Department of Mathematical Sciences

MSc. in Modern Applications of Mathematics

The Dynamics of Crowds

Gareth William Parry

Supervisors: Prof. C. J. Budd & Dr. C. J. K. Williams

September 2007
Abstract

Crowds pose an interesting example of a complex system in which emergent behaviour is

observed out of the interaction of many individual agents. This behaviour can be very

important in the safe design of sports and other stadia, especially in the case of possible

emergency situations. For this reason crowd dynamics is of great interest to architects.

The purpose of this project is to review the current literature on crowd dynamics and to

look at some simulations of different types of crowd behaviour in certain specific

situations namely: the meeting of two crowds in a ’scramble crossing’, the motion of a

crowd though an exit and the response of a crowd to an emergency (such as a fire). The

student will implement a differential equation model of the crowd crossing problem. In

particular, they will study the emergent behaviour that arises in the scramble crossing

problem and see how this depends on various parameters relating to the individual actions

of the members of the crowd.

Declaration: I hereby certify that the work in this document is my own, unless otherwise

referenced.

Signed ............................................................ Gareth Parry

i
Acknowledgements
I would like to thank Professor Budd in particular for his help and input whilst writing

this thesis. The lively debate and insightful questions raised in our meetings was one of

the particularly enjoyable aspects of the project. I would also like to thank Dr Chris

Williams and Odysseas Kontovourkis for their help in providing an Architects insight

into pedestrian dynamics.

Movies and Code

Submitted along with this thesis are copies of several movies created by the simulator and

converted to *.avi files. Also included on the CD are the final MATLAB script and

function files (hard copies are in Appendix A) along with several *.mat files which are

saved data from different simulations. The movie and *.mat files are named with the

number of pedestrians simulated followed by the geometry the flow was simulated in. So

if 500 pedestrians were modelled in a Scramble Crossing then the associated files would

be: Movie: 500scramblecrossing.avi

Data: 500scramblecrossing.mat

Finally a *.PDF version of this thesis is also provided for submission to a plagiarism

prevention website.

ii
Contents

Introduction....................................................................................................................... 1

An Overview of Models for the Simulation of Pedestrian Dynamics........................... 3

2.1 Pedestrian Modelling approaches ............................................................................. 4
2.2 The Gas-Kinetic Model of Pedestrian Flows............................................................ 5
2.3 The Magnetic Force Model....................................................................................... 9
2.4 Queuing Systems .................................................................................................... 12
2.5 Cellular Automata................................................................................................... 18
2.6 The Social Force Model.......................................................................................... 22
2.6.1 The Driving Force............................................................................................ 23
2.6.2 Pedestrian Interactions ..................................................................................... 24
2.6.3 Boundary Interactions...................................................................................... 27
2.6.4 Attractive Interactions...................................................................................... 27
2.6.5 Individuality and Random Behaviour Fluctuations ......................................... 28
2.6.6 Another Formulation........................................................................................ 28

Self-Organisation Phenomena in Pedestrian Flows..................................................... 30

3.1 Intersecting Flows ................................................................................................... 30
3.2 Bottlenecks.............................................................................................................. 31
3.3 Lane Formation ....................................................................................................... 31
3.4 Shockwaves............................................................................................................. 32

Implementation of the Social Force Model................................................................... 33

4.1 Numerical Solution of the Social Force Model ...................................................... 33
4.2 The Desired Destination and Waypoints ................................................................ 37
4.3 The Desired Velocity .............................................................................................. 39

Simulation and Analysis of the Social Force Model .................................................... 41

5.1 The Model Constants .............................................................................................. 41
5.2 Pedestrian Counterflows ......................................................................................... 44
5.3 A Ninety Degree Corner ......................................................................................... 47
5.4 The Scramble Crossing ........................................................................................... 51
5.5 Model Shortcomings ............................................................................................... 54

Summary and Future Directions ................................................................................... 56

iii
Bibliography .................................................................................................................... 58

Appendix A ...................................................................................................................... 59

MATLAB Codes.............................................................................................................. 59
A.1 SFMDFPIBI.m ....................................................................................................... 59
A.2 fun5.m ........................................................................................................................................... 63

iv
List of Figures
Figure 1: Acceleration force A acting on a to avoid collision with b............................... 12
Figure 2: Planar Graph G′ (V ′, E ′ ) and dual graph G (V , E ) ............................................ 14

Figure 3: Lane switching behaviour in Blue/Adler model................................................ 21

Figure 4: Emerging Lane Formation................................................................................. 21
Figure 5: Distance between α and β ............................................................................... 25

Figure 6: Fαβ characterises anisotropic behaviour ............................................................ 26

Figure 7: Striping phenomena in Intersecting flows......................................................... 30

Figure 8: illustration of pedestrian behaviour causing shockwaves ................................. 32
Figure 9: Cputime for several solvers ............................................................................... 34
Figure 10: Verlet - Sphere................................................................................................. 35


Figure 11: Pedestrian paths with a single point for p , desired destination ...................... 37
Figure 12: Pedestrian paths using a desired destination array .......................................... 37
Figure 13: Pedestrians path with and without a waypoint ................................................ 38
Figure 14: The difference between a fixed desired velocity and a variable one............... 40
Figure 15: An example of pedestrian flows in the test room (Not to Scale) .................... 42
Figure 16: The initial and final pedestrian positions created for a counterflow ............... 45
Figure 17: These figures show how flow oscillates at a bottleneck ................................. 46
Figure 18: The Speed of a single Pedestrian around a ninety degree corner .................... 47
Figure 19: The simulation of a ninety degree corner........................................................ 48
Figure 20: The overall mean speed of pedestrians around a ninety degree corner........... 49
Figure 21: Plots of the speed characteristics of a Scramble crossing ............................... 51
Figure 22: Variance and Standard Deviation of Pedestrian in a Scramble Crossing........ 52
Figure 23: Regions where Lattice Constants are calculated..............................................53
Figure 24: The New Verlet-Sphere................................................................................... 55

v
List of Tables
Table 1: Simulation Parameters ........................................................................................ 43
Table 2: Lattice Constants ................................................................................................ 54

vi
Chapter 1

Introduction
The design and planning of large pedestrian areas has become increasingly important in

recent times with the vast number of facilities having to cope with large volumes of

pedestrian traffic. Notable examples of such buildings are airports, shopping centres,

night clubs and sports stadia which have become larger and more commonplace in recent

times. One of the key considerations in the architecture of these situations is the

behaviour of the people that use them, in particular their reaction in panic situations.

With the prevalence of these types of structures and tragic events which have resulted in

the massive loss of life, the modelling, simulation and understanding of pedestrian

movement in emergency egress situations is a necessity.

In many instances the fatalities and injuries in emergency evacuations were not caused by

the hazard leading to the need for egress but the actions of the crowd itself. Stampedes

are caused both by real hazards, such as fire, but also the behaviour of the crowd, perhaps

in forcing its way out of a stadium. The most obvious example of such a situation is the

Hillsborough disaster, when on April 15th, 1989; ninety-six Liverpool FC fans lost their

lives resulting in the conversion of many football stadiums in the United Kingdom to all-

seater and the removal of barriers at the front of stands.

