Вы находитесь на странице: 1из 12

Computers & Industrial Engineering 59 (2010) 736747

Contents lists available at ScienceDirect

Computers & Industrial Engineering


journal homepage: www.elsevier.com/locate/caie

Simulating operational behaviors of pedestrian navigation q


John M. Usher *, Lesley Strawderman
Department of Industrial and Systems Engineering, Mississippi State University, P.O. Box 9542, MS 39762, USA

a r t i c l e

i n f o

Article history:
Received 10 March 2009
Received in revised form 19 July 2010
Accepted 28 July 2010
Available online 3 August 2010
Keywords:
Pedestrian simulation
Pedestrian navigation behavior
Micro-simulation

a b s t r a c t
Navigation is an innate ability for humans, but simulating this capability in a virtual environment is no
easy task and has been of interest to researchers for over a decade. This paper describes the development
of ISAPT, an individual-based Intermodal Simulator for the Analysis of Pedestrian Trafc. ISAPTs development is based on the observed behaviors of pedestrians reported from the literature and simulates
the strategies employed by pedestrians for collision avoidance, including changes in speed and trajectory,
passing strategies, and distance between objects. The implementation of these behaviors and strategies is
described in the paper along with the results from a validation study. These results illustrate that the
micro-level simulation of individual pedestrians gives ISAPT the ability to reproduce identied macrolevel pedestrian behavior, as well as the capability to reproduce the operational statistics of an observed
pedestrian corridor. Such functionality is necessary to support the use of simulation as a tool for designers and planners in the design and evaluation of intermodal facilities.
2010 Elsevier Ltd. All rights reserved.

1. Introduction
Capturing realistic pedestrian behavior in simulation is useful
for evaluation and planning in building design (Daamen, Bovy, &
Hoogendoorn, 2001; Okazaki, 1979), urban design (Jiang, 1999),
land use (Parker, Manson, Janssen, Hoffmann, & Deadman, 2003),
marketing (Borgers & Timmermans, 1986), passenger movement
(Zhang, Han, & Li, 2008) and trafc operations (Cetin, Nagel, Raney,
& Voellmy, 2002). The goal of our research is to develop a system
that can be used as an aid for designers and planners in the evaluation and operation of intermodal facilities. Based on a review of
published texts on architectural design and our conversation with
architects, there appears to be no real science associated with the
design of such facilities. It appears that design decisions are made
based on experiential knowledge concerning operations within
such facilities. It is believed that the availability of a simulation
system that realistically models the behavior of pedestrians and
crowds can be used as a tool to evaluate designs of new or existing
intermodal facilities, in terms of their impact on pedestrian trafc
performance in terms of such measures as level of comfort and
time to destination. The availability of performance measures in
concert with visualization will provide a powerful evaluation tool
for both designers and planners.
When planning pedestrian environments it is necessary that
designers take into consideration how pedestrians will respond to

This manuscript was processed by Area Editor Paul Savory.


* Corresponding author. Tel.: +1 662 325 7624; fax: +1 662 325 7618.
E-mail addresses: usher@ise.msstate.edu (J.M. Usher), strawderman@ise.
msstate.edu (L. Strawderman).
0360-8352/$ - see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cie.2010.07.030

their environment as they navigate in order to complete their individual and sometimes joint missions. It is a unique problem given
that you have a crowd of persons traveling to unique destinations
for various purposes. Some may be in a hurry to get to one specic
destination, while others may be visiting intermediate destinations
for the express purpose of wasting time prior to reaching their intended destination. As well, as crowd density increases, the path
a person takes to reach their destination becomes convoluted as
they make numerous variations in their journey while navigating
the crowds. Overall, the environment is one that is highly dynamic.
Navigation is an innate ability for humans, but simulating this
capability in a virtual environment is no easy task. This paper reports on operational behaviors of pedestrian trafc that have been
reported in the literature and then describes the implementation of
these observed behaviors within the simulation system. The system is referred to as the Intermodal Simulator for the Analysis of
Pedestrian Trafc (ISAPT). The discussion focuses on the development aimed at simulating behavioral traits exhibited by pedestrians as they navigate their environment. These reported behaviors
are described in concert with the strategy used for simulating
them. Examples from the simulation are provided to illustrate
these behaviors as they occur during the pedestrian simulation.
As well, a human study of a pedestrian corridor was performed
to provide another basis for validation. The results of this study
and its simulation are presented.
2. Background
When a pedestrian enters an environment, such as an airport,
they have an overall purpose in mind. One possible purpose may

J.M. Usher, L. Strawderman / Computers & Industrial Engineering 59 (2010) 736747

be to ensure they reach their gate in a timely manner for the scheduled departure of a ight. This overall purpose will be carried out
taking into account one or more sub-objectives that are required,
or desired, in order to satisfy the overall goal. These objectives
might include such tasks as the requirement that they check their
luggage and/or get something to eat prior to arriving at their gate.
Therefore, navigation may entail the visitation of one or more
intermediate stops prior to arriving at their nal destination. While
navigating a selected path, the pedestrian will alter their route
based on such factors as crowd density and constraints imposed
by the architecture itself (e.g., walls, columns, etc.) and its contents
(e.g., furniture, planters, etc.). Given the static nature of these
items, a pedestrian is able to plan ahead, making slight modications to their direction as they move along a selected route (Bierlaire, Antonini, & Weber, 2003). While the pedestrian is following
their selected path, they will need to make modications to their
movement based on interactions with unforeseen moving obstacles that will likely cross their path as they travel. These obstacles
principally represent other pedestrians but may include such mobile items as courtesy vehicles.
In order to safely traverse their path, real-time reactionary decisions are made that steer the pedestrian to avoid collisions. Therefore, collision detection and avoidance represents a critical
component of navigation. The concept of reactive navigation does
not necessarily use predened paths. A pedestrian navigates based
on its reaction to items within the environment. Such methods include the use of social force elds, rule based methods, and XZT
space methods. Even though these methods permit a pedestrian
to navigate, the pedestrian still requires some overall goal or direction that motivates it to move. Therefore, it is not uncommon to
see these reactive methods used in combination with a dened target that a pedestrian is trying to reach, or a path that it is following.
The social force eld approach of reactive navigation involves
steering a pedestrian by the application of a combination of forces
that arise from the pedestrians interaction with the environment.
This approach is sometimes referred to by the name, particle systems, in that the pedestrians in the system are each represented
as a particle of a given mass moving in a specied direction at a dened velocity as a result of the combined effect of the surrounding
forces (Braun, Musse, de Oliveira, & Bodmann, 2003). For example,
a target position can be represented as an attractive force and
obstacles as repulsive forces. Such combined forces that dynamically change as the pedestrian moves through the environment result in a random path arising as they traverse the environment.
Helbing and Molnar (1995), Helbing and Molnar (1997) were one
of the rst to propose such an approach for pedestrian modeling.
Other examples of social force eld approaches include Lamarche
and Donikian (2004), that of Metoyer and Hodgins (2003) to support visualization of pedestrian movement as an enhancement to
the presentation of architectural and urban designs, and Heigeas,
Luciani, Thollot, and Castagne (2003) for visual rendering of
crowds in the ancient Greek agora of Argos.
A second approach involves the use of rule-based systems. Such
systems have been described as a fast approach (Soteris & Yiorgos,
2006), while at the same time touted as inappropriate for use with
large crowds (Heigeas et al., 2003). However, Loscos, Marchal, and
Meyer (2003) developed a rule-based system that is able to simulate large crowds up to 10,000 pedestrians using a 2D grid to represent the environment where each cell is either empty or
occupied by a pedestrian or obstacle (i.e., building). Pedestrian
navigation involves deciding which one of eight possible locations
to move to next. In this framework, collision detection and avoidance is implemented by each pedestrian exploring the grid up to 5
tiles ahead to identify potential collisions. A problem with this approach is that angular changes in direction are limited to 45 and 90
angles. This leads to unrealistic movement of pedestrians an affor-

737

dance Loscos is willing to take in order to be able reduce computational complexity.


