spatial-temporal patterns and pedestrian simulation
TRANSCRIPT
Spatial-temporal patterns and pedestriansimulation
By Nan Hu*, Suiping Zhou, Zhongke Wu, Mingquan Zhouand Benjamin Eng Keong Cho* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
In this paper, we propose a framework for modeling lower-level pedestrian navigational
behaviors. We aim not only to generate realistic simulation results but also to make
our framework flexible and extendible, and easy to use for model developers. A
divide-and-conquer methodology is first adopted to divide the complex navigational
behaviors into three levels, which allows us to focus on the intermediate level. We then
propose a pattern-based framework for modeling pedestrian navigational behavior at this
level. In our framework, spatial-temporal patterns are used to represent the situational
perception, and a pattern-matching mechanism is proposed to model the navigational
choices of individual pedestrians. To demonstrate the effectiveness of our framework, a
computational model is constructed to simulate pedestrian behaviors in a corridor with
medium to relatively high density of pedestrians. Simulation results with this model are
summarized in this paper. Copyright # 2010 John Wiley & Sons, Ltd.
KEY WORDS: navigational behavior, pedestrian simulation, spatial-temporal patterns
Introduction
Modeling human-like pedestrian navigational beha-
viors in a virtual environment is still a challenging task
to date. In real life, various cognitive and navigational
activities are involved for pedestrian to reach their
destinations in a complex dynamic environment with
various static obstacles and moving objects. From a
computational model point of view, these activities can
be broadly classified into two levels. At the higher level,
path planning is involved to plan a rough path for a
pedestrian to move through the environment from a
starting location to a destination. At the lower level,
maneuvers are carried out by the pedestrian to navigate
along the path while keeping away from potential
collisions with other objects or pedestrians in the
environment.
The higher-level navigational behavior simulation has
been generally well studied 1–3. On the other hand,
lower-level pedestrian navigational behaviors involve
dynamic interaction between pedestrians and the
environment and among themselves in relatively short
distances. It represents a more challenging, yet inter-
esting issue in navigational behavior modeling and
simulation. Some steering mechanisms 4–7 have been
studied and applied in different models and seemingly
realistic steering behaviors have been generated.
However, a salient feature of realistic steering behaviors
can hardly be reflected by existing approaches: ped-
estrians can effectively steer even in a relatively dense
crowd without much dedicated effort. This is an
important feature for a computational model to achieve
high degree of realism that can be observed in real life
situations.
This paper presents a novel pattern-based framework
to address this challenge in modeling lower-level
pedestrian navigational behaviors. In this framework,
an agent’s perceived information and subjective judg-
ment on the situation are summarized as spatial-temporal
patterns. An agent will match the perceived patterns
with some prototypical cases in its experience. That is,
an agent’s decision-making process is essentially treated
as a pattern-matching process.
COMPUTER ANIMATION AND VIRTUAL WORLDS
Comp. Anim. Virtual Worlds 2010; 21: 387–399Published online 24 May 2010 in Wiley InterScience
(www.interscience.wiley.com) DOI: 10.1002/cav.341* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
*Correspondence to: N. Hu, Parallel and Distributed Comput-ing Center, School of Computer Engineering, Nanyang Tech-nological University, Singapore.E-mail: [email protected]
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Copyright # 2010 John Wiley & Sons, Ltd.
This pattern-based framework provides good poten-
tial to reflect the salient feature of pedestrian steering
behaviors in two aspects:
� First, the use of patterns in modeling pedestrians’
perception and judgment on the situation reflects
the fact that pedestrians usually process information
in a bulk/parallel manner. It has endowed our frame-
workwith the human-like capability of effective situa-
tional assessment.
� The pattern-matching mechanism imitates the effi-
cient decision-making process of human beings based
on their experience 8,9. Note that different expertise
levels result in different efficiency in pedestrians’
steering behaviors. For example, most adults are
experienced walkers in normal situations. They keep
sufficient patterns to help them perceive, judge the
situation rapidly, and avoid the potential collision
efficiently based on corresponding empirical instruc-
tions. Thus, they are observed to move smoothly even
in crowded places. On the contrary, children, who
have acquired few patterns and corresponding
empirical instructions in their experiences, must often
resort to some reactive actions to resolve collisions in a
close distance. Thus, their movements in crowded
places often look less smooth compared to adults
and are more likely to bump into others.
This paper is organized as follows: Section 2 describes
some related work on pedestrian modeling and simu-
lation. Our modeling methodology and the proposed
pattern-based modeling framework for pedestrian simu-
lation are presented in Section 3. Based on this framework,
wedescribe a computationalmodel to simulate pedestrian
behaviors in a corridor scenariowithmedium to relatively
high density of pedestrians. Test cases and simulation
results are shown in Section 5. Section 6 concludes this
paper and also provides ideas for future work.
