uniformity in the looks, movement and manner-
isms of individuals, instead of the variety of behaviours that
can be observed in real crowds. My work focuses on improv-
ing the real-time behaviour of individuals in a crowd in such
a way that watching an individual for an extended period
of time will not destroy the illusion of "natural" behaviour,
and that a wide range of dierent behaviours can be ob-
served within a crowd. I plan to achieve this by developing
new algorithms for path planning and following, combined
with a high-level description of the world which in
uences
behaviour.
2. RELATED WORK
Real-time crowd simulation.
The dynamics of crowds have been studied extensively
in the past, and a variety of crowd simulation approaches
exist. They range from early social forces models and
uid
simulations [4] to cellular automata [5], rule-based systems
[6] and agent systems. A recent example is shown on Figure
2.
Path following.
Many path planning methods exist and are used in games,
like grid based A* methods, navigation meshes, and
ock-
ing. I base my work on the Indicative Route Method (IRM).
It can be used to steer characters though a complex virtual
environment in real-time [7]. It creates a path based on a
Voronoi diagram of a static environment, and an additional
corridor which denes the area that a character can walk in
while reacting to local conditions. Characters move towards
a goal following an attraction point that moves along the
path in the center of the corridor. The IRM only supports
goal-oriented behaviour, and draws characters towards the
center of a road, while people tend to walk on the sides. It
does also not deal with turning corners in a realistic manner,
nor take into account the dierent movements of people that
are familiar or unfamiliar with an environment, people wan-
dering around as tourists, people stopping or people striding
purposefully, e.g. on the way to a meeting or home. Similar
approaches with the same limitations have been proposed
by [8].
Crowd diversity.
Visual cues are the most common focus when trying to
achieve diversity in a crowd while minimizing the number
of 3D character models used: [9] have found that by chang-
ing the colour of characters' clothes, especially on the torso,
the time it takes onlookers to detect clones in a crowd is
extended. [10] have studied the eect of desynchronization
and speed variation of motions of characters on clone per-
ception, while [11] added a variety of props such as hats and
bags to make a crowd look more diverse. Those approaches
vary the look of characters, but not their behaviour.
To create complex behaviours, agent systems are often em-
ployed. In those systems, individual agents contain be-
havioural models, where dierent behaviours are triggered
by the agent's internal states. The internal state can be
the result of an emotion model or a set of rules [12]. Agent
models of sucient complexity are too computationally de-
manding to achieve real-time speed for large crowds and are
therefore not feasible for computer games. One approach
to solve the complexity problem was proposed by [13], who
move the "intelligence" from the agents into objects in the
scene, which tell approaching agents how to interact with
them, but this does not scale to inter-agent interaction.
3. RESEARCH GOALS
My goal with this research is to fully automatically create
and simulate crowds of people in real-time based on high-
level descriptions. In particular, I want to achieve the fol-
lowing sub goals:
 To create new algorithms for crowd simulation based
on path planning by expanding and enhancing exist-
ing methods like the IRM. I will focus on more realistic
behaviour when turning corners, crossing open spaces
and keeping distances from mobile and immobile ob-
stacles, with the explicit constraint of real-time frame
rates.
 To give the people in the crowd a range of varied
behaviours, from the commonly found goal-oriented
behaviour to distinctly dierent behaviours like shop-
ping, wandering aimlessly or exploring.
 To handle dynamic changes in the world. This includes
moving obstacles which completely block a path, as
well as more indistinct in
uences on path choice such
as crowd density and visibility in an area.
 To extend the variety of inter-character collision avoid-
ance behaviours related to personality, gender, age and
other properties of the characters.
To achieve those goals, I will begin by studying real-world
crowds and their behaviour using observations, tracking peo-
ple in videos, and controlled experiments in our motion cap-
ture lab. I will then derive computable models from the
observations, design and implement algorithms simulating
these models, and evaluated them in terms of quality and
eciency.
4. RELEVANCE OF RESEARCH
The simulation of large crowds of people is increasingly
gaining importance in video games, with crowds starting
to become an integral part of game-play, instead of serving
as the little observed backdrop where errors do not matter
much. As soon as a player can interact with a crowd, or in-
dividuals therein, it is imperative that the crowd members
behave in believable ways. Erratic behaviour will destroy
the immersion in a virtual world and game creators cannot
depend on suspension of disbelief regarding behaviour in an
environment that oers ever more realistic graphics. They
have to keep up with the development in other areas to cre-
ate a balance that is essential for a high-quality overall game
experience.
Another area where crowds are gaining importance are avatar-
based online 3D social networks, e.g. Second Life [14]. A
257