Expanding on the typical rule-based approach used by 2D gridbased systems, Tecchia, Loscos, and Chrysanthou (2002) incorporates a height map whereby each cell has a specied height value
indicating the elevation of the plane represented by that cell. If the
change in elevation is not within the capabilities of the pedestrian
then this cell is treated as an obstacle around which they must navigate. As well, Tecchia et al. (2002) divide the behavioral rules
across four layers. The rst two layers focus on rules related to collision detection with obstacles (Layer 1) and other pedestrians
(Layer 2). The third layer encodes behavioral rules that may be
associated with a cell, and the fourth layer provides support for
the environment to react with a pedestrian (e.g., bus responds to
person at a stop). A novel aspect of the another rule-based approach of presented by Niederberger and Gross (2003) is that they
distinguish between the possible actions (e.g., movements to follow a path) and reactions (e.g., collision avoidance) of an agent providing top priority for reactions when the system selects from a
queue the next action to perform.
Soteris and Yiorgos (2006) also employ a 2D grid for navigation,
but extend the typical rule-based approach for navigation with the
use of a ow grid as a perception mechanism to measure overall
crowd density. The idea is that a pedestrian will choose to move
through areas of low density as they travel to their destination.
The purpose of this global vision mechanism is to read the collective behavior of the crowd; thereby, eliminating the need for each
pedestrian to have to consider each neighbor when making a decision. This perception mechanism works in simulation but one must
ask in reality how well a person is able to perceive crowd densities
at some distance from their current location.
The third method, the XZT (or XYT) approach, considers how
pedestrian ow in a 2D (XZ) coordinate system will change over
time (T). This detailed extrapolation of each pedestrians trajectory
through space permits the determination of future collisions
among the pedestrians. Feurtey (2000) used this approach to
determine whether a pedestrian should either slow down, make
direction adjustments, or speed up in order to avoid collision. A novel factor in his approach is the ability to consider multiple pedestrians in the analysis using a cost function to arrive at the best
response.
Sakuma, Mukai, and Kuriyama (2005) extend the concept of collision avoidance by working to include human limits as dened by
historical psychological experimentation, specically memory
capacity. However, they do not provide any details on how this
memory inuences the rules used for collision avoidance. Unlike
other studies, Sakuma et al. (2005) does model crowd densitys
inuence on pedestrian speed.
The ISAPT system presented in this paper makes use of an
agent-based, rule-based approach to navigation, but not with the
restrictions afforded by a 2D grid based environment. The system
employs a 3D spatially continuous domain to describe the position,
movement, and velocity of each pedestrian. As well, the characteristics of each pedestrian (e.g. age, size, mobility, etc.) are allowed to
inuence their behavior on an individual level better representing
the diverse pedestrian crowds that are commonly found within
intermodal facilities.

3. Reported pedestrian behavior


Behavioral studies found in the literature dene a number of
strategies used by pedestrians as they navigate through a crowd,
as summarized in Table 1. Related to collision avoidance, pedestrians tend to either change their trajectory or change their speed.
They also have a number of strategies that are employed when

738

J.M. Usher, L. Strawderman / Computers & Industrial Engineering 59 (2010) 736747

Table 1
Observed pedestrian behaviors from literature.
Category

Behavior

Source

Changes in speed

Each pedestrian has a current, maximum, and preferred speed

B
C
D

Pedestrian speed is constant unless conditions for a change exist


Speed change is the preferred method of collision avoidance
Speed decreases with an increase in crowd density

Decisions regarding speed change are dependent upon an individuals goals

Haklay, Osullivan, and Thurstain


Goodwin (2001)
Willis, Kukla, Kerridge, and Hine (2002)
Blue, Embrechts, and Adler (1997)
Daamen and Hoogendoorn (2003b) and
Fruin (1971)
Helbing and Molnar (1997)

F
G

Antonini, Bierlaire, and Weber (2006)


Goffman (1971)

H
I

Preference is given to maintaining their current direction


Once a trajectory change has been made, pedestrians tend to return to moving in the direction
of their original path
Pedestrians choose paths that minimize the need for angular displacements
Pedestrians prefer a smooth non-linear trajectory to an acute linear trajectory.

Passing strategies

J
K
L

Pedestrians tend to pass on the right in opposite-direction passing


Pedestrians pass on both the left and right with equal probability in uni-directional ow
If a head-on collision is imminent, both pedestrians tend to make a side-step

Goffman (1971)
Daamen and Hoogendoorn (2003b)
Helbing, Molnar, Farkas, and Bolay (2001)

Distance between
objects

M
N

Pedestrians keep minimum distance from others in crowds (territorial effect)


Distance depends on type of pedestrian and type of obstruction

Bierlaire et al. (2003)


Willis et al. (2002)

Changes in direction

passing other pedestrians. These collision avoidance patterns are


impacted not only by the individuals, but also the density of the
crowd (Bierlaire et al., 2003).
When a pedestrian realizes that a collision with another pedestrian is imminent, they generally decide a course of action. The two
most common choices are to change their trajectory or to change
their speed. Behavioral studies have shown that pedestrians prefer
changes in speed (Blue et al., 1997). The reason behind this is likely
due to the fact that it takes more physical effort to change ones
trajectory. It is much easier, both physically and mentally, for pedestrian to make a speed change. Each pedestrian has a preferred,
maximum, and current walking speed. These speeds are used to
make decisions about acceleration or deceleration as they relate
to collision avoidance (Haklay et al., 2001). A pedestrian will put
a high emphasis on attaining and maintaining their preferred
speed. It should be noted, however, that pedestrians will alter from
their preferred speed to accomplish their goal (reaching their destination). These acceleration and decelerations are made not only
to avoid collisions, but also to achieve time-dependent sub-goals.
The decision to change speeds is dependent upon the pedestrians
goals, urgency, and situation awareness. The speed of the pedestrian will remain constant unless conditions for a change of speed
are presented (Willis et al., 2002). A pedestrian will keep their current speed unless an external factor (another pedestrian, obstacle,
or time) presents a cause for change.
A pedestrians speed is impacted not only by their goals and urgency but also the surrounding crowd. With an increase in crowd
density, a pedestrians average speed is decreased. This decrease
in speed is not caused simply by a lack of physical space. The speed
decrease is likely compounded by the fact that pedestrians will be
faced with many more decision points when navigating through a
dense crowd. They will have to be very situation-aware, taking in a
high amount of information, to navigate the crowd. The time to observe the surroundings and make corresponding decisions will often slow pedestrians movement.
Daamen and Hoogendoorn (2003b) studied pedestrian speed at
two densities: high density of 0.6 pedestrians/m2 and low density
of 0.15 pedestrians/m2. The speed for each density was 1.47 m/s
and 1.57 m/s, respectively. Similarly, Fruin (1971) found a decrease
in speed with increasing density. They reported an approximate
speed of 1.14 m/s at a density of 0.6 pedestrians/m2 and a speed
of 1.40 m/s at a density of 0.21 pedestrians/m2. Though the data
are not replicated between these two studies, the emergent trend
is apparent.