RelatedWork
There have been many attempts to simulate the lower-
level navigational behaviors of pedestrians, where
collision avoidance plays a central role. In general,
existingwork can be broadly classified into three classes:
force-based, rule-based, and example-based models.
Force-based approaches essentially treat an individ-
ual as a physical entity and impose some artificial forces
on the entity to control its motion. A typical example of
this approach is Helbing’s social force model 10, where
some artificial forces have been introduced to constrain
the movement of an individual in a similar way as
Newtonian mechanics. Although plausible simulation
results have been observed by such model, we feel that
the treatment of an individual as a physical entity is
oversimplified. In particular, these social forces are not
adequate to reflect the richness in pedestrian interaction.
In agent-based pedestrian simulations, various forms of
potential field have been used to reflect agents’ personal
preferences and decision-making in their steering
behaviors 4,11–13. In these works, collision avoidance is
generally achieved through a series of comparisons
between an agent and each potential threat. Spatial
information is processed in a separate and sequential
manner, while real pedestrians seem to process spatial
information in a bulk and parallel manner.
Rule-based models achieve collision avoidance based
on certain pre-defined rules 3,5,14. In References 15–17, a
layered or multi-resolution approach is used to model
an individual’s steering behaviors in a crowd based on
certain collision avoidance rules. In general, these rules
are tightly coupled with conditions that have been
foreseen, thus the models do not react well to
exceptional situations that inevitably arise in a high-
density crowd simulation. Furthermore, it is difficult for
users to specify such rules for different situations.
To overcome the downside of the rule-based approach,
example-based approach has been proposed and applied
in crowd simulation recently 18,19. Real-life examples of
the moving trajectories of individuals within a crowd are
recorded, extracted, and stored. Agents in the simulation
analyze the current situation and compare with the stored
examples and apply the moving trajectories of the
matching example. Though seemingly natural behaviors
can be generated, such approach synthesizes the paths
for the agents. Only external factors such as the
temporal-spatial information that influence the naviga-
tional behaviors can be extracted from the image-based
examples. Unique traits of individual persons are not
easy to be modeled with such approach.
Our work aims to imitate how real pedestrians make
steering decisions corresponding to the situations in real
life, and to generate plausible navigational behaviors
naturally. Geometricmodels such as 20,21 make use of the
space–time concept and discretize the surrounding
environment into regions. In our pattern-based
approach, spatial-temporal information is processed in
a bulk/parallel manner through the recognition of
spatial-temporal patterns from the situation. Behavior
diversity is naturally and easily reflected in the different
prototypical patterns that are stored as an individual’s
experience cases.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Copyright # 2010 John Wiley & Sons, Ltd. 388 Comp. Anim. Virtual Worlds 2010; 21: 387–399
DOI: 10.1002/cav
N. HU ET AL.* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Methodologies andModelingApproaches
We believe that realistic lower-level navigational behaviors
can be generated by imitating how pedestrians make
navigational decisions in various real life situations and
modeling how they execute such decisions in loco-
motion. We first make two assumptions to describe and
explain the lower-level navigational behaviors, and then
investigate and model pedestrians’ decision-making
processes based on these assumptions.
Two Assumptions
Assumption 1: Experienced pedestrians proactively
apply certain simple steering strategies to minimize
the chances of performing instinctive reactions to avoid
imminent collisions.
It is amazing that most pedestrians can avoid collisions
efficientlywithout dedicated effort even in crowded areas.
Various steering behaviors such as following, overtaking can
be frequently observed in different situations. One of the
distinctive characteristics of these behaviors is that
proactive planning is involved. We refer to such a plan
of actions to avoid collision as a steering strategy. In
addition to such planned strategies, we also observe (less
frequently) some instinctive reactions such as standing still
suddenly or sliding body significantly that pedestrians resort
to in order to avoid imminent collisions. We do not
observe such instinctive reactions frequently for experi-
enced adults as they are equipped with various steering
strategies and tend to proactively apply these steering
strategies to prevent potential collisions in advance, thus
to avoid close contact with other pedestrians 22–24. While
for young kids with less experience of steering, they may
have to resort to instinctive reactions to avoid imminent
collisions, which also make their movement less efficient.
This assumption has two implications for pedestrian
modeling and simulation:
� First, we may divide the lower-level navigational
behavioral model into some (high-level) steering strat-
egies and (low-level) instinctive reactions.
� Second, instinctive reactions are relatively simple and
largely similar in various situations. In addition, with
good steering strategies, instinctive reactions seldom
occur. Therefore, our design focuses on steering strategies.
Assumption 2: A limited number of steering strategies
are enough for pedestrians to move smoothly in real-life
situations through scheduling and executing these
steering strategies properly based on their experience.