1
The key aim of this thesis is to explore the current theories and models relating to

pedestrian flows and implement two of these schemes. There is a massive breadth of

literature relating to such models which have been developed over the last forty years.

The model this thesis will focus on and implement is the Social Force Model which was

developed D. Helbing and P. Molnár[1],[2].

The ultimate aim of this is to explore the emergent behaviour of pedestrian systems,

particularly in the case of a scramble crossing and the expected striping effects. Other

models will be discussed and outlined in some detail to highlight the variety of theories

suggested to simulate pedestrian dynamics. The scheme proposed by V. J. Blue and J. L.

Adler, their Cellular Automata[3] model, will be explored in some detail as initially it

was intended that this model also be implemented in the MSc project.

2
Chapter 2

An Overview of Models for the Simulation of Pedestrian

Dynamics

The simulation of pedestrian movement has being explored in a variety of ways. In this

chapter the aim will be to provide an outline of the most prominent of these models along

with an exposition of the governing equations of each scheme. In general these models

describe the forces each pedestrian feels; treating each pedestrian as a particle in a larger

system and using Newton’s Second Law to evaluate position, velocity and acceleration.

A numerical solver then can be used over a discrete timestep providing a velocity and

position update. One of the key differences between pedestrian traffic models and other

roadway-based traffic systems is that pedestrian locations are not restricted to a single

dimension. Whilst many of the models are initially based on vehicle traffic systems, these

are only one dimensional models and unlike pedestrian movement subject to a number of

laws and restrictions governing traffic. Pedestrian movement is inherently more

changeable than that of vehicles, as unlike vehicle flow which is controlled by well

defined lane markings with lane change and passing opportunities restricted, there are no

such restrictions on pedestrian walkways. Pedestrian interaction is also markedly

different to that of cars since safety concerns are much less, clearly pedestrians can

actually touch each other without incident, which is certainly not true of moving vehicles.

Also pedestrians often move in pair or clusters, such as couples or family groups, whilst

such attractive influences are rare in vehicle traffic. This leads to interactions between

people that have to be considered in any model, examples of which are bumping into

3
each other, exchange of places or bypass when pedestrian density is high as opposed to

sidestepping which would be analogous to the behaviour of a car. The velocity and

acceleration characteristics of pedestrians are also very different, with each pedestrian

having their own desired and maximum speeds. They are also able to accelerate to full

speed from standstill almost immediately and can change speed more rapidly allowing

them to take advantage of gaps in traffic when they arise.

2.1 Pedestrian Modelling approaches

Pedestrian flow models can be classified in different ways depending upon how the

scheme treats the pedestrians and the level of detail of the models. These classifications

are:

1. Microscopic models, which consider individual pedestrian behaviour separately.

The pedestrian behaviour in these models is often described by their interactions

with other pedestrians in the system.

2. Mesoscopic models, which do not consider each pedestrian individually but the

overriding characteristics, such as velocity distributions. The pedestrian behaviour

is described microscopically though not specifically but rather in terms of velocity

distributions.

3. Macroscopic models, which do not make distinctions between individual

pedestrians nor describe their individual behaviour but consider the flow in terms

of density, average velocity and flow patterns.

As previously discussed, the main focus of this thesis will be on the Social Force Model

and Cellular Automata which are both examples of microscopic models. Whilst these are

4
computationally demanding, both the individual and emergent group behaviour of

pedestrians is of interest.

2.2 The Gas-Kinetic Model of Pedestrian Flows

This model was first suggested by L. F. Henderson[4], who treats large pedestrian crowds

like molecules in a dilute gas. Whilst there are some seemingly random fluctuations in the

movements of people in large crowds, the fact that each individual has a mass and

velocity suggests that the classical Maxwell-Boltzmann statistics could be used to

describe the motion of a crowd using a density function f ( x , v , t ) . Henderson only

applies the Maxwell-Boltzmann equations to the so called gaseous phase. This is when

the crowd is moving and has a low particle density, defined as the number of persons per

unit area. If this is small then each individual is assumed to be able to move at their

desired speed. However the model becomes problematic when boundary interactions

occur and particle density increases, for example at a doorway. This is described as a

phase transformation to a densely packed crowd liquid phase. To describe the crowd gas

several assumptions are made, firstly that movement takes place on a continuous plane

and that at time t each of the N pedestrians has a position (x,y) and velocity (Vx,Vy).

Secondly the crowd is considered to be homogeneous, that is each particle will have the

same mass and probability of velocity components. This homogeneity is analogous to

chemical purity in molecular systems, although Henderson suggests that this

homogeneity may not be a fair assumption due to what he describes as sexual

inhomogeneity, the idea that men and women behave differently in crowds. This

nonuniformity may extend beyond gender and also be attributed to other social and

environmental factors such, as the age of each pedestrian. The next assumption is that the

5
particles are independent of each other in position and velocity components with position

and velocity of the individual pedestrian being uncorrelated. Finally, the crowd is

assumed to be in equilibrium and that it can be treated as a statistical ensemble of any

individual. The Maxwell-Boltzmann equations for the described assumptions are as

follows, with the probability density function P(Vx), for a single fluctuating velocity

component, Vx, is

1 dN vx 1 § 1 V2 ·
P (Vx ) ≡ = exp ¨ − 2 x ¸ (2.2.1)
N dVx 2π vr .m.s © 2 vr .m.s ¹

Where vr .m.s is the standard deviation of the speed v ≡ V . The expression for Vy is

analogous and can be combined to get the resultant velocity, V

1 dNV 1 § 1 V2 ·
P (V ) ≡ = exp ¨ − 2 ¸ (2.2.2)
N dV 2π vr2.m.s © 2 vr .m.s ¹

With the probability density function for speed

1 dN v π v 2 § π v2 ·
P(V ) ≡ = exp ¨ − 2 ¸
with v = π / 2vr .m.s (2.2.3)
N dv 4 v2 © 4v ¹

These functions can be extended to situations where there may be a superimposed flux

upon the system, for instance a crowd moving along a corridor. (2.2.1) Is shifted in this

case with a new velocity component Vx′ ≡ Vx − Vx

1 dNVx′ 1 § 1 V ′2 ·
P (Vx′ ) ≡ = exp ¨ − 2x ¸ (2.2.4)
N dVx′ 2π vr .m.s © 2 vr .m.s ¹

With analogous treatment of (2.2.2) and (2.2.3).