Turner and Penn (2002)


Bierlaire et al. (2003)

Though pedestrians prefer changes in speed, they often are


forced to change their trajectory. Pedestrians have a tendency to
choose paths to their destination that minimize the need for angular displacements (Turner & Penn, 2002). Additionally, Bierlaire
et al. (2003) have demonstrated that pedestrians prefer a smooth
non-linear path as opposed to a linear acute path. That is, they will
not make their direction changes on a dime. The changes in trajectory will be more gradual and smooth. Pedestrians have a strong
preference of keeping their current direction as they move toward
their goal destination (Antonini et al., 2006).
A simulation model created by Blue et al. (1997) calculated that
18% of steps in a pedestrians pathway included a trajectory adjustment. Though no indication was provided regarding the extent of
the trajectory change (e.g. 10 or 90), the nding is critical. The desire for a pedestrian to remain on a straight path is circumvented
by an overriding need to change the trajectory in 18% of their steps.
It is important to note that although pedestrians may change their
trajectory, they tend to move back to their original pathway once
the collision or obstruction has been passed (Goffman, 1971).
When pedestrians are faced with collision avoidance, they often
make a decision to pass another pedestrian. While this may seem
similar to the passing of automobiles on a roadway, there are some
distinct differences. No rules exist for appropriate pedestrian passing behavior. For instance, there is nothing prohibiting a pedestrian
from crossing over into opposite owing trafc in order to pass
someone. However, in the United States, pedestrians do have a tendency to pass on the right when completing an opposite-direction
pass (Goffman, 1971). When a pass needs to be made in uni-directional ow, pedestrians tend to pass on both the left and the right
with equal probability (Daamen & Hoogendoorn, 2003b). Helbing,
Farkas and Bolay (2001) observed that when a head-on collision is
imminent, both pedestrians tend to make a side-step in order to
avoid collision.
Another important facet of pedestrian behavior is that of spacing. Individual pedestrians tend to keep a minimum distance from
others in the crowd. This is known as the territorial effect (Bierlaire
et al., 2003). Willis et al. (2002) have found that the actual distance
between people or objects depends both on the type of pedestrian
and the type of obstruction. Pedestrians take into account their
familiarity with the surrounding pedestrians, uncertainty of the
others actions, and prioritization of trajectories when maintaining
distance to other pedestrians. The distance kept from other obstacles is likely dependent on the maneuvering capabilities of the
individual pedestrian.

J.M. Usher, L. Strawderman / Computers & Industrial Engineering 59 (2010) 736747

4. Simulation system
The ISAPT system is an OpenGL-based application written in the
C++ programming language with an aim at supporting cross-platform use. The simulation was developed using the OpenSteer toolkit (Reynolds, 1999) that was originally created to aid in the
development of autonomous characters in animation and gaming
applications. It provides a foundation for the development of an
individual behavior-based simulation making use of an existing
graphical architecture. The application provides a basic framework
in which behaviors can be developed for 2D and 3D applications.
A pedestrian is represented in the simulation as a point mass
approximation providing it with the capability for linear momentum, but no rotational momentum. A pedestrian is dened in terms
of their position, mass, velocity, direction, and limitations (i.e.,
maximum force and speed). Since we are simulating humans, a pedestrians velocity arises from forces that are self-generated and
modied by changes in these forces (referred to as steering forces).
At this time, no consideration is given to forces that may arise
externally due to collisions with other pedestrians or obstacles.
The simulation employs a xed time advance mechanism with
each pedestrians position and velocity being updated in each iteration (time step).
The movement of each pedestrian within the simulation is dictated by a steering force vector that is the compilation of a set of
individual steering forces derived from individual behaviors programmed into the system. This steering force vector is resolved
into a position and velocity for the pedestrian. Unlike the 2D
grid-based approaches reviewed earlier that permit only discrete
positioning and incremental linear and angular movement of a pedestrian in the environment, the ISAPT system utilizes a threedimensional (3D) spatially continuous domain to describe the position and movement of a pedestrian. Each pedestrians position
(x(t)) is dened in terms of a 3D point in space, and vectors are employed to represent their velocity (v(t)) and acceleration (a(t)) all
dened with respect to some point in time (t) within the simulation. The Euler equations are used to simulate the physics of realistic movement of a pedestrian. Therefore, the velocity and
position of a pedestrian at each iteration is determined using the
equations:

~
v t ~
v t  1 ~at  Dt
xt xt  1 ~
v t  Dt

1
2

The acceleration of the pedestrian is computed taking into account


the overall steering force vector (F(t)) generated from the simulated
behaviors of the system and the pedestrian mass (m(t)):

~
at ~
Ft=mt

The mass of the pedestrian is depicted as a function of time given


the need to represent such tasks as baggage check or claim where
a pedestrian either reduces or increases their mass based on the
activity performed. (At this time, changes in geometry of the pedestrian area ignored.) Using these equations, the movement of every
pedestrian in the simulation is updated in response to a change in
their steering force vector prior to incrementing the simulation
clock.
The basic behaviors, currently employed within ISAPT, that contribute to the overall steering force include moving forward (in a
specied direction), path following, target seeking, braking, as well
as collision detection and avoidance of both stationary and moving
obstacles (e.g., pedestrians, benches, etc.). In this paper, the discussion will focus on the implementation of the steering force behaviors associated with collision detection and avoidance ignoring, for
the sake of brevity, the topic of dealing with stationary obstacles
since this represents a much easier task than avoiding moving

739

obstacles and utilizes much of the same logic. The other behaviors
listed have not been modied from those offered within the OpenSteer framework and are described in Reynolds (1999).
5. Pedestrian behavior and simulation
The individual basic behaviors mentioned earlier represent the
basic building blocks ISAPT uses as a means for simulating more
complex macro-level behavior (e.g., formation of lanes in pedestrian trafc). However, a problem arises in determining how best
to combine the steering force contribution of each of these behaviors to determine the overall steering force vector applied by the
pedestrian. Depending on the situation, you may want only one
behavior to execute (e.g., brake at destination) or for several to
be applied in parallel (e.g., follow the path while avoiding obstacles). To execute just one behavior means that in a single iteration
only that one behavior would determine the overall steering force.
For times when it is desired that behaviors be applied in parallel,
the system would sum their individual steering force contributions.
It is possible that a situation may arise where the effects of parallel behaviors may cancel out resulting in no overall change. This
may result in an inability to avoid a collision. Therefore, some
mechanism is needed to determine how best to combine the
behaviors considered. After experimenting with various possible
methods for combining behaviors (weighted sums, lters, priorities, etc.), Reynolds (1999) found that the technique of prioritized
dithering was most useful. This involves prioritizing the behaviors,
and then if a behaviors consideration results in no action, then the
next most important behavior is considered. This continues until
either a response is determined or all behaviors have been considered. At this time, this mechanism has been implemented using the
logic shown below.
if (near Target)
applyBrake
steeringForce = 0
else if (need to avoid pedestrians)
steeringForce += Avoid Moving Collision
else if (need to avoid obstacles)
steeringForce += Avoid Stationary Collision
else
steeringForce += Follow Path
steeringForce += Seek Target
end if

This logic is not perfect and can still result in situations where a person may collide with another person or obstacle. However, watching any crowd of persons in actions reveals this is not an unusual
behavior in reality. The best strategy for combining these behaviors
is not settled at this point and represents a continuing research effort that will be formally explored in the near future. At this time,
this paper focuses on the specic mechanisms and strategies implemented in ISAPT for behavioral simulation at the operational level.
Examples are provided to illustrate the systems capability to correctly simulate the reported observed pedestrian behaviors summarized in Table 1.
The overall approach ISAPT uses involves each pedestrian evaluating their best move taking into account the future position of
the other pedestrians in their immediate vicinity. The system does
not try to predict changes that other pedestrians may make, but
works to create a response that is the least disruptive and will
allow other pedestrians to continue in their current paths. Using
the anisotrophy property (like Hoogendoorn & Bovy, 2001) the