Though pedestrians perform many different
steering behaviors under different situations, a limited
number of steering strategies are commonly observed
in daily life such as following, overtaking, and side-
avoiding 22–24. A person’s steering strategies are usually
acquired from his/her past experience. In terms of
steering behaviors, experience influences pedestrians
from two aspects. Firstly, individual pedestrians
with different steering experiences may store different
sets of steering strategies in their long-term memory.
Secondly, even with the same set of steering strategies,
different experiences may cause pedestrians to select
and execute steering strategies in different ways. This
suggests that even equipped with the same limited
number of steering strategies, different pedestrians may
make different decisions on when to use which steering
strategies, thus they may demonstrate different steering
behaviors.
This assumption also has two implications on the
pedestrian modeling and simulation:
� First, it suggests a divide-and-conquer methodology
such that we may investigate limited sets of steering
strategies to generate complex behaviors in different
situations.
� Second, it suggests a modeling approach that empha-
sizes the role of experience in pedestrian decision
making. In this regard, pattern-based approach is a
good attempt, which will be discussed later.
Divide-and-ConquerMethodology
Based on the two assumptions, a divide-and-conquer
methodology is proposed for the design of ourmodeling
framework. As shown in Figure 1, the complex lower-
level pedestrian navigational behaviors are composed of
the default behaviors and the maneuvers for collision
avoidance. Maneuvers can be further divided into two
parts as steering behaviors and instinctive reactions and
our framework will focus on modeling the steering
behaviors due to Assumption 1. Based onAssumption 2,
steering behaviors consist of a limited number of
steering strategies from a strategy pool. Pedestrians
select strategies from the strategy pool and schedule
them in a proper sequence to achieve collision
avoidance. Each selected steering strategy is executed
through a sequence of locomotive actions (e.g., move,
stop, turn, etc.).
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Copyright # 2010 John Wiley & Sons, Ltd. 389 Comp. Anim. Virtual Worlds 2010; 21: 387–399
DOI: 10.1002/cav
SPATIAL-TEMPORAL PATTERNS* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
As shown in Figure 1, complex navigational behaviors
are divided into three levels, which allow a modeler to
focus on different levels in the hierarchy to meet
different design requirements. The divide-and-conquer
methodology has two main advantages:
� First, it naturally reflects the way real pedestrians
make decisions during navigation, e.g., pedestrians
will not consider how to execute a steering strategy
until it selects it.
� Second, it simplifies the modeling of complex naviga-
tional behavior. On one hand, it allows model devel-
opers to focus on behaviors with less complexity at a
time in top-down manner. On the other, it avoids the
complicated process of fine tuning of various influ-
ential factors that directly affect the locomotion of the
agent. Note that navigational behaviors are often and
naturally described in some higher levels than the
locomotive level in daily life. For example, we say
‘‘A is overtaking B’’ instead of ‘‘A increases his speed by
how many units and turns his direction by how much angle
to the left/right side of B.’’ In our framework, steering
behaviors lie in the intermediate level, which balances
the mentioned modeling problems of the higher and
lower level behaviors.
Pattern-based Framework
Figure 2 shows our modeling framework which follows
the divide-and-conquer methodology. We address two
important phases in the cognitive process that are
involved in selecting and executing steering strategies:
formation of situational awareness and pattern matching.
Pedestrians need to form up situation awareness about
the current situation. According to Endsley 25, there are
three levels of situation awareness: a person’s perception
of environmental elements within a volume of time and
space, the comprehension of their meaning, and the
projection of their status in the near future. In our
framework, agents continuously assess the environment
during navigation through sensory inputs (e.g., vision,
audition, etc.). Based on the sensed information,
comprehension and prediction aremade on the situation
for the current moment t as well as for a short period of
time Tp in the near future.
In the framework, we have proposed to use spatial-
temporal patterns to represent some prototypical spatial-
temporal configurations that pedestrians rely on to
quickly recognize the current situation. Figure 3 illus-
trates the pattern-matching mechanism that we pro-
posed to model the cognitive processes of a pedestrian.
Figure 1. Divide-and-conquer methodology.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Copyright # 2010 John Wiley & Sons, Ltd. 390 Comp. Anim. Virtual Worlds 2010; 21: 387–399
DOI: 10.1002/cav
N. HU ET AL.* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
We first describe the steering strategy selection process.
From the situation awareness module, an agent will
recognize certain global spatial pattern as the perceptive
description and overall judgment on the current
situation. Together with its intention1, the agent tries
to match the global spatial pattern with certain
prototypical patterns (X1–Xn) from its experience
instances (E1–En). The matched prototypical pattern Xi
serves as the cue to retrieve the associated empirical
steering strategy Si. In case no matched pattern is found,
the agent needs to make decisions through a trial-and-
error process. Such case usually happens for a child,
who lacks the experience in recognizing the global
spatial patterns. Nevertheless, the match does not need
to be perfect as we will discuss in the next section.