Clearly this is a somewhat simplistic view of pedestrian dynamics, but it is the starting

point for S. Hoogendoorn and P. H. L. Bovy’s formulation [5]. In the simplest case where

6
there is no distinction between pedestrian types, the gas-kinetic equations represent

describe the dynamics of the generalised phase-space density ρ = ρ ( t , x , v , w ) defined

by r ( t , x ) f ( v , w | t , x ) . Here r ( t , x ) is a multidimensional density which reflects the

expected number of traffic entities per unit volume at ( t , x ) , with f ( v , w | t , x ) the joint

probability density function of the velocity v and continuous attributes w which reflects

characteristics of traffic flow and its constituent entities such as desired velocity. The gas-

kinetic equation in n dimensions is

(I ) ( II ) (
IV )
  (V )
 
 (
III )
    
∂ρ § ∂ρ · § ∂ρ ·
+ ∇ x .( ρ v ) + ∇ v .( ρ A) + ∇w .( ρ B ) = ¨ ¸ +¨ ¸ (2.2.5)
∂t © ∂t ¹event © ∂t ¹cond

This equation shows how the phase-space density changes, with term (I) being

convection, (II) acceleration, (III) the adaptation of continuous attributes, (IV) event

based noncontinuum processes and (V) the condition-based noncontinuum processes. The

so-called pedestrian phase-space density (P-PSD – this is consistent with Hoogendoorn’s

notation) ρ ( t , x , v , w ) conforms to the setup laid out in the introduction to this chapter,

that is it is a two dimensional system to describe the pedestrian flow where each

pedestrian may have different velocity, desired velocity and acceleration considerations.

With this in mind, pedestrian density ( r , x ) is defined to be the expected number of

pedestrians per unit area.

In (2.2.5) terms (I) – (III) are the continuum processes, that is they are smooth and

describe the change in the spatial distribution of the pedestrians (in terms of density).

They represent continuous changes in the independent variables x , v , w . Terms (IV) and

7
(V) describe the noncontinuum processes of pedestrian interaction – Hoogendoorn calls

this a stimulus-response mechanism: “an event causes a remedial manoeuvre of the

impeded pedestrian” and as such aren’t continuous but occur when pedestrians engage

with each other. The current research omits terms (III) and (V). The resulting equation to

describe the P-PSD is

 (I )
  ( II ) ( IV )

+ −

( ρ v1 ) + ( ρ v2 ) + ( ρ A1 ) + ( ρ A2 ) = §¨ ¸· + §¨ ¸·
∂ρ ∂ ∂ ∂ ∂ ∂ρ ∂ρ
+ (2.2.6)
∂t ∂x1 ∂x2 ∂v1 ∂v2 © ∂t ¹event © ∂t ¹event

With A1 and A2 describing the acceleration laws of the system and the terms in (IV)

describing noncontinuum events which increase or decrease the phase-space density

respectively. Hoogendoorn and Bovy describe in detail the derivation of each term,

however it is sufficient for this thesis to describe the model in general without detailing

specifics.

One area of interest in the derivation however is the role of transition probabilities in the

pedestrian interactions, as these probabilities describe how the pedestrian’s direct

environment affects their behaviour. Three types of pedestrian interaction are

distinguished, which correspond to the stimulus-response mechanisms:

1. One-sided interaction – where pedestrian p catches up to a slower moving

pedestrian q. Here p is held up by q but the converse isn’t true.

2. Two-sided interaction – where two pedestrian travelling in opposite directions

meet head on and hold each other up.

8
 3. Passive interaction – this is the same as 1 except the pedestrian being considered

is the one being caught up and so will not take any action as their progress is

unimpeded.

Transition probabilities are then used to describe the expected behaviour of each

pedestrian in the system in one of the three interactions described.

The model also touches on factors that are relevant in several models, particularly the

pedestrian’s spatial requirements and different classes of pedestrian. The Gas-Kinetic

model traditionally assumes that the particles in the system are infinitesimally small, this

however is not necessarily a valid assumption for modelling pedestrian flows since the

amount of space each pedestrian occupies is of dominant importance. The pedestrian

classification is the same problem Henderson touched on, in that gender, age or other

demographic characteristics may affect the behaviour of individuals in the system.

Hoogendoorn and Bovy then solve the formulated problem using the Monte Carlo

method in a variety of simple cases such as unidirectional, bidirectional and crossing

pedestrian flows. Their results are reasonable and reproduce the expected velocity-density

relations qualitatively.

2.3 The Magnetic Force Model

The next model was developed by S. Okazaki and S. Matsushita[6] and as the title

suggests treats the pedestrians within the system as charged objects within the resulting

magnetic field. Each pedestrian in the system is given a positive charge and destinations

such as doorways or service counters a negative charge. Clearly the attractive nature of

9
opposite magnetic charges results in the pedestrians moving towards these destination

points. This magnetic effect also means that pedestrians exert a repulsive force upon each

other which physically corresponds to them avoiding collisions with each other and

objects such as boundaries.

Once the room has been setup, it is inputted as a series of vertexes described in the

Cartesian coordinates. With the destination details configured the model then requires a

detailed amount of input data to produce realistic results before the simulation can be

started. The required data is the desired destination, initial position, initial velocity, the

pedestrian orientation, time that the pedestrian starts walking and their method of walk.

This final input is the most interesting as the simulation can implement a wayfinding

technique where the pedestrians will follow a sequence of points (corners in the case of

this simulator) until they have an unobstructed route, that is no boundaries in their path,

to their desired destination. Another point of interest in this model is the velocity input,

which is a maximum velocity. This is simply because if there were no upper bound on the

velocity the pedestrians would accelerate without limit according to Coulomb’s Law.

Since the pedestrians are treated as magnetic objects, the appropriate force law is

Coulomb’s Law

1 Q1.Q2
F= (2.3.1)
4πε s 2

Q1 and Q2 are the signed values for the magnetic charge of the objects they represent, s is

the distance between the two particles and k, the so called Coulomb constant, is the

constant term 4πε1 . If Q1 represents a positively charged pedestrian, a, say and Q2 a

10
negatively charged pole representing the desired destination. Since a two dimensional

setup is being considered the force is a vector quantity resulting in the following form of

Coulomb’s Law

Q1.Q2
F = k. .sˆ (2.3.2)
s3

Here ŝ is the unit vector pointing from Q1 to Q2 . If the charges have the same sign, as in a

pedestrian-pedestrian interaction, then the resulting force is positive which corresponds to

a repulsive interaction. Conversely if the signs are different, in the case of a pedestrian-

destination interaction, there is a negative force and so an attractive force is felt which

leads to the particle accelerating towards the destination. When more than two charges

are present in the system the forces are superimposed, that is the force between any pair is

the sum of all the exerted forces from the component charges.

The model also incorporates another force which acts upon the pedestrians in order to

simulate the collision avoidance characteristics of the crowd. If two pedestrians come

within a certain distance of each other then the new force is exerted upon the pedestrians.

That is, if a, intersects within a specific area of another pedestrian then a feels the

following repulsive force and resulting acceleration to cause a change of direction and

thus preventing a collision. This acceleration is represented by

dva
= tan β .cos α .va (2.3.3)
dt

Where va is the velocity of pedestrian a, β is the angle between the relative velocity (RV)

of a to pedestrian b and the contacting line from the position of pedestrian a to the circle

around pedestrian b. This circle is the “Pedestrian Area Module” which is equivalent to

11
the “Territorial Sphere” in Helbing’s Social Force Model [1]. Finally α is the angle

between relative velocity of pedestrian a (RV) to the pedestrian b and the velocity of

pedestrian a.