740

J.M. Usher, L. Strawderman / Computers & Industrial Engineering 59 (2010) 736747

pedestrian only considers those pedestrians in front of them for


collision avoidance.
5.1. Changes in speed
As stated previously, the basic drive of a pedestrian within the
ISAPT system is to move forward from its origin to its destination.
In the system, each pedestrian is dened by a set of attributes that
includes a mean (desired) and maximum speed. These values vary
from one pedestrian to the next according to a normal distribution
and are inuenced by the pedestrians age and gender in satisfaction of behavior A.1 A pedestrians current speed will be the same
as their mean speed and remain constant, as per behavior B, unless
the environmental conditions warrant a need to make a change.
Such conditions prompting a change by the pedestrian may be as
simple as an impending collision with another pedestrian, or as complex as the presence of a crowd out of which a pedestrian is unable
to navigate requiring that they adjust their speed (s(t)) to move with
the ow of the other persons in their immediate vicinity. This adjusted speed is calculated as:

st st  1 

 
t1
t2

where t1, time until the pedestrian will collide with their closest
threat; t2, lead time, the time period into the future within which
potential collisions are considered; v(t) = s(t)  (forward direction
vector).
In ISAPT, the collision detection mechanism explores the immediate vicinity of each pedestrian looking for other pedestrians that
represent potential future threats, the system extrapolates the future paths of the pedestrians in the vicinity exploring successive
time periods out in the future. For example, the position of each
pedestrian in the vicinity Dt time periods out in the future would
be determined by:

xt Dt xt v t  Dt

A collision is then tagged for those cases where the distance between these two pedestrians is less than the sum of their radii plus
a set personal space attribute of each pedestrian. This evaluation
assumes that each pedestrian is represented in the 2D plane by a
circle, but could be modied to consider other spatial geometries.
The system performs this evaluation for each pedestrian in the
vicinity, at each successive time period out in the future, up to
the specic lead time of the pedestrian under evaluation
(Dt = t2). After examining the collision potential of all other pedestrians in the immediate area, the system will return the time, t1,
representing the time until collision with the most imminent pedestrian. Given the two values, t1 and t2, their ratio provides a convenient factor for adjusting the speed to deal with an impending
threat. This capability of dealing with collision avoidance by
changing speed matches that of behavior C. The decision to change
speed as opposed to changing direction (a behavior discussed later)
in response to a potential collision is handled using probabilistic
settings with speed changes occurring x% of the time and direction
changes (1  x)% of the time with x > (1  x). A value of 80% is used
for x in the current implementation, but future studies are needed
to determine an appropriate value for this setting.
When a pedestrian is in an area consisting of several potential
threats, the same algorithm mentioned above results in a natural
reduction in speed to account for the crowded conditions illustrating the observed behavior D. Therefore, as crowd density increases
the pedestrians trailing behind others will begin to make adjust1
References to the individual behaviors listed in Table 1 will be indicated by the
use of italics (e.g., behavior A).

ments in their speeds to account for loss in mobility. Behavior E


is not currently considered by the system at this point in its development, but will come into play as the system capabilities expand
to include consideration of each pedestrians purpose and agenda,
as mentioned earlier.
5.2. Changes in direction
Behavior F has a pedestrian giving preference to maintaining
their current direction. As mentioned earlier, the basic drive of a
pedestrian is to move forward to its destination by moving in the
direction of a chosen path. The pedestrian will maintain the given
direction of their path until an impending threat arises from other
moving or unforeseen stationary obstacles. The path they are on
will already have considered visible stationary obstacles when it
was generated as a part of path planning, but sometimes a modication to this path may be needed if an unexpected stationary
obstacle is encountered (i.e., a piece of luggage left by someone
in the middle of the oor). In accordance with behavior G, if a direction change is made due to a stationary or moving object, the pedestrian will return to a trajectory in the direction of their
original path. This is illustrated in Fig. 1a where as two pedestrians
are approaching one another in opposing ow, pedestrian #2
makes a direction change to avoid pedestrian #1. The overall historical path movement is visible in Fig. 1b indicating that pedestrian #2 has returned to their original direction and will slowly
begin to migrate (as indicated by the acceleration vector pointing
down and to the left) back to the original path as they continue
on toward their destination. This return is guided by the contribution of two steering behaviors, one that seeks to move along the
path and the other that contributes to seeking the target destination. Each of these is weighted to impact their contribution.
In ISAPT, the avoidance behavior of the pedestrian depends on
the distance between the pedestrians. In a manner similar to
(Sakuma et al., 2005) and Heigeas et al. (2003), ISAPT recognizes
two boundaries and like the psychophysical volume of Heigeas
et al. (2003) the size of these volumes is variable. The outer boundary, termed the sensory boundary, is dened by the lead time of the
pedestrian (dened earlier as t2). Instead of using a set distance for
a boundary as in (Sakuma et al., 2005), each pedestrian has a personal lead time, and if another pedestrian poses a future threat in
that a collision will take place within that lead time period, then an
avoidance maneuver is prescribed by that pedestrian. This use of
time is more appropriate that a set distance since it takes into account the speed of both pedestrians; thereby, allowing a pedestrian to compensate for how quickly the threat is approaching
giving them sufcient time to react. This lead time is adjusted in
response to changes in crowd density as prescribed by Fruin
(1971). The inner boundary, termed the personal boundary, is provided for those cases where an adjustment is made (by the pedestrian themselves or perhaps someone else) that results in a sudden
encounter of another pedestrian collision threat deep inside the

Fig. 1. Changes in trajectory.

J.M. Usher, L. Strawderman / Computers & Industrial Engineering 59 (2010) 736747

741

sensory boundary. Such an encounter requires a quick drastic


change (e.g., stopping, or side-stepping) to avoid a collision.
In the scenario of Fig. 1, pedestrian #2 has a longer lead time
(larger sensory distance) than #1 (3 s versus 2.5 s); therefore, it detects and reacts to pedestrian #1 before pedestrian #1 ever reacts.
This is the reason why the path of pedestrian #1 is unaltered. If
they have the same lead time, then both pedestrians will notice
each other and both make subsequent adjustments to their path.
Behaviors H and I are a function of the method used to determine what overall paths a pedestrian will follow as they navigate
to their destination. The idea is that they will choose a path that
will result in a streamlined route from origin to destination. This
is the idea of smoothing a path of sharp edges. This behavior is
not embodied at the collision detection and avoidance level, but
manifested at the strategic and tactical planning levels where
paths are determined and re-evaluated on a periodic basis. At the
operational level, this is supported by the implementation of the
physics of pedestrian movement in that sharp turns at normal
walking speed are not really possible, non-linear trajectories require that the pedestrian rst slow or stop, and then change direction prior to moving again.
5.3. Passing strategies
The behavioral strategies for passing (behaviors J through L) are
implemented within the steering system of ISAPT with observance
of behaviors J and L as discussed above. As mentioned, when a pedestrian within the sensory boundary represents a collision threat
then a temporary modication of trajectory is considered. When
deciding on how a pedestrian should avoid a potential collision,
most systems only consider the position and direction of the one
pedestrian threat they are facing ignoring any other pedestrians
in the immediate area. Aside from Feurteys (2000) use of a complex analytical approach that does take into account multiple
pedestrians in determining a path, only some of the 2D grid-based
approaches will make limited consideration of others in their rulebased approach, but even in these cases the extent of consideration
is not robust and limited by the number of rules employed. Osaragi
(2004) employs an equation to calculate a measure of collision
that is used to determine if a pedestrian should move left or right
to avoid a collision. This measure divides the forward eld of an
agent in two and computes a measure for each side taking into account those pedestrians on a side that are within a dened region.
The measure is computed as the sum of the product of the time and
distance separating the pedestrian in question from each other pedestrian. The pedestrian will them move to the side with a lower
value of the measure. Hoogendoorn and Bovy (2001), employ a
method that takes into account density, likelihood of interaction,
destination, and preference in the computation of a utility value
for deciding which side to choose.
In a similar manner, the ISAPT system has the capability to consider all pedestrians in the immediate area of the threat. The system rst looks ahead to the identied pedestrian threat and
divides the eld of view (FOV) into two halves (left and right). It
then computes a density measure for each area that takes into account not only the distance, but also the relative direction of each
pedestrian in each half of the FOV. The equation is:

Density measure

X directioni 
i

distancei

Pedestrians moving in the same direction are favored over those


moving in the opposite direction. The further away a pedestrian is
the less impact it has on the measure. Following consideration of
all other pedestrians in the area, the pedestrians avoidance maneuver will be to the side with the smaller density measure. The density

Fig. 2. Passing strategy employing density measure.