Once a steering strategy Si is selected, the associated
empirical instructions Yi on how to execute steering
strategy Si will be retrieved. We model the empirical
instructions Yi in the form of a set of qualitative rules
(Ri1–Rik). These rules serve as the benchmarks to instruct
the agent to execute the specific steering strategy Si in a
particular way based on certain prototypical conditions.
Such conditions of the rules are usually kept in the
experience cases of the agent as certain patterns (P1–Pk)
about the situation. For example, one of the rules in Yi
could be ‘‘If there is enough available space on the right hand
side of the target, then overtake the target from the right hand
side.’’ The condition of this rule can be perceived as a
prototypical spatial pattern that reflects the particular
spatial configuration of the situation. The agent will then
match the various local patterns in the current situation
with the prototypical patterns. Based on the match, the
resultant execution Ak will be applied to the current
situation.
Implementation of aComputational Model
Based on the proposed framework, we have imple-
mented a computational model for simulating ped-
estrian steering behaviors in a medium to relatively
high-density scenario. We choose this scenario since
higher density in general will create higher opportunity
of potential collisions in the simulation, which may
involve various interesting steering behaviors. Such
scenario also represents many real-life situations such as
a corridor or a passway at a city link. The steering
strategy pool contains the three commonly observed
strategies that pedestrians usually adopt: following,
overtaking, and side-avoiding. We will focus on the
development of the key components of this compu-
tational model such as situation awareness and pattern-
matching process in this section.
Situational Awareness
Wemodel the three levels of situation awareness (SA) of
an agent.
Level1SA: perception. Currently, we only consider
the vision inputs as pedestrians mainly reply on visual
spatial information during navigation. Figure 4 illus-
trates the vision and attention system in the model.
In our current design, the vision of an agent is
constraint to the vision angle of 1208 and vision range of
around 5m in a relatively dense crowd, which is a
reasonable estimation with sufficient simplicity. How-
ever, vision angle can be further extended as the agent
actively perceives the environment, where head and eye
rotation is involved. Our model can easily adapt to
model such feature by includingmore columns to record
Figure 2. Modeling framework.
1The intention of an agent is generated from the higher-levelcognitive process and it indicates the steering strategy that theagent prefers to apply in the current situation.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Copyright # 2010 John Wiley & Sons, Ltd. 391 Comp. Anim. Virtual Worlds 2010; 21: 387–399
DOI: 10.1002/cav
SPATIAL-TEMPORAL PATTERNS* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
the spatial information as will be discussed later. Vision
center indicates the moving direction of the agent. The
fan-shape originated from the agent and enclosed by the
vision range indicates the spatial information available
to the agent at each simulation frame. Information
collectedwith theAttention range of an agentwill be used
by the agent for decision making.
Level 2 SA: comprehension. To imitate the continu-
ous spatial information perceived by pedestrians, we
represent agents’ perception on the current spatial
information through radial intersection scanning. The
space within the vision range as shown in Figure 4 is
partitioned into 12 equal sectors by a total of 13 radials
including two boundaries eradiated from the agent. An
integer array is then used to represent the processed
spatial information within the attention range based on
the intersection status of the 13 radials. The value of this
array is designed as:
0: if the radial does not hit any agent
1: if the radial hits an agent with the same moving
direction
�1: if the radial hits an agent with the opposite
moving direction
If more than one agent is hit by the same radial within
the attention range, the closer one to the observing agent
will be considered. The processed spatial information as
shown in the Figure 4 is thus represented as:
SAðt; 0Þ¼ ½000� 1011100000�
From this array, we may easily infer that ‘‘there is
available space in my right front direction.’’ We will discuss
on how to form patterns from this spatial representation
in more detail in the next subsection.
Level 3 SA: projection. In addition to the spatial
information obtained from the Level 2 SA, agents further
assess the situation through mental prediction. In our
design, prediction is based on the current status (e.g.,Figure 4. Agents’ vision and attention system.
Figure 3. Pattern-matching mechanism.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Copyright # 2010 John Wiley & Sons, Ltd. 392 Comp. Anim. Virtual Worlds 2010; 21: 387–399
DOI: 10.1002/cav
N. HU ET AL.* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
relative speed, relative direction, etc.) of the perceived
information. The spatial information in the predicted
frames will be extrapolated. Note that only the spatial
information of other agents within the vision range at
the current framewill be predicted to represent the basic
predictive ability of human in real life. For example, a
pedestrian usually cannot predict the information of
others behind a wall that he/she cannot see.