Figure 1: Acceleration force A acting on a to avoid collision with b

This model was used to simulate to simulate an escape from fire on one floor of an office

building, to plot the movement of pedestrians in part of an underground railway station

and pedestrian flows in a hotel lobby. The model can be used to evaluate how long an

emergency escape might take, the behaviour of pedestrians in queuing situations – that is

the number in each queue, the length of their wait and movement processes.

2.4 Queuing Systems

Queuing theory, which is generally considered to be a branch of operations research,

describes pedestrian flows in terms of probability functions. The pedestrian will arrive at

12
a given node, which represents a server, with a certain probability. They will then spend a

certain amount of time being served, at a shop till for example, and then continue on to

their next destination, leaving the queue. A queuing system is comprised of three

elements, the pedestrian’s arrival in the queue, the service mechanism and the service

discipline. A queuing discipline determines the manner in which the exchange handles

calls from customers. Examples are

• First In, First Out – This principle states that customers are served one at a time

and that the customer that has been waiting longest is served first

• Last In First Out – This principle also serves customers one at a time, however

the customer with the shortest waiting time will be served first

• Processor Sharing – Customers are served equally. Network capacity is shared

between customers and they all effectively experience the same delay

Queuing is handled by control processes within exchanges, which can be modelled using

state equations. Queuing systems use Markov Chains which model the system in each

state where Incoming traffic to these systems is modelled via a Poisson distribution. The

stochastic process in a queuing system is the population of a particular room.

The first model to be discussed was formulated by S. J. Yuhaski, Jr and J. Macgregor

Smith[7] which develops a state dependent queuing model for the congestion effects of

movement through circulation systems of a building. Circulation systems are, in this case,

the pathways of movement such as corridors, stairways and ramps. The problem they

describe is that of crowded pedestrian flows in confined spaces and the paper describes

the following as “crucial aspects to movement Systems”

13
 1. The service rate of the movements system decays with increasing traffic

2. The amount of available space within the movement system is finite

These characteristics are the basis for the construction of their model and their aim to

“best capture the congestion effects”. The first step is to generate a representation for the

facility; this is its floor plan. This can be represented by a planar graph G′ (V ′, E ′ ) (this is

consistent with their notation) and so the queuing network is the Dual Graph of G ′ . There

are then two distinct types of spatial entity which must be distinguished, that is the

activity network and the circulation network. So V is partitioned into two sets V = { A, S } .

The set A = { A1 , A2 ,..., AN } is the set of activity nodes which represent so called “activity

areas”, say a department of a shop. The set S = {S1 , S2 ,..., SM } then is the circulation

nodes which are the movement pathways joining activity areas.

Figure 2: Planar Graph G′ (V ′, E ′ ) and dual graph G (V , E )

14
One feature of modelling pedestrian facilities is that transitions within a network are not

virtually instantaneous as in other systems (telephone networks for example) and so the

set S is required. Vertices from this set represent additional nodes within a facility which

handle the flow of pedestrians from Ai to Ak without interrupting service.

The model makes the following assumptions; firstly that there are J customer classes,

which wish to use the facility, drawn from an infinite population and each class has K

generating sources. Then the average arrival rate of type j per unit time from source k is

λ jk ( j = 1, 2,...J ; k = 1, 2,...K ) . The next assumption is that each pedestrian of type jk will

follow a deterministic routing vector through the facility, called a “customer chain”. This

vector has elements rjkl ( l = 1, 2,...L jk ) where the lth element represents the destination of

the pedestrian to the next resource after they have been served at their previous

destination. So, customers of the type jk enter a system of queues in independent Poisson

streams at a rate λ jk and follow a sequence of queues before leaving the facility. A

Poisson stream of arrivals corresponds to arrivals at random. In a Poisson stream

successive customers arrive after intervals which are independently exponentially

distributed.

The next stage is to model a single corridor as a queuing system. The corridor is given a

maximum capacity of C = [5LW] where L is the corridor length and W the width. It is

then assumed that pedestrians enter the corridor with the behaviour of a Poisson stream of

rate λ , and the time each person occupies the corridor is exponentially distributed with

rate µn . Thus, there is a state dependent service rate, meaning it is a function of the

15
number of pedestrians in the corridor. The model now uses the Chapman-Kolmogorov

steady-state difference equations for the state probabilities p1 , p2 ,..., pc

λ0 λ1...λn −1
pn = p (2.4.1)
µ1µ 2 ...µn 0

Such that

1 C
ª λ λ ...λ º
= 1 + ¦ « 0 1 n −1 » (2.4.2)
p0 n =1 ¬ µ1 µ 2 ...µ n ¼

Here the arrival rates are not influenced by the number in the queue so

λ = λ0 = λ1 = ... = λC . The paper suggests two congestion models with approximate

overall walking-speed Vn , firstly the linear relation

1.5
Vn = (C +1 − n) (2.4.3)
C

Or the exponential relation, which may be more accurate, is

ª § n − 1 ·γ º
Vn = A exp « − ¨ ¸ » (2.4.4)
¬« © β ¹ ¼»

A is the amplitude with parameters β and γ are called the scale and shape parameters. In

both models the service rate rn , is the average of the inverse of the time it takes for the

pedestrians to travel the length of the corridor

Vn
rn = (2.4.5)
L

With overall service rate

µn = nrn (2.4.6)

The overall service rate for the linear model then becomes

16
 1.5
µn = (C + 1 − n ) (2.4.7)
C

Then the Chapman-Kolmogorov equations become

λn
pn = n
p0 (2.4.8)
§ A · n

¨ ¸
© LC ¹
∏ ( C − i + 1) i
i =1

And

1 C
kn
= 1+ ¦ n
(2.4.9)
p0 n =1
∏ ( C − i + 1) i
i =1

With k n = λ LC
A . Similarly the Chapman-Kolmogorov equations for the exponential model

are

λn
pn = p0 (2.4.10)
n
§ A· ­° ª § i − 1 ·γ º ½°
∏ i ¨ ¸ exp ® « − ¨
i =1 © L ¹
¸ »¾
°¯ ¬« © β ¹ ¼» °¿

And

­ ½
° °
C ° °
1 λn
= 1+ ¦ ® ¾ (2.4.11)
p0 n =1 ° n § A· ª § i −1 · º °
γ

° ∏ i ¨© L ¸¹ exp « − ¨ β ¸ » °
¯ i =1 «¬ © ¹ »¼ ¿

Another notable Queuing Model is the one developed by Løvas[8] introduced a similar

stochastic model where pedestrians can be modeled in a queuing network. This model is

setup similarly where Nodes in the network represent rooms and links the doors. Each

pedestrian will then select a new node with a given probability. The model can evaluate

several performance measures such as the mean number of persons in a node and the

mean number of safe evacuees. Løvas has developed a tool called EVACSIM which has

17
simulated egress in several different setups. The system visualizes the movement of the

pedestrians and provides qualitative information about behaviour at bottlenecks in the

queuing systems.