considerations of individual pedestrians are what contribute to lane


formation in pedestrian trafc. This behavior is illustrated in Fig. 2
where one can see that the pedestrian has favored (as it should) the
side with those pedestrians moving in their same direction. Even
though passing to the left of the threat (pedestrian #5) is possible,
this would place the pedestrian in a ow counter to their own. In
the situation where the density values are equal, then if the threat
is approaching the pedestrian from the front, a right side pass is favored (as per behavior J) and if the pedestrian is approaching the
threat from behind, then passing left or right is chosen with equal
probability (as per behavior K).
5.4. Distance between objects
Behavior M involves keeping a minimum distance from other
pedestrians. This is controlled by that inner boundary that was
mentioned earlier. When the inner boundary of a pedestrian is
penetrated, then using the logic discussed in Section 5, they will
slow down, stop, or move aside depending on the position of obstacles in its immediate area. This will continue until something in
their immediate environment changes (e.g., another pedestrian
moves out of the way or changes their speed). As per behavior N,
each pedestrian has their own value for this inner boundary representing their personal comfort zone, which is currently set to values on the order of 00.5 m. However, at this time, this distance is
not a function of the obstruction type.
6. Validation
The validity of ISAPT will be demonstrated in two ways. First,
emergent crowd behaviors reported in the literature will be compared to those demonstrated with ISAPT. Due to the fact that the
simulation is modeled at an individual pedestrian level, any emergent collective behavior that has been previously veried serves to
validate the model. Additionally, specic relationships between
pedestrian metrics (e.g. density, ow) will be demonstrated. Once
again, these relationships have been reported in the literature and
will be used to test the validity of ISAPT. The last test will involve

742

J.M. Usher, L. Strawderman / Computers & Industrial Engineering 59 (2010) 736747

comparing the results of a human study of a pedestrian sidewalk


with a simulation of the same system.
6.1. Emergent collective behaviors
A variety of crowd behaviors have been documented in the literature, based on not only simulation studies but empirical data as
well. The most widely referenced collective behaviors include lane
formation, intersection striping, bottleneck negotiation, aisle
usage, and the impact of density on pedestrian speed.
For the system runs covered in this section, ISAPT was set up to
simulate pedestrian trafc moving either uni- or bidirectionally in
a hallway that is 8 m wide. The region of interest for observation
was a 26 m long hallway containing no other entry or exit points.
The pedestrians were initially randomly distributed throughout
the hallway in both the horizontal and vertical direction. The desired velocity of each pedestrian was generated from a normal distribution with a mean of 1.2 m/s and a standard deviation of 0.2 m/
s. The number of pedestrians in the system was kept constant in
each simulation run. When a pedestrian exits the region of interest,
the system would reintroduce that pedestrian back at the start of
the hallway with their initial position across the width of the hallway (y-value) at the entry point being determined using Eq. (7)
where the current position is altered by a normally distributed value (mean zero and variance of 1) whose value is constrained to
0.5 m.

NewY-Position CurrentY-Position ClipN0; 1; 0:5; 0:5

6.1.1. Speed as a function of density


Given the many reports on the fact that pedestrian walking
speed is most signicantly impacted by trafc density, we desired
to see if the micro-simulation of pedestrian trafc would reproduce this behavior. The number of pedestrians was kept constant
in each simulation run and ranged from 20 to 220 pedestrians. Given that the desired speed (free ow) of each pedestrian varied, the
cycle time of each pedestrian was logged after an initial warm-up
period. The average cycle time of each pedestrian was then computed and expressed as a fraction of their free-ow speed.
The results of these runs are shown in Fig. 3 where the average
fraction of free-ow speed for all pedestrians is plotted versus the
available area per pedestrian in the system. These results illustrate
that ISAPT correctly simulated the trend whereby pedestrian trafc
slows as trafc density increases. An early published report by Fruin (1971) illustrating the impact trafc density has on speed for pedestrian in a walkway is shown in the same gure for comparison.
The graph represents the results tting a curve to the observations
reported by Fruin. Note the similarity in the shape of the curves
with the greatest difference being 10% at high density (1 m2/s).

6.1.2. Lane formation


When pedestrian density exceeds a critical value, dynamic lanes
emerge. The pedestrian lanes consist of pedestrians that share the
same intended direction and approximately the same velocity. The
number of lanes formed depends on the density of pedestrian trafc and the width of the walkway. The formation of lanes is a result
of self-organized pedestrian behavior. Pedestrians prefer to walk
behind another pedestrian rather than making their own path. By
joining a travel lane, a pedestrian is able to minimize interactions
that require avoidance maneuvers. This leads to more efcient travel for pedestrians (Daamen & Hoogendoorn, 2003a; Daamen &
Hoogendoorn, 2003b; Goffman, 1971; Helbing, Buzna, Johansson,
& Werner, 2005; Helbing & Molnar, 1997; Helbing et al., 2001;
Weng, Shen, Yuan, & Fan, 2007).
When two lanes are formed, the division generally occurs in the
center of the walkway. Using the same hallway setup as was simulated above, experiments were run to determine if the micro-level
simulation of pedestrian behavior would result in the formation of
lanes. The previous setup was modied such that the number of
pedestrians would be equally divided between those traveling in
each of the two directions in the hallway. The initial conditions were
such that the pedestrians were randomly placed on their side of the
hallway resulting in an initial mass chaos of pedestrian movement
as they strive to navigate to the opposite end. Fig. 4 illustrates the
initial conditions for a run with 80 total pedestrians in the system.
Multiple simulation runs were made with different numbers of
pedestrians in the system to illustrate the various lane formation
behaviors that would result. Table 2 shows the results from running
ISAPT for various crowd densities. This table lists the direction of
each lane, its average width, and the number of pedestrians occupying the lane. Three lanes form in most cases, but the number of
lanes and the direction of the lanes vary. Fig. 5 illustrates the case
where two lane form as a result of the interaction of the 80 pedestrians within the system, while Fig. 6 shows three lanes for the case
of 120 pedestrians. Lane formation occurs quite quickly with the
initial semblance of lanes appearing within 30 s and clear lane distinctions with almost no deviation at around 70100 s into the run.

6.1.3. Speed distribution across a hallway


Pedestrians who have a faster speed than the average lane travel tend towards the outer edges of the walkways (Helbing et al.,

Fig. 4. Initial conditions for bidirectional ow of 80 pedestrians.

Table 2
Lane formation results.
Number of pedestrians No. lanes Lane: ow direction/width (m)/occupancy

Fig. 3. Speed as a function of trafc density.

40
60
80
80
80
100
120
140
200

3
3
3
3
2
3
3
3
2

Top lane

Middle lane

Bottom lane

Right/2/12
Right/2.5/21
Right/2/17
Left/1/8
Right/4/40
Left/2/26
Right/2.5/31
Right/2/33
Right/4/100

Left/4.5/20
Left/4.5/30
Left/4/40
Right/4.5/40

Right/1.5/8
Right/1/9
Right/2/23
Left/2.5/32
Left/4/40
Left/2/24
Right/2/29
Right/2/37
Left/4/100

Right/4/50
Left/3.5/60
Left/4/70

J.M. Usher, L. Strawderman / Computers & Industrial Engineering 59 (2010) 736747

743

ior at these intersections depends on the number of walking directions present. When only two trafc directions are present at an
intersection, striping formations emerge. If the two directions are
exactly opposite, the striping becomes lane formation. If, however,
the two directions are not exactly opposite, pedestrians form
stripes to proceed through the intersection. This is most noticeable
when two wide streams of pedestrians intersect. According to
Helbing,
Fig. 5. Lane formation for 80 pedestrians.

Fig. 6. Lane formation for 120 pedestrians.