The 1D integer array in Level 2 SA can be extended to
a 2D array by taking the spatial information in the
predicted frames into account. For example, the Level 3
SA with three predicted frames is represented as:
SAðt; 3Þ ¼0 0 0 �1 0 1 1 1 0 0 0 0 0�1 �1 �1 0 0 1 1 1 0 0 0 0 0�1 �1 0 0 1 1 1 1 1 0 0 0 0�1 0 0 0 1 1 1 1 1 0 0 0 0
2664
3775
Such 2D representation is not only capable of
representing the continuous spatial information in a
discrete form, but also capturing the dynamic change of
spatial information in the temporal domain automati-
cally. Furthermore, the array data structure provides
good feasibility for formulating the pattern-matching
process in computer program as will be discussed.
Though the static obstacles are not included in the
example as demonstrated in Figure 4, it is worth
mentioning here that we handle the static obstacles in
the formation of the three levels of SA in the similar way
as we process the dynamic obstacles (e.g., other agents).
The only difference lies in the representation of the
spatial information for the perceived static obstacles
within the attention range of the observing agent. We
may assign a distinctive value (e.g., 2) to represent that a
direction is blocked by static obstacles.
Pattern-matching Processes
Steering Strategy Selection. We have proposed a
simple steering strategy selection scheme based on the
agent’s general situational judgment and intention.
This scheme is based on a fuzzy pattern-matching
mechanism.
The general situational judgment usually involves the
immediate recognition of certain global spatial pattern at
the current time instance without further prediction.
Thus, only the first row of 2D representation of the
situation awareness is used in this process. Such global
spatial pattern is characterized by the spatial features
such as themoving direction aswell as the relative positions
of other agents to the observing agent. The value sets for
these spatial features are formulated into fuzzy sets as:
A¼ {same, mostly same, mostly opposite, opposite}
B¼ {front is fully blocked, front center is blocked with
available space on either side, front center is not blocked}
The value set for intention of the agent is defined as a
crisp set as:
C¼ {overtaking, following, side-avoiding}
The value set of the result of the selection is defined as
another crisp set as:
Y¼ {overtaking, following, side-avoiding, instinctive reac-
tions}
The relation between these variables can be described
using some fuzzy rules in the form of Y¼ f (A, B, C). The
fuzzy rules reflect different decisions on the strategy
selection of different pedestrians according to their
different experiences. For instance, one of such rules
corresponding to the example shown in Figure 4 is ‘‘if the
front center is blocked by agents with same moving direction
and there is available space on either side, and the preferred
strategy is overtaking, then the strategy overtaking is
selected.’’ By aggregating all of such rules, a check-up
table can be built to imitate the experience instances of
pedestrians.
Steering Strategy Execution. Once a specific steer-
ing strategy is selected, the agent needs to properly
execute the selected strategy. Issues such as ‘‘how to
choose an overtaking (or following) target’’ and ‘‘how to
overtake (or follow) the target’’ need to be resolved to
guide the agent’s locomotive actions. In this subsection,
we discuss the pattern recognition, pattern matching,
and instruction retrieval processes that have been
implemented in the model.
Local Pattern Recognition. As shown in Figure 5, we
consider three spatial features of the spatial configur-
ation: relative distance (sd), relative direction (sa), and the
spatial span (sw, e.g., the spatial width perpendicular to
the observing agent’s moving direction).
Based on these features, a spatial pattern can be
specified ps (sd, sa, sw). Note that the parameters (e.g., d1,
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Copyright # 2010 John Wiley & Sons, Ltd. 393 Comp. Anim. Virtual Worlds 2010; 21: 387–399
DOI: 10.1002/cav
SPATIAL-TEMPORAL PATTERNS* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
d2, u1, u2, w) are for visual illustration purpose only. In
actual implementation, we make use of the situation
awareness in the form of a 2D integer array based on the
following value assigning scheme:
� sd is set according to the row index of a particular row
in the array, where spatial information grouping are
performed. Note that an agent only processes spatial
information (e.g., set to 0, 1, or �1 etc.) within its
attention range when forming its situation awareness.
Thus, relative distance can be measured in terms of
the number of predicted frames. In our design, the
particular frame is identified as the first row in the 2D
array that contains the least number of 0s with the
seventh (e.g., vision center) column is not a 0. In this
row, the available space for the observing agent is
most limited. Thus, there is least number of choices
available and the least amount of time for the agent to
execute the steering strategy according to the com-
prehension and prediction on the situation at the
current time instance.
� sa is set according to the direction deviation from the
observing agent’s vision center. It can be calculated
easily as the differences between the column index of
the group of spatial information to the interest of the
observing agent and 7 (e.g., column index of the vision
center).
� sw is set based on the continuity of the group of spatial
information at the particular frame as discussed above
in setting sd. Specifically, the value equals to the
number of continuous 1 or �1 or 0 at the particular
row of the 2D array.