2.5 Cellular Automata

A cellular automaton is a discrete model studied in computability theory, mathematics,

and theoretical biology. It consists of a regular grid of cells, each in one of a finite

number of states. Time is discretised and the state of a cell at time t is a function of the

states of a finite number of cells (called its neighbourhood) at time t − 1. These

neighbours are a selection of cells relative to the specified cell, and do not change

(though the cell itself may be in its neighbourhood, it is not usually considered a

neighbour). Every cell has the same rule for updating, based on the values in this

neighbourhood. Each time the rules are applied to the whole grid a new generation is

created. The Cellular Automata method can be used to simulate pedestrian flow. It is fast

and relatively simple, with the walkway represented by a cellular grid. In this

representation each cell within the grid is occupied by one pedestrian. The pedestrian

flow is modelled by a set of governing rules which differ according to the particular

model being considered.

The most notable Cellular Automata model is the one developed by V. J. Blue and J. L.

Adler ([3],[9]) which models the walkway as a circular lattice (closed loop) with width

W, length G and class L = WG. Each cell within the lattice is assigned a label

L ( i, j ) with1 ≤ i ≤ W and 1 ≤ j ≤ G . The density of pedestrians on the walkway is

determined at the start of the simulation and remains constant throughout. The

18
Microsimulation continues in discrete timesteps ti for i = 1, 2,..., T . The lane assignments

and speed updates change the position of all the pedestrians within the lattice in four

stages determined by the local rules which are applied to each individual on the walkway.

The stages are

1. A set of lane change rules determining the lane for each pedestrian on the lattice

2. The pedestrians are moved in to the assigned lanes

3. A set of rules is applied to find the allowable speed of each pedestrian based on

the available gap ahead and the pedestrians desired speed

4. Forward movements based on the allowed speeds are made

The rule sets for each stage are

Lane Change (parallel update 1 – stages 1 and 2)

1) Eliminate conflicts: If two walkers are adjacent then they cannot sidestep into each
other
a) If a cell is available between two walkers then assign it to one of then with a
50/50 split

2) Identify gaps: The lane (same or left/right adjacent) is chosen which best advances
forward movement upto v_max according to the gap computation subprocedure that
follows the step forward update

a) For Dynamic Multiple Lanes (DML):

i) Step out of lane if a walker in the opposite direction is within 8 cells by
assigning gap = 0
ii) Step behind a pedestrian moving in the same direction when avoiding a
collision with an oncoming walker by choosing any available lane with
gap_same = 1 when gap = 1

b) Ties of equal maximum gaps ahead are resolved according to:

i) If the 2-way tie is between adjacent lanes then a 50/50 split between which
lane is chosen
ii) If the 2-way tie is between the current and an adjacent lane then the walker
stays in lane
iii) If there is a 3-way tie between the current and adjacent lanes then the
pedestrian stays in lane

19
3) Move (this is stage 2): Each pedestrian pn is moved 0, +1 or -1 lateral sidesteps after
1) – 3) is completed

Step forward (parallel update 2 – stages 3 and 4)

1) Update velocity: Let v ( pn ) = gap where gap is from the subprocedure below

2) Exchanges: IF gap = 0 or 1 AND gap = gap_opp (cell occupied by an opposite

moving pedestrian) THEN with probability p_exchg,
v ( pn ) = gap + 1 ELSE v ( pn ) = 0 . That is the walker and the opposite moving
pedestrian either exchange cells with a predefined probability p_exchg or stop when
they meet each other and sidestep at the next timestep

3) Move (this is stage 4): each pedestrian pn is moved v ( p n ) cells forward

Gap Computation Subprocedure

1) Same direction: Look ahead a up to 8 cells (8 = 2* the largest v_max) IF an occupied

cell is found with a pedestrian moving in the same direction THEN set gap_same to
the number of cells between the two pedestrians ELSE set gap_same = 8

2) Opposite Direction: IF an occupied cell is found with an opposite moving pedestrian

THEN set gap_opp to (0.5*number of cells between the pedestrians) ELSE gap_opp
=4

3) Assign gap = MIN(gap_same, gap_opp, v_max)

The lane switching procedure is illustrated in Figure 3. For pedestrian 1, the adjacent

cells on both the left and right are available. The unoccupied distance on however is

uniquely maximal in the present lane so no switch is required. Pedestrian 4 conflicts to

the right with pedestrian 6 and in this instance pedestrian 6 is given access to this cell,

hence 4 can only sidestep to the left. The unoccupied distance is largest in this lane and

so 4 will switch. The change is represented by 4’. Pedestrian 7 cannot switch lanes since

both adjacent lanes are occupied and thus the lane change procedure is terminated.

20
 Figure 3: Lane switching behaviour in Blue/Adler model

This model was successfully applied to both uni and bi-directional flows. The results

reproduce observed phenomena particularly the expected lane formation which is shown

in figure 4 where grey cells are left moving pedestrians and the black right moving

pedestrians

Figure 4: Emerging Lane Formation

21
Another interesting model was developed by J. Dijkstra, J. Jessurun and H. Timmermans

[10] which looks at pedestrian movement with a shopping centre. The rule set for this

model is

1) Check decision point: If a pedestrian has passed a decision point (the end of an

activity or node in the network) then got 3)

2) Check the cell type – examine the behaviour of the pedestrian and the walkers desired

direction followed by a change into that direction then a decision point will be passed

3) If the cell is free then the pedestrian can move into that cell, otherwise got 4)

4) If the cell to the left/right isn’t occupied then move there

In this model the movement is directed only toward the destination and can only change

at decision points with pedestrian interactions not being considered.

2.6 The Social Force Model

This model has been developed primarily by D. Helbing and P. Molnár ([1], [2], [11]).

They describe the idea of social forces in the context of ordinary pedestrian behaviour.

That is, in general a pedestrian will be used to the majority of situations that confront

them and they have prescribed behavioural strategies based on previous experience of

similar situations. As such they will react in the best way, that is the most efficient for

them and as such pedestrian movement is automatic and thus predictable. The Social

Force Model is a Microsimulation of each individuals behaviour and so each

pedestrian, α , in the system can be represented by a point rα ( t ) in space, which changes

continuously with speed being governed by the equation of motion

drα ( t )

= vα ( t ) (2.6.1)
dt

22


Similarly, the speed vα ( t ) is continuously changing and thus acceleration is governed by

the social forces fα ( t ) , which represent the sum of the different influences upon the

individual pedestrian (that is environment and other pedestrians). There is also a

consideration for random fluctuations within the system which account for random


behavioural fluctuations which gives rise to ξα ( t ) . So, the acceleration obtained is

dvα


= fα ( t ) + ξα ( t ) (2.6.2)
dt

The model being implemented in this thesis takes into account an acceleration force





fα0 ( vα ) , repulsive effects of boundaries fα B ( rα ) , repulsive interactions with other

pedestrians fαβ ( rα , vα , rβ , vβ ) and attraction effects fα i ( rα , ri , t ) leading to

fα ( t ) = fα0 ( vα ) + fα B ( rα ) + ¦ fαβ ( rα , vα , rβ , vβ ) + ¦ fα i ( rα , ri , t ) (2.6.3)
(≠ )
β α i

2.6.1 The Driving Force

As the name suggests, the driving force, is the component of the social forces which

describes each individuals desire to move to their intended destination with some desired


velocity vα0 . The desired direction of motion is given by eα and deviations of the actual

velocity vα from the desired velocity vα0 ( t ) = vα0 ( t ) eα ( t ) are corrected within the so

called relaxation time τ α . The equation which describes this motivation is

1


fα0 = ( vα0 ( t ) eα ( t ) − vα ( t ) ) (2.6.4)
τα

The desired direction of the pedestrian is described by

p − rα
eα ( t ) =

(2.6.5)
p − rα

23



Where p is the desired destination and rα the current position.