2001). This was tested using the same hallway setup for the case of
80 pedestrians moving in two lanes of bidirectional ow. Speed
data of each pedestrian was collected as they passed through a
0.1 m wide vertical zone halfway down the hallway. This zone
was further subdivided into 0.2 m wide blocks and the speed and
position of each pedestrian whose center passed through this block
was averaged. Graphs of these results are shown in Fig. 7 and illustrate the existence of a speed gradient supporting the observed
behavior.
6.1.4. Intersection striping
At intersections, pedestrians are forced to interact with other
pedestrians traveling in different directions. The emergent behav-

Stripes are a segregation phenomenon, but not a stationary


one. The stripes are density waves moving in the direction of
the sum of the directional vectors of both intersecting ows.
Stripes extend sidewards into the direction that is perpendicular to their direction of motion. Therefore, pedestrians move forward with the stripes and sidewards within the stripes
(Helbing et al., 2005).
The emergence of stripes allows pedestrians to move through
the intersection without the need to stop. Similar to lane formation, striping at intersections maximizes travel efciency by limiting obstructions and increasing average speed. To test this
behavior, an intersection between two hallways was modeled in
ISAPT with the method of pedestrian generation being identical
to that used in previous experiments and the initial position of
each pedestrian of the 140 pedestrians randomly generated. To
best illustrate this behavior we rst removed the variability in
the walking speed of the pedestrians. This would mean that once
the crowd reached steady state it would be easier to see the striping phenomena if it occurred. Fig. 8 shows the results indicating
very clear vertical stripes in the crowds motion. Fig. 9 illustrates
the results when variability in walking speed is permitted. One
can see that the simulated crowd still exhibits the striping behavior although not as clearly as before.
6.2. Pedestrian sidewalk study
An additional method for validation was replicating empirical
data with the simulation of a pedestrian corridor. To achieve this,
video footage of an outdoor sidewalk was captured. The sidewalks
area of interest measured 11.3 m long by 3 m wide. There were no
obstructions on either side of the sidewalk. The location of the
sidewalk was in front of a campus building and observations were
recorded during a break between classes. Therefore, the majority
of the 51 pedestrians in the footage were students. Of the 51

Fig. 7. Speed variation across hallway.

Fig. 8. Stripe formation with no variability.

744

J.M. Usher, L. Strawderman / Computers & Industrial Engineering 59 (2010) 736747

The desired (mean) and max speeds for each agent are set to
the observed value for each individual pedestrian they
represent.
 Case 2: The time between arrivals of agents is the same as in
Case 1. The desired speeds of the agents are normally distributed with those walking in each specic direction taken from
a different distribution whose parameters match those given
in Table 3. The max speeds of each agent are calculated by multiplying their desired speed by the ratio value from Table 3 for
their particular direction (e.g., the max speed of an agent walking left will be 2.12  desired speed).
 Case 3: Time between arrivals of all agents follows an exponential distribution with a mean value equivalent to the mean
observed value (1.59 s), and each agents desired and max
speeds are dened using the same procedure as in Case 2.

Fig. 9. Stripe formation with variability.

pedestrians, 18 entered the sidewalk on the left (moving right) and


33 entered on the right (moving left). The footage also captured
data for one bicyclist.
Video footage was captured using a high denition digital video
camera and saved as image stacks. This resulted in a compilation of
901 images. The micro-level behavior of each pedestrian was captured by tracking pedestrian coordinates frame by frame resulting
in over 4000 points used in the analysis. These coordinates were
then used to calculate pedestrians paths, speeds, and interactions.
Additional details on the empirical data collection and analysis
methods employed can be found in (Lee, Strawderman, & Usher,
2008).
When comparing the empirical results to the simulation results,
we will use the term pedestrian to refer to the observed pedestrian in the system (empirical data), and the term agent to represent the simulated pedestrians (simulation data). The goal of this
validation phase is to show that the simulation accurately depicts
the sidewalk system. This is shown by comparing the simulation
data to the empirical data. The characteristics of the pedestrians
speed are given in Table 3.
6.2.1. Simulation setup and experimentation
A sidewalk of the same dimensions as mentioned above was
constructed in the simulation making use of penetrable barriers
to dene the edges of the sidewalk. These types of barriers provide
a boundary for the agent that acts as a guide but which can be penetrated if desired for emergency purposes (i.e., overcrowding).
ISAPT has the capability to make use of varying levels of detail with
respect to the assigned walking speeds for agents, their interarrival
times, as well as origin and destination positions. Three different
experiments (cases) were setup to test the simulation varying the
level of detail provided to the simulation. The performance statistic
of interest is the mean time in the system for the pedestrians.
 Case 1: The time between arrivals of each agent to the system,
from either direction, is taken directly from the observed data.

Case 1 represents simulating a pedestrian area where there is an


abundance of pedestrian data available. Given that this is not often
the situation, Case 2 relaxes the data requirements by making use
of a selected distribution to characterize the desired and max
speed of each pedestrian that enters the system. Case 3, then represents the most popular scenario for the use of such a tool, where
distributions are commonly selected to represent both the interarrival times and speed of the pedestrians. In each experiment, the
same number of agents (51 pedestrians and one bicyclist) were
generated and the simulation was run until all agents exited the region of interest.
6.2.2. Simulation results
Ten replications were run for each of the three cases and the results are shown in Tables 4 and 5. The results in Table 4 report the
statistics on the time in the system from both the experimental
and simulation studies. Overall, the simulation did an excellent
job of predicting the time in the system behavior for this pedestrian corridor. Errors on the overall mean time in the system ranged from 0% to 1.71% and errors on the variance range from 2%
to 20% depending on the level of detail of the input data used in
the simulation. It is interesting to see that relaxing the data
requirements between Cases 2 and 3, did not have much impact
the accuracy of the results. The last column of data does a pair wise
comparison, looking at the absolute value of the difference between each of the pedestrians observed in the study and that of
their corresponding simulated agent. For Case 1, the mean of this
difference is in the vicinity of just a 1/20 of a second with the results of Cases 2 and 3 not being much different from each other
with mean differences of approximately of 1 s. At worst, this represents an 813% error when looking at the time in the system for
the individual pedestrians within the simulation.
The time in the system for the observed pedestrians was normally distributed as were the results of the replicated experiments
across each pedestrian. Given this, a paired t-test was performed
comparing the mean time across the 10 replications for each of
the 51 agents with their pedestrian counterpart to get an idea if
a signicant difference existed. The results shown in Table 5 and
indicate additional support of ISAPTs ability to model the macrolevel system behavior of the pedestrian trafc.

Table 3
Pedestrian statistics.
Walking direction

Left
Right
Both

Pedestrian speed (m/s)

Max speed (m/s)

Max speed/mean speed

Mean

Variance

Mean

Variance

Mean

Variance

1.30
1.43
1.34

0.0083
0.0285
0.0189

2.75
3.18
2.90

0.115
0.236
0.197

2.12
2.23
2.16

0.049
0.047
0.050

745

J.M. Usher, L. Strawderman / Computers & Industrial Engineering 59 (2010) 736747


Table 4
Simulation results.
Case

Walking direction

Number of pedestrians

Simulation
Time in system (s)

Observation
Time in system (s)

Error (%)

Mean
1

Right
Left
Both

18
33
51

8.08
8.78
8.53

Right
Left
Both

18
33
51

Right
Left
Both

18
33
51

Variance

Mean

Variance

Mean

Variance

Mean

Variance

1.22
0.52
0.86

8.06
8.77
8.52

1.01
0.52
0.80

0.25
0.11
0.12

20.79
0.00
7.50

0.063
0.030
0.042

0.041
0.012
0.022

8.02
8.79
8.52

1.03
0.49
0.82

8.06
8.77
8.52

1.01
0.52
0.80

0.50
0.23
0.00

1.98
5.77
2.50

1.108
0.722
0.865

2.282
0.875
1.359

8.17
8.92
8.65

0.98
0.46
0.78

8.06
8.77
8.52

1.01
0.52
0.80

1.36
1.71
1.53

2.97
11.54
2.50

1.119
0.718
0.870

2.123
0.876
1.302

Table 5
Results of paired t-test.

Per agent
|TSim  TObs|

Table 7
Deviation in travel path for Cases 1 and 3.