PatternMatching. In general, pattern matching is the
act of checking for the presence of the constituents of a
given pattern. Thus, pattern-matching process can be
realized through a series of comparisons between paired
features that are used to define the pattern. We calculate
the similarity value for each pair of the features as
g j ¼ Gjðfj; fej Þ, where Gjð�Þ represents the function to
measure the similarity value between the pair of features
of type j. fj and fej represent the features of type j that are
used in the definition of the recognized pattern in the
current situation and the prototypical pattern stored in
the experience of the agent, respectively.
The similarity between the two patterns can then be
obtained by aggregating the similarity values for all the
features as g ¼ Pwj � g j, wherewj represents the relative
importance of different features.
Empirical Instruction Retrieval. Several recognized
local patterns may be matched with the same proto-
typical patterns with different similarity values. Our
empirical instruction retrieval mechanism is not solely
based on the pattern similarity value, but also considers
the adaptation effort to apply the instruction in the
current situation. The adaptation effort is measured as
the aggregated change in moving speed and direction.
Some pedestrian studies 22,26 provide guideline for the
specific settings in the aggregation. By considering the
adaptation effort, not all the local patterns are necess-
arily compared with the prototypical patterns in the
empirical instruction for the specific steering strategy
execution. This makes our model more efficient in terms
of runtime performance.
Simulation Results
We have implemented the aforementioned compu-
tational model in liteC provided by GameStudio A7.
Figure 6 shows some different views of a testing scenario.
We have tested themodel by tracking the navigational
behaviors of some individual agents in some typical
testing cases and analyzing their steering choices in such
cases together with their perceived spatial-temporal
patterns. In particular, we examine the execution
processes of individual steering strategies (following,
overtaking, and side-avoiding) as well as the autonomous
scheduling of these strategies.
Figure 7 shows a typical test case for our model. The
visionary radials are drawn in black lines. The personal
space of the observing agent as well as those who have
entered into the attention range of the observing agent is
shown as a cylinder. We specify the individual
parameters (e.g., personal space, number of predicted
frames, etc.) and the rules relating prototypical cases and
Figure 5. Spatial pattern.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Copyright # 2010 John Wiley & Sons, Ltd. 394 Comp. Anim. Virtual Worlds 2010; 21: 387–399
DOI: 10.1002/cav
N. HU ET AL.* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
corresponding instructions based on some existing
pedestrian studies and our own experiences as ped-
estrians in real life. For example, a typical rule for
overtaking that we have implemented is ‘‘if there is
available space on the right hand side, then overtake from right
hand side.’’ It should be pointed out that these parameters
and rules can be easily changed according to different
requirements based on different set of experiences.
In the testing case shown in Figure 7, we set the
intention of the observing agent as ‘‘overtaking,’’ which
indicates that the agent will always attempt to overtake
others if the situation allows it to do so. In the beginning
(see Figure 7a), it notices the group of two agents in front
and the wall on the left-hand side blocks its way. In
Figure 7b, it decides to overtake from the right-hand side
of the group in front. Later in Figure 7c, its newly
perceived spatial-temporal pattern indicates that there is
no available space in front for the next few frames. Note
that in Figure 7c, the agents on the right front of the
observing agent are out of the attention range of the
observing agent at the current frame. However, since
prediction is involved in the proposed pattern recog-
nition process, the agent can actually perceives the
situation as ‘‘no available space for overtaking’’ and will
change its steering strategy from overtaking to following
temporarily. The agent executes the following steering
strategy in Figure 7d and e. In Figure 7e, the newly
perceived spatial-temporal pattern indicates that there
will be available space on the left-hand side since the
agent occupying the space currently is overtaking others
and will make the space available. Thus, the overtaking
steering strategy is selected by the agent. Figure 7f and g
illustrates the overtaking process.
In Figure 8, we have tested cases with agents moving
in different directions. We have observed that agents
avoid each other in a cooperative manner, and
temporary lane-formation phenomenon can also be
observed during this process. Due to page limit, we will
Figure 6. Different views of simulation cases.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Copyright # 2010 John Wiley & Sons, Ltd. 395 Comp. Anim. Virtual Worlds 2010; 21: 387–399
DOI: 10.1002/cav
SPATIAL-TEMPORAL PATTERNS* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
not explain this test case in detail. More testing results
can be found from the supplementary videos.
Though a quantitative validation of the model is still
in progress, the current simulation results are promising
in the way that:
� Our model is capable of replicating some interesting
pedestrian behaviors in real life situations.
� By integrating our model with some higher level
cognitive layers, macroscopic phenomena such as
lane-formation and leader-following could be
achieved.
� The model naturally follows the way pedestrians
make decisions based on their perceived spatial-
temporal patterns and experiences. The model makes
it easier for developers to specify the rules that reflect
individual experiences to cater for different appli-
cations.