It is often the case that pedestrians may be delayed, at bottlenecks or doorways for

example, which leads to an increase in the desired speed over the course of time. One

way to describe this effect is to implement the following

vα0 ( t ) = ª¬1 − nα ( t ) º¼ vα0 ( 0 ) + nα ( t ) vαmax (2.6.6)

With, vαmax the maximum desired velocity and vα0 ( 0 ) the initial one. The parameter

vα ( t )
nα ( t ) = 1 − (2.6.7)
vα0 ( t )

Describes the impatience of the pedestrian to reach their destination, with v ( t ) the

average speed into the desired direction of motion. This particular effect may lead to a

crowd developing pushy behaviour thus increasing the pressure within the crowd, which

may lead to clogging effects which may have disastrous consequences.

2.6.2 Pedestrian Interactions

The repulsive force term fαβ ( rα , vα , rβ , vβ ) describes the interactions between two

pedestrians α and β , and the desire of pedestrian α to keep a certain distance from β .

This term is described by

ª ( rαβ − dαβ ) º
ª ( rαβ − dαβ ) º

fαβ ( t ) = Aα1 exp « » n . F
αβ αβ + A2
α exp « » nαβ (2.6.8)
«¬ Bα1 »¼ «¬ Bα2 »¼

The first term of this force describes the tendency to respect the private sphere of each

individual and also helps to avoid collisions if there are sudden changes within the

system. The second term accounts for the behaviour of physical interaction in high

24
densities and pushy crowds if so called frictional effects are ignored. A second

formulation which includes frictional effects will be briefly introduced later. The

parameters Aαi and Bαi denote the interaction strength and range respectively. These

parameters are often dependent on cultural influences, for example the personal space

expected may vary depending on the society being modelled. The parameter dαβ is the

distance between the centres of mass of the pedestrians being considered, rαβ is the sum

of the radii of pedestrians α and β and nαβ is the normalised vector pointing from β to α

xα ( t ) − xβ ( t )
nαβ = (2.6.9)
dαβ ( t )

Where xα ( t ) is the point of the centre of mass of α at time t , and similarly for β . Figure

5 shows visually these parameters.

β


eα ( t )
rβ
nαβ

ϕαβ dαβ − ( rα + rβ )
rα

Figure 5: Distance between α and β

Each pedestrian in the system is modelled to have a solid core, rather than just be a point.

However this does not have a great impact on the results of the simulation and in [2] the

pedestrians were considered to be point particles without any problem. The final

25
parameter in (2.6.8) is Fαβ accounts for the anisotropic behaviour of the pedestrians. In

this context it means that the actions of those in front of our pedestrian have a greater

impact on their movement than those behind. Pedestrians can often predict what may

occur infront of them and can react accordingly. Thus Fαβ is a factor which gives the

pedestrians within view greater influence than those out of view. This is done by having

this factor depend on the angle, ϕ , between the walkers desired direction of movement

and the direction of the pedestrian exerting the repulsive force. This is also shown in

figure 5.

1 + cos (ϕαβ )
Fαβ = λα + (1 − λα ) (2.6.10)
2

λα is the potential of the anisotropic character of the pedestrians, with

vα ( t )
cos (ϕαβ ) = − nαβ ( t ) .eα ( t ) where e ( t ) =
(2.6.11)
vα ( t )

Setting λα < 1 creates the aforementioned anisotropic character, with a pedestrian directly

infront having the greatest effect as ϕ = 0 so Fαβ = 1 and pedestrians in the range

0 < ϕ < π2  32π < ϕ < 2π have the strongest influence. Figure 6 shows this for various λ

Figure 6: Fαβ characterises anisotropic behaviour

26
2.6.3 Boundary Interactions
The treatment of boundaries within the model is analogous to that of other pedestrians,

excluding the anisotropic effects of the pedestrian interactions. This leads to

§ r − dα B ·

fα B ( rα ) = AαΒ exp ¨ α ¸ nα B (2.6.12)
© Bα B ¹


Where dα B is the distance between the boundary and the pedestrian and nα B the normal

vector pointing from the boundary to α . In most situations there is more than one

boundary to consider leading to a question as to which boundaries influence to account

for. There are three possible ways of considering the boundary interaction

1. Superposition: All boundaries influence the pedestrian so the forces are summed

2. Shortest distance: Only the closet boundary element is considered

3. Biggest impact: only the boundary with the largest impact is considered

In most geometries the biggest impact and shortest distance model are equivalent; this is

often an appropriate choice too however this may not however be reasonable for angled

passageways and superposition may be better. It was decided that in this project only the

nearest boundary element would be considered, again this was the approach of [2] and

produced realistic results in that case.

2.6.4 Attractive Interactions

Often pedestrians demonstrate certain joining characteristics, such as families or groups

of tourists, and will wish to move through the walkway together. Other instances may be

shops, window displays or performances in the street. These two cases are however

separate, in the former the attractive force is constant and independent of time reflecting

the desire of these groups to remain together over the whole time interval. In the latter

27
case however the attraction is time dependent as the pedestrian will not wish to be late

and clearly will ignore the attraction if this is the case. Such attractions have a similar

modelling to pedestrian interactions however the attraction range Bα i is typically larger

with a smaller, negative, time dependent interaction strength Aα i .

2.6.5 Individuality and Random Behaviour Fluctuations

As previously mentioned, each pedestrian may display some random behaviour arising

from accidental or deliberate changes to the expected and optimal actions. This force


ξα ( t ) is Gaussian distributed and perpendicular to the desired direction. One such

formulation would be

ξα = eα ( t ) , fα ( t ) Xeαperp (2.6.13)

Here X ∝ N ( 0, σ 2 ) with probability density

1 1 § − x2 ·
f ( x) = . exp ¨ 2 ¸ (2.6.14)
2π σ © 2σ ¹

2.6.6 Another Formulation

As previously mentioned, there is another formulation of the Social Force model which

takes into account the ‘sliding friction force’ in the pedestrian and boundary interaction

terms. For the pedestrian interactions the model becomes

­° § rαβ − dαβ · ½°


¸ .Fαβ + k .g ( rαβ − dαβ ) ¾ nαβ + κ .g ( rαβ − dαβ ) ∆vαβ tαβ (2.6.15)
t
fαβ = ® Aα exp ¨
¯° © Bα ¹ ¿°



Here tαβ = ( − nαβ
2 1
, nαβ ) is the tangential direction and ∆vαβt = ( v
β − v
α ) .tαβ the relative
tangential velocity. The values k , κ are given constants of the system and the function

28
g ( x ) is zero when its argument is negative and equal to its argument otherwise. In this

case

­°0 if dαβ > rαβ

g ( rαβ − dαβ ) = ® (2.6.16)
°̄rαβ − dαβ else

Physically this means that the ‘sliding friction force’ is only felt if the pedestrians are

touching. Again boundary interactions are analogous, so

­° § r − dα B · ½°




fα B = ® Aα B exp ¨ α ¸ + k .g ( rα − dα B ) ¾ nα B − κ .g ( rα − dα B ) ( vα .tα B ) t gaB (2.6.17)
°¯ © Bα B ¹ °¿

This alternative model will not be implemented (although in principle once the first

model is coded the changes required are not difficult) and is only included for

completeness. Similarly the attractive interactions and random fluctuations described

earlier will not be implemented either.

29
Chapter 3

Self-Organisation Phenomena in Pedestrian Flows

One of the main objectives of this thesis is to explore and simulate emergent behaviours

of pedestrian systems. If a pedestrian flow has certain conditions such phenomena are

often observable. Of particular interest are

• Lane formation and striping effects

• Oscillatory flows and clogging effects at doorways

• Shock waves in dense crowds

These self-organisation effects are patterns of behaviour that are not externally planned or

organised by an outside source such as traffic signals or behavioural conventions.

3.1 Intersecting Flows

Intersecting flows are commonplace in many pedestrian facilities and are nearly

unavoidable as alternatives, such as bridges, are costly and generally impractical. Of

particular interest are Scramble Crossings which have been observed to display striping

effects when they intersect. A paper by Dzubiella and Löwen [12] exploring the pattern

forming in such systems describe these stripes as density waves moving into the direction

of the sum of the directional vectors of both flows.

Figure 7: Striping phenomena in Intersecting flows

30
The stripe formation reduces the number of obstructing interactions and maximisation of

average pedestrian speeds.

3.2 Bottlenecks

Bottlenecks such as doorways are also common in pedestrian areas and the way

pedestrians interact at them is of interest. Often the flows through bottlenecks are

irregular and inefficient. If there is an opposite flow into a bottleneck it is often the case

that the flow of pedestrians is oscillating and unidirectional as opposed to the

bidirectional flow in an ordinary corridor without a bottleneck. Similar to lane formation,

which will be discussed in the next section, it is typical for groups to move through the

bottle neck since it is easier to follow someone than to move against them. This leads to a

pressure increase, due in part to impatience, on the side of the bottleneck where flow is

halted and a decrease on the side that is flowing. When the difference between the two

pressures is large enough the flow will change. Thus the process is reversed and the flows

will oscillate.

3.3 Lane Formation

A bidirectional system, or counterflow, in everyday conditions has pedestrians moving in

opposite directions which are unevenly distributed over the walkway. One of the most

commonly observed phenomena is lane formation where pedestrians form into lanes of

unidirectional flow. This effect leads to a reduction in the number of evasive manoeuvres

each pedestrian need to make and so increases the efficiency of the walkway. In this case,

improved efficiency means that the average pedestrian speed is maximal and necessary

avoidance and braking actions minimal. The number of lanes formed is dependent on the

31
geometry of the walkway. Initially the lanes are small channels, created when opposite

moving pedestrians meet, which then merge with each other to form larger lanes.

3.4 Shockwaves

When the flow of pedestrians is slowed, the shockwave effect can be observed. The

velocity of each pedestrian, as described by (2.6.6) for the Social Force model, is

dependent upon their average speed. Physically this corresponds to the walker becoming

impatient and the desire to increase their velocity becomes stronger than the desire to

keep a certain distance from the pedestrian infront. This effect is cumulative and when

one pedestrian moves forward those behind him also move and thus the shockwave effect

occurs.

Figure 8: illustration of pedestrian behaviour causing shockwaves

32
Chapter 4

Implementation of the Social Force Model

One of the key aims of this project was to successfully implement the Social Force

Model, as described by Helbing [1], and attempt to do some quantitative analysis of the

results along with reproducing the qualitative behaviours of large pedestrian flows as

described in chapter 3. In this chapter the implementation of this model will be explored

with particular emphasis given to detailing some of the problems observed with the

model and how they were overcome.

4.1 Numerical Solution of the Social Force Model

The Social Force Model described in chapter 2 is a first order Ordinary Differential

Equation (ODE) which needs to be solved numerically to calculate the velocity and

position of the pedestrians within the system. Importantly the model is an initial value

problem, for both the velocity and displacement equations, which means that there are

several excellent MATLAB solvers available for this type of problem. The ODE

described by the Social Force model with initial conditions

t
rα ( 0 ) = ( rα1 ( 0 ) , rα2 ( 0 ) )
(4.1.1)

t
vα ( 0 ) = ( vα1 ( 0 ) , vα2 ( 0 ) )

Is a non-stiff system, as all the elements evolve on a similar time-scale. As this is the case

the suggested solver initially was ode45 which is extremely efficient for these types of

problem. This solver implements a Runge Kutta method of the fourth order and was

developed by Dormand and Prince. However, an unspotted error in the initial code

33
actually meant that it appeared that the system was stiff and so a different solver was

chosen, namely ode15s which is a Gear solver and generally very efficient for stiff

systems. This error was subsequently spotted in the de-bugging procedure and obviously

raised the question as to whether ode15s was an appropriate solver for the system. To

choose the best solver some tests were run to calculate the time each one was taking to

simulate an intersecting pedestrian flow. Each solver was tested with systems of 10 to

100 pedestrians in increments of 10 and the MATLAB function cputime was used to

calculate the time taken.

Plot of the cputime of several MATLAB solvers

500
ode23
ode23t
450 ode45
ode15s

400

350

300
cputime

250

200

150

100

50

0
10 20 30 40 50 60 70 80 90 100
Number of pestrians in the System

Figure 9: Cputime for several solvers – clearly ode23 and ode45 are the best solvers for
the Social Force Model

Whilst the stiff solver produced accurate results, they were no more accurate than the

non-stiff solvers. As such in the final implementation ode45 was chosen. The efficiency

of the system is of paramount importance as the Social Force model is of n 2 complexity.

34
This complexity comes from the pedestrian interaction equation (2.6.8), where the effect

of every other individual in the system upon α must be calculated and then summed.

This then has to be done for each pedestrian leading to O ( n 2 ) calculations, this

corresponds to a double for loop in the function file. This algorithm can infact be

reduced to an order n system, however this is not straightforward and time constraints

prevent its implementation. The first step of this reduction however was included in the

final code, which is to use a Verlet-Sphere [13]. For each pedestrian a circle, of radius r,

surrounding them is introduced with r being the radius of the interaction length of the

social force model. This radius is the maximum distance between any two pedestrian for

which their social force interaction, as defined by (2.6.8), will be calculated. This

represents physically the idea that pedestrians beyond a certain distance away will not

effect the decision making of the individual α , which is both intuitively and

mathematically reasonable, as (2.6.8) becomes negligible for r = rαβ − dαβ  O (10 ) .