Case

t value

95% CI

P-value

Case 1

1
2
3

0.56
0.02
1.08

(0.030, 0.053)
(0.262, 0.129)
(0.114, 0.381)

0.575
0.981
0.285

|Dx|

|Dy|

|Dx|

|Dy|

0.0937
0.00546
0.487
0.0310

0.0949
0.00422
0.357
0.0257

0.4070
0.0725
1.3460
0.0757

0.0878
0.00284
0.2639
0.0224

To provide a more complete picture, the length of time required


for all pedestrians to exit the system was also considered. This
length of time is inuenced by the interarrival time characteristics
along with the mean and max speed of the pedestrians and their
interaction during their journey through the pedestrian corridor.
Table 6 reports these results for the three cases indicating errors
in the range from 0% to 1.4%.
In addition to the macro-level behavior, there is also interest in
the ability of ISAPT to capture the micro-level behavior of a pedestrian as they walk from origin to destination. Given the data collected in the sidewalk study, it is possible to map out the path
taken by a pedestrian for direct comparison with their corresponding simulated agent. The position of each pedestrian was extracted
from the video data and a corresponding position for each agent
was recorded during each replication of the simulation. The x-axis
represents a pedestrians position along the length of the sidewalk
and the y-axis represents their position relative to the width of the
sidewalk (with the point (0, 0) at the top left of the frame). In the
simulation, the x and y positions of each agent were averaged
across the 10 replications of the simulation at each instance in time
(Dt = 0.1 s) and compared with the positions of their corresponding
pedestrian from the video study at that same point in time. For
each agent, the mean of the difference in both the x and y position
values between each pedestrian and agent for their entire path
were computed. Table 7 lists the statistics on this data for the results of Case 1 and Case 3. (Given the similarity between Cases 2
and 3, no analysis was performed on the Case 2 data.) These results
illustrate how well the simulation is able to represent the micro-level navigational behavior of the pedestrians. These results could
probably be improved further if the experimental data is smoothed
to account for any possible error on the part of the operator when
interpreting the video position of each pedestrian.
Looking at some of the specic pedestrian interactions gives us
an idea of the navigational capability of ISAPT. In the video footage
Table 6
Time last pedestrian exits the system.
Case

Simulation

Observation

Error (%)

1
2
3

90.10 s
89.69
88.85

90.10 s
90.10
90.10

0.00
0.46
1.39

Mean
Variance
Max
Min

Case 3

it is easy to discern that both pedestrians 9 and 13 alter their paths


to avoid a potential head-on collision with each other. Fig. 10a
shows a graph of their x and y positions for the paths of both the
pedestrian and agent, while Fig. 10b and c shows their respective
x and y positions over time. Given the orientation of Fig. 10a, pedestrian 9 is traveling from left to right and pedestrian 13 walks
from right to left. In this gure you can see the adjustments made
to their paths in response to a potential collision. At the beginning
of their journey you can see that pedestrian 9 begins to make an
adjustment for pedestrian 13 long before the encounter. From
Fig. 10b we can see that the change by pedestrian 9 begins at
approximately 16.5 s into the study period, while pedestrian 13 begins their change at 19.6 s. Fig. 10c shows that they will actually
pass each other at approximately 21.3 s into the observation period. Using Fig. 10c and extrapolating the x-position of pedestrian 13
at 16.5 s, it was determined that pedestrian 9 begins their path
change while 7.6 m away from pedestrian 13, whereas pedestrian
13 begins their change while only 5.5 m separates them from impact. This illustrates the variability in each pedestrians response
time. In ISAPT this parameter can be individually set and altered
for each pedestrian. This is an important capability given the variability of this value across cultures and the fact that it is a function
of crowd density.
Given that the time of collision would have been at 21.3 s, you
can see in Fig. 10b that at this point in time, the y-axis separation
between the two pedestrians is 0.93 m what that between the
agents is 0.68 m. The path followed by pedestrian 13 shows two
adjustments. The rst at x = 11 m is in response to pedestrian 9
and another that begins at approximately 5 m (24 s) for a bicyclist
that enters the observation area at time 29.8 s. The paths taken by
the agents exhibits similar behavior to the pedestrians. Examining
the agents paths we see that agent 13 responds to agent 9 with
their initial path closely following that of their pedestrian counterpart. However, an overcorrection results that is a combination of
the approaching bicyclist (mentioned above) and their goal of
reaching their target position at the exit from the system. (Each
agent is given a destination when they enter the system that is
based on the actual exit point of their pedestrian counterpart. This
destination contributes to the overall steering behavior of the

746

J.M. Usher, L. Strawderman / Computers & Industrial Engineering 59 (2010) 736747

(a)

(a)

(b)

(b)

Fig. 11. Travel paths of from-behind passing pedestrians/agents 12 and 15.

(c)

in the plots of each pedestrian and its corresponding agent. Overall,


the maximum value of the x and y-axis deviations of agent 12 are
0.10 and 0.13 m, while the x and y-axis deviations for agent 15 are
0.25 and 0.15 m.
7. Conclusions

Fig. 10. Travel paths of head-on passing pedestrians/agents 9 and 13.

pedestrian.) Looking at the path of agent 9 in Fig. 10a, a slight


correction that occurs at approximately 7 m along the x-axis is a
navigational correction made by the agent to ensure that they
not walk off the sidewalk. Overall, the maximum value of the
x and y-axis deviations of agent 9 are 0.25 and 0.28 m, while the
x and y-axis deviations for agent 13 are 0.36 and 0.68 m.
As opposed to a head-on collision, another popular dynamic between pedestrians is when two pedestrians are traveling in the
same direction but at different speeds such that the faster pedestrian will maneuver to pass the slower one. This behavior was observed in the interaction between pedestrians 12 and 15 as they
walked from right to left (high value on x-axis to low value). Pedestrian 12 was traveling at a mean rate of 1.01 m/s, while pedestrian
15 walked at the mean rate of 1.36 m/s. Fig. 11a shows that the
point at which they would pass one another (intersect on the xaxis) falls outside of the observation region, occurring at an
approximately x-axis value of 2 (at around 32 s into the study).
From Fig. 11b we can extrapolate that the distance separating their
centers when they pass will be approximately 0.65 m, indicating a
close proximity when passing occurs. In this gure you can see that
the passing behavior of the pedestrian is accurately simulated by
the agent along with the actual travel path as seen by the overlap

ISAPT provides a system with the capability to simulate pedestrian trafc within intermodal facilities. Some of the unique characteristics of ISAPT include the fact that as opposed to the
discrete 2D grid-based systems that currently exist, ISAPT provides
a 3D spatially continuous domain to represent the position and
movement of pedestrians. The attributes of each pedestrian can
be set independently taking into account variations in their weight,
age, size, and mobility. Pedestrian speeds are allowed to vary on a
continuous scale following observed distributions characterized in
terms their preferred and maximum speeds. As opposed to analytical methods that are grounded in force theory alone, ISAPT has the
advantage in that the individual steering behaviors were developed based on the micro-level pedestrian behaviors that have been
reported in the literature. The animation of pedestrian trafc takes
place in real time permitting simultaneous display of the trafc
ow while the simulation is running.
The results of the pedestrian validation study indicate ISAPTs
ability to model pedestrian trafc at both the macro and micro-level. From the individual behaviors of the simulated agents emergent collective behaviors arose that were validated by published
studies of pedestrian behavior. These behaviors including lane formation, intersection striping, distribution of speeds across a corridor, and the reproduction of the speed versus density graphs from
the literature. As well, based on a simulation of an observed pedestrian corridor, the simulation was able to predict the macro-level
performance of the system while also exhibiting observed microlevel behavior of the individual pedestrians within the system.
The analytical results demonstrate ISAPTs ability to generate accurate trafc performance parameters (mean time in system, trafc

J.M. Usher, L. Strawderman / Computers & Industrial Engineering 59 (2010) 736747

ow, etc.) that are commonly used in the evaluation of intermodal


facility designs and layouts.
As can be expected in pedestrian trafc, the video from the pedestrian study showed several of the pedestrians walking together
in groups of two. The simulation, at this time, is not able to directly
simulate this behavior in providing a group type of response to
external stimuli. However, in Case 1 of the sidewalk validation
study, since the mean and max speeds of the agents match those
of the observed pedestrians, then in the simulation this grouping
behavior is seen visually, but the behavior deviates if the group
is required to respond to a collision threat since a group response
is not currently programmed into the system.
Another limitation of the current system is that if the origin and
destination of a pedestrian are separated across a large distance
and the architecture/layout of static obstacles between them is
somewhat complex, then reactive navigation can sometimes result
in an unusual path for the pedestrian. This problem requires that
the user dene paths composed of intermediate points between
the origin and destination to overcome the complexity of the buildings architecture.
Future work includes conducting additional human studies to
validate the simulation model on a larger scale. Work is in progress
on the simulation of an airport lobby area involving a host of services to travelers. To handle this complex system, ISAPT will be expanded to include the concept of pedestrian agendas incorporating
more advanced pedestrian maneuvers and planning tactics to handle activity scheduling and route planning within more complex
building architectures.
Acknowledgement
This project is sponsored by the US Department of Transportation (Grant No. DTOS59-06-G-00041).
References
Antonini, G., Bierlaire, M., & Weber, M. (2006). Discrete choice models of pedestrian
walking behavior. Transportation Research Part B: Methodological, 40(8),
667687.
Bierlaire, M., Antonini, G., & Weber, M. (2003). Behavioral dynamics for pedestrians.
In K. Axhausen (Ed.), Moving through nets: The physical and social dimensions of
travel (pp. 118). Elsevier.
Blue, V. J., Embrechts, M. J., & Adler, J. L. (1997). Cellular automata modeling of
pedestrian movements. In IEEE int. conference on systems, man, and cybernetics
(pp. 23202323).
Borgers, A., & Timmermans, H. (1986). A model of pedestrian route choice and
demand for retail facilities within inner-city shopping areas. Geographical
Analysis, 18(2), 115128.
Braun, A., Musse, S. R., de Oliveira, L. P. L., & Bodmann, B. E. J. (2003). Modeling
individual behaviors in crowd simulation. In Proc. of the 16th int. conf. on
computer animation and social agents (CASA03), 89 May 2003 (pp. 143148).
Cetin, N., Nagel, K., Raney, B., & Voellmy, A. (2002). Large-scale multi-agent
transportation simulations. In Proc. of the annual congress of the European
Regional Science Association (ERSA), Dortmund, Germany, August 2002.
Daamen, W., Bovy, P. H. L., & Hoogendoorn, S. P. (2001). Modelling pedestrians in
transfer stations. In M. Schreckenberg & S. D. Sharma (Eds.), Pedestrian and
evacuation dynamics (pp. 5974). Duisburg: Springer Verlag.
Daamen, W., & Hoogendoorn, S. P. (2003a). Experimental research of pedestrian
walking behavior. In Transportation Research Board annual meeting (pp. 116).
National Academy Press.
Daamen, W., & Hoogendoorn, S. P. (2003b). Qualitative results from pedestrian
laboratory experiments. In E. R. Gales (Ed.), Pedestrian and evacuation dynamics
(pp. 121132). London: CMS Press.

747

Feurtey, F. (2000). Simulating the collision avoidance of pedestrians. Dept. of


Electronic Engineering, University of Tokyo, Tokyo, Masters.
Fruin, J. (1971). Pedestrian planning and design. Metropolitan Association of Urban
Designers and Environmental Planners, Elevator World.
Goffman, E. (1971). Relations in public: Microstudies of the public order. New York:
Basic Books.
Haklay, M., OSullivan, D., & ThurstainGoodwin, M. (2001). So go down town:
Simulating pedestrian movement in town centres. Environment and Planning B:
Planning and Design, 28(3), 343359.
Heigeas, L., Luciani, A., Thollot, J., & Castagne, N. (2003). A physically-based particle
model of emergent crowd behaviors. In Proc. of GraphiCon 2003, Moscow,
September 510, 2003.
Helbing, D., Buzna, L., Johansson, A., & Werner, T. (2005). Self-organized pedestrian
crowd dynamics: Experiments, simulations, and design solutions.
Transportation Science, 39(1), 124.
Helbing, D., & Molnar, P. (1995). Social force model for pedestrian dynamics.
Physical Review E, 51(5), 42824286.
Helbing, D., & Molnar, P. (1997). Self organization phenomena in pedestrian crowds.
In F. Schweitzer & H. Haken (Eds.), Self organization of complex structure: From
individual to collective dynamics (pp. 569577). Gordon and Breach: Amsterdam.
Helbing, D., Molnar, P., Farkas, I. J., & Bolay, K. (2001). Self-organizing pedestrian
movement. Environment and Planning B: Planning and Design, 28, 361383.
Hoogendoorn, S. P., & Bovy, P. H. L. (2001). Nonlocal continuous-space microscopic
simulation of pedestrian ows with local choice behavior. Transportation
Research Record, 1776(01-2095), 201210.
Jiang, B. (1999). SimPed: Simulating pedestrian ows in a virtual urban
environment. Journal of Geographic Information and Decision Analysis, 3(1),
2130.
Lamarche, F., & Donikian, S. (2004). Crowd of virtual humans: A new approach for
real time navigation in complex and structured environments. In
EUROGRAPHICS 2004 (pp. 509518).
Lee, H., Strawderman, L., & Usher, J. M. (2008). Utilizing video footage for the
analysis of pedestrian behavior. In Industrial engineering research conference,
IERC, Vancouver, BC, May 1721.
Loscos, C., Marchal, D., & Meyer, A. (2003). Intuitive crowd behavior in dense urban
environments using local laws. In Proc. of the theory and practice of computer
graphics (TPCG03), June 35, 2003 (pp. 122129).
Metoyer, R. A., & Hodgins, J. K. (2003). Reactive pedestrian path following from
examples. The Visual Computer, 20(10), 635649.
Niederberger, C., & Gross, M. (2003). Hierarchical and heterogenous reactive agents
for real-time applications. In Eurographics 2003, CD-ROM.
Okazaki, S. (1979). A study of pedestrian movement in architectural space. Part 1:
Pedestrian movement by the application of magnetic models. Transactions of AIJ,
283, 111119.
Osaragi, T. (2004). Modeling of pedestrian behavior and its applications to spatial
evaluation. In Int. conf. on autonomous agents and multiagent systems
(AAMAS04), New York, NY (pp. 834841).
Parker, D. C., Manson, S. M., Janssen, M., Hoffmann, M., & Deadman, P. (2003). Multiagent systems for the simulation of land-use and land-cover change: A review.
Annals of the Association of American Geographers, 93(2), 314337.
Reynolds, C. (1999). Steering behaviors for autonomous characters. In Proc. of game
developers conference, San Jose, CA.
Sakuma, T., Mukai, T., & Kuriyama, S. (2005). Psychological model for animating
crowded pedestrians. Computer Animation and Virtual Worlds, 16, 343351.
Soteris, S., & Yiorgos, C. (2006). Crowd self-organization, streaming and short path
smoothing. In WSCG2006, 14th int. conf. in central Europe on computer graphics,
visualization and computer vision, Plzen, Czech Republic, January 30February 3,
2006.
Tecchia, F., Loscos, C., & Chrysanthou, Y. (2002). Visualizing crowds in real-time.
Computer Graphics Forum, 21(4), 753765.
Turner, A., & Penn, A. (2002). Encoding natural movement as an agent-based
system: An investigation into human pedestrian behaviour in the built
environment. Environment and Planning B: Planning and Design, 29, 473490.
Weng, W. G., Shen, S. F., Yuan, H. Y., & Fan, W. C. (2007). A behavior-based model for
pedestrian counter ow. Physica A: Statistical and Theoretical Physics, 375(2),
668678.
Willis, A., Kukla, R., Kerridge, J., & Hine, J. (2002). Laying the foundations: The use of
video footage to explore pedestrian dynamics in PEDFLOW. In M. Schreckenberg
& S. D. Sharma (Eds.), Pedestrian and evacuation dynamics (pp. 181186). Berlin:
Springer.
Zhang, Q., Han, B., & Li, D. (2008). Modeling and simulation of passenger alighting
and boarding movement in Beijing metro stations. Transportation Research Part
C: Emerging Technologies, 16(5), 635649.

Вам также может понравиться