Certain limitations of our current implementation can
also be found from these test cases. First, since single
thread programming is used currently, all the agents
make decisions at the same frame, which may some-
times cause unrealistic decisions. Second, the evenly
distributed visionary radials may cause difficulties for
pattern specifications to describe groups of agents at the
peripheral area within the attention range of the
observing agent. Thirdly, the static 1208 of vision angle
may cause an agent to miss some spatial information
that could be detected by real pedestrians. As discussed,
this problem can be partially solved by expanding the
number of columns in the 2D array form of situation
Figure 7. Automatic steering strategy scheduling between overtaking and following strategies.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Copyright # 2010 John Wiley & Sons, Ltd. 396 Comp. Anim. Virtual Worlds 2010; 21: 387–399
DOI: 10.1002/cav
N. HU ET AL.* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
awareness. Research such as Reference 24 that focuses on
modeling the gaze attention of the agents could provide
idea for further improvement on our model as well.
Conclusion and FutureWork
We present a pattern-based framework that aims to
imitate how real pedestrians perceive andmake steering
decisions in daily-life situations. This new framework
has the potential to improve the realism of pedestrian
simulation. As the spatial-temporal information percep-
tion and reasoning processes of this pattern-based
approach are more natural and closer to how we make
steering decisions in daily-life, the framework makes it
easier for model developers to specify various
parameters and behavioral rules in simulating ped-
estrian behaviors in different situations. With this
pattern-based approach, the complex cognitive pro-
cesses involved in steering decision making are
essentially transformed to pattern-matching processes
between the perceived spatial-temporal patterns and the
prototypical cases in agents’ experience.
The current simulation results are promising. We plan
to further refine some of the key components in our
framework, especially the situation awareness and the
representation of spatial-temporal patterns. Quantitat-
ive validation of the model will also be conducted.
Nevertheless, we hope that the idea presented in this
paper can direct researchers in this area with a fresh
perspective.
ACKNOWLEDGEMENTS
This work is supported in part by the Singapore National
Research Foundation under Grant NRF2007IDM-IDM002-052.
References
1. Paris S, Bonvalet N, Donikian S. Environmental abstractionand path planning techniques for realistic crowd simu-lation. Computer Animation and Virtual Worlds 2006; 17:325–335.
2. Pettre J, Ciechomski PdH, Maım J, Yersin B, Laumond J-P,Thalmann D. Real-time navigating crowds: scalable simu-
Figure 8. Lane formation emerged from following and side-avoiding strategies.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Copyright # 2010 John Wiley & Sons, Ltd. 397 Comp. Anim. Virtual Worlds 2010; 21: 387–399
DOI: 10.1002/cav
SPATIAL-TEMPORAL PATTERNS* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
lation and rendering.Computer Animation and VirtualWorlds2006; 17(3–4): 445–455.
3. Shao W, Terzopoulos D. Autonomous pedestrians. Proceed-ings of the 2005 ACM SIGGRAPH/Eurographics Symposium onComputer Animation. ACM: Los Angeles, California, 2005.
4. Kapadia M, Singh S, Hewlett W, Faloutsos P. Egocentricaffordance fields in pedestrian steering. I3D ’09: Proceedingsof the 2009 Symposium on Interactive 3D Graphics and Games.ACM: Boston, Massachusetts, 2009.
5. Loscos C, Marchal D, Meyer A. Intuitive Crowd Behaviourin Dense Urban Environments using Local Laws. Proceed-ings of the Theory and Practice of Computer Graphics 2003. IEEEComputer Society: Washington, DC, 2003.
6. PelechanoN, Allbeck JM, Badler NI. Controlling individualagents in high-density crowd simulation. Proceedings of the2007 ACM SIGGRAPH/Eurographics Symposium on ComputerAnimation. Eurographics Association: San Diego, Califor-nia, 2007.
7. Pelechano N, O’Brien K, Silverman B, Badler N. Crowdsimulation incorporating agent psychological models, rolesand communication. Proceedings of the First InternationalWorkshop on Crowd Simulation, 2005.
8. Zsambok CE, Klein GA. Naturalistic Decision-making. Lawr-ence Erlbaum Associates: New Jersey, 1997.
9. Klein GA. Sources of Power: How People Make Decisions. MITPress: Cambridge, MA, 1998.
10. Helbing D, Molnar P. Social force model for pedestriandynamics. Physical Review 1995; 51(5): 4282–4286.
11. Reynolds CW. Steering behaviors for autonomous charac-ters. Proceedings of Game Developer Conference, San Fransico,CA, 1999; 763–782.
12. RonaldN, Sterling L, KirleyM. An agent-based approach tomodelling pedestrian behaviour. International Journal ofSimulation: Systems, Science and Technology 2007; 8(1): 25–39.