Figure 10: Verlet - Sphere

35
The final code for the simulation solves 4n differential equations, where n is the number

of pedestrians in the system. To solve, ode15s requires a vector input with each

component corresponding to a differential equation. The first 2n differential equations are

the x and y velocity components for pedestrians 1 to n and the second 2n differential

equations the x and y displacement values for pedestrians 1 to n. That is

d ªvα ( t ) º ª fα ( t ) º
1 1

« »=« » for α = 1,..., n (4.1.2)

dt «¬vα2 ( t ) »¼ «¬ fα2 ( t ) »¼

and

d ª rα ( t ) º ª vα ( t ) º
1 1

« »=« » for α = 1,..., n (4.1.3)

dt «¬ rα2 ( t ) »¼ «¬ vα2 ( t ) »¼

The following pseudo-code describes how the differential equations are formulated in the

function file that ode15s solves.

FOR every pedestrian in the system do

Calculate desired velocity according to (2.6.6)
Calculate desired destination
Calculate desired direction according to (2.6.5)
Calculate driving force (2.6.4)
FOR every other pedestrian do
Check distance between pedestrians
IF this distance < radius of Verlet-Sphere
Calculate Social Force (2.6.8)
Add Social Force to driving force
FOR each boundary element do
Check distance to element
Calculate the influence of the closest element from (2.6.12)
Add to the Social and driving force
END

Set first 2n vector components as (4.1.2)

Set second 2n vector components as (4.1.3)

The simulation parameters are set outside of the function file in the associated script file.

These parameters include the constants of each term, the desired destination, initial

36
position and velocities and the boundary array which describes the geometry of the room

being considered.

4.2 The Desired Destination and Waypoints

In the earlier version of the code, the desired destination was set as a single point which

the pedestrian would head for. Whilst setting the desired destination in this way did not

prevent the simulations from displaying the emergent characteristics expected, it leads to

simulated behaviour which was qualitatively unreasonable. A good example to illuminate

this would be two pedestrians moving in opposite directions meeting in a corridor.

Figure 11: Pedestrian paths with a single point for p , desired destination

In figure 11, the behaviour of pedestrians with fixed desired destinations is show.

However, when walking down a corridor the destination is only the opposite end and not

a fixed point. Thus the expected behaviour would be for the pedestrian to take the

shortest path to the end of the corridor and as such they would re-evaluate their desired

destination in the event of interaction with another walker. To implement this behaviour

the desired destination was re-evaluated with each iteration. This was achieved setting a


desired destination array and then calculating the closest point of this and using that as p .

Figure 12: Pedestrian paths using a desired destination array

37
In Figure 12 the desired destination arrays are shown at each end of the corridor and the

new pedestrian paths. In this small test case the difference may seem to be unimportant

but its value in the modelling come as the number of pedestrians in the system increase. If

the desired destination remained as a single point then individuals maybe simulated

trying to force their way through a dense crowd to reach the specified point instead of

taking the more efficient route. The principle at the core of all pedestrian modelling is

that individuals will try to use the most efficient route for themselves and that the model

must reflect the decision making abilities of humans and their capability to change and

adapt to the system around them.

Another problem found in earlier version of the code was so called Hunting or Chattering

effects. This phenomenon occurs when the direct line between the pedestrian and their

desired destination is blocked by the boundary of the setup being simulated. If the model

isn’t modified to counter this effect the walker will head directly towards the desired

destination until they reach the boundary. The repulsive interaction with the boundary is

(obviously) stronger than the driving force and the pedestrian will become stuck on the

wall. Clearly this behaviour is not consistent with observed pedestrian movement and so

a solution must be found.

Figure 13: Pedestrians path with and without a waypoint

38
The solution implement in here is to use a waypoint system. In Figure 13 the blue dashed

line illustrates the desired path of the pedestrian in the original code and the red line the

desired path after a waypoint is included. The process of implementing waypoints is

relatively straight forward, the pseudo-code is as follows:

FOR each pedestrian do

Check the distance from current position to waypoint (i)
IF the distance to waypoint (i) < specified distance
Set desired destination as the waypoint (i + 1)
ELSE
Set desired destination as waypoint (i)
END
END

Again it is important to note that waypoints are intuitively consistent with pedestrian

behaviour, in particular a pedestrians immediate destination is always within their line of

sight. That is, an individual will always travel to a place they can see until their ultimate

destination is in view. In the case of angled passage ways, corners will be appropriate

waypoints since moving directly toward them will provide the shortest path to the overall

goal.

4.3 The Desired Velocity

One of the self-organisation phenomena of interest are shockwaves in dense crowds. This

effect comes from an impatience factor of the modelled pedestrians. This is described in

the formulation of the desired velocity, (2.6.6). The only extra calculation required is the

average velocity, which is

rα ( t ) − rα ( 0 )
vα = (4.3.1)
t

39
In earlier versions of the code the desired velocity was fixed, the following graph shows

the acceleration of an unobstructed pedestrian with both fixed and variable desired

velocities.

1.5
Fixed Desired Velocity
Variable Desired Velocity

1
Speed

0.5

0
0 1 2 3 4 5 6 7 8 9 10
Time

Figure 14: The difference between a fixed desired velocity and a variable one

As Figure 14 shows, the variable desired velocity means that the pedestrian will want to

travel faster if they are impeded. The graph shows that this effect means the pedestrian

will move slightly faster than the initial desired velocity before settling into a comfortable

walking speed. In the case of the fixed desired velocity the walker reaches the desired

velocity smoothly and without going faster than this value. Certainly the variable desired

velocity describes shockwave effects well but this effect in the acceleration phase is also

reasonable.

40
Chapter 5

Simulation and Analysis of the Social Force Model

In this chapter the code created to implement the Social Force Model will be run for a

variety of different pedestrian flows. These will be a counterflow within a corridor, with

and without a bottleneck, a ninety degree corner and finally a Scramble Crossing. The

aim will be to use the data from the simulation to do some quantitative analysis of

pedestrian flows to supplement the qualitative observations laid out in chapter 3.

Obviously it is expected that the model will recreate self organisation phenomena and

where this is the case it will be highlighted.

5.1 The Model Constants

The Social Force Model has several parameters in each of the driving force (2.6.4),

pedestrian interaction (2.6.8), and boundary interaction terms (2.6.12). Thus, several

parameters must be set in the startup phase of the simulation, in this instance in the script

file for the model.

The driving force parameters are consistent with those of Helbing [1] where the desired

velocity is set to 1.34 ms-1 with a relaxation time, τ α , of 0.5. The relaxation time is the

time taken to correct disturbance in movement (e.g. obstacles or avoidance manoeuvres).

Helbing also gives values for the parameters in the pedestrian and boundary interaction

terms. To verify that these values were reasonable the simulation was run for a variety of

parameters in a test room and the resulting pedestrian