13. Treuille A, Cooper S, Popovi Z. Continuum crowds. ACMTransactions on Graphics 2006; 25(3): 1160–1168.
14. Lamarche F, Donikian S. Crowd of virtual humans: a newapproach for real time navigation in complex and struc-tured environments. Computer Graphics Forum 2004; 23(3):509–518.
15. Ulicny B, Thalmann D. Crowd simulation for interactivevirtual environments and VR training systems. Proceedingsof the Eurographic Workshop on Computer Animation andSimulation. Springer-Verlag New York, Inc.: Manchester,UK, 2001.
16. Ulicny B, ThalmannD. Towards interactive real-time crowdbehaviour. Computer Graphics Forum 2002; 21: 767–775.
17. Musse SR, Thalmann D. Hierarchical model for real timesimulation of virtual human crowds. IEEE Transactions onVisualization and Computer Graphics 2001; 7: 152–164.
18. Lee KH, Choi MG, Hong Q, Lee J. Group behavior fromvideo: a data-driven approach to crowd simulation. Pro-ceedings of the 2007 ACM SIGGRAPH/Eurographics Sym-posium on Computer Animation. Eurographics Association:San Diego, California, 2007; 109–118.
19. Lerner A, Chrysanthou Y, Lischinski D. Crowds byexample. Computer Graphics Forum 2007; 26(3): 655–664.
20. Feurtey F. Simulating the collision avoidance behavior ofpedestrians. School of Engineering. University of Tokyo:Tokyo, 2000.
21. Paris S, Pettre J, Donikian S. Pedestrian reactive navigationfor crowd simulation: a predictive approach. ComputerGraphics Forum 2007; 26: 665–674.
22. Dalton RC. The secret is to follow your nose: route pathselection and angularity. Environment and Behavior 2003;35(1): 107–131.
23. Goffman E. Relations in Public: Microstudies of the PublicOrder. Basic Books: New York, 1971.
24. Wolff M. Notes on the behaviour of pedestrians. In People inPlaces: The Sociology of the Familiar, Birenbaum A, Sagar E(ed.). Praeger: New York, 1973; 35–48.
25. Endsley MR. Measurement of situation awareness indynamic systems. Human Factors 1995; 37(1): 65–84.
26. Collett P, Marsh P. Patterns of public behavior: collisionavoidance on a pedestrian crossing. Non-verbal Communi-cation, Interaction and Gesture, Approaches to Semiotics, 1981;199–217.
Authors’ biographies:
Nan Hu is currently a Ph.D. candidate in the School ofComputer Engineering at Nanyang Technological Uni-versity (NTU), Singapore. He received his B.Eng. inComputer Engineering from NTU in 2007. His currentresearch interests include decision-making of ped-estrians during navigation, situation awareness,spatial-temporal pattern representation, and matchingmechanism in modeling crowds of pedestrians.
Suiping Zhou is currently an Assistant Professor in theSchool of Computer Engineering at Nanyang Techno-logical University (Singapore). Previously, he worked asan Engineer in Beijing Simulation Center, China Aero-space Corporation, and then joined Weizmann Instituteof Science (Israel) as a Post-doctoral fellow. He receivedhis B.Eng., M.Eng., and Ph.D in Electrical Engineeringfrom Beijing University of Aeronautics andAstronautics(P.R. China). He is a member of IEEE and his current
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Copyright # 2010 John Wiley & Sons, Ltd. 398 Comp. Anim. Virtual Worlds 2010; 21: 387–399
DOI: 10.1002/cav
N. HU ET AL.* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
research interests include large-scale distributed inter-active applications (e.g., MMOGs), parallel/distributedsystems, and human behavior representation in model-ing and simulation. He has published over 70 peerreviewed articles in these areas. He is currently anAssociate Editor of the International Journal of Compu-ter Games Technology. He has served as technical pro-gram committee member of many internationalconferences and workshops in computer games andvirtual environments.
Zhongke Wu is a Professor in College of InformationScience and Technology, Beijing Normal University,P.R. China. He received B.Sc. in Mathematics fromPeking University in China in 1988, and M.Eng. andPh.D. in CAD/CAM from Beijing University of Aero-nautics & Astronautics, China, in 1991 and 1995 respect-ively. His current research interests include computergraphics, geometric modeling, CAD/CAM, volumegraphics and medical imaging, scientific visualization,animation and virtual reality.
Mingquan Zhou is the Dean of College of InformationScience and Technology, Beijing Normal University,P.R. China. His current research interests include com-puter graphics, 3D visualization. He has received 11awards of province andministry. He has published over300 important papers in this field.
Cho Eng Keong Benjamin is currently an undergradu-ate student in School of Computer Engineering atNanyang Technological University.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Copyright # 2010 John Wiley & Sons, Ltd. 399 Comp. Anim. Virtual Worlds 2010; 21: 387–399
DOI: 10.1002/cav
SPATIAL-TEMPORAL PATTERNS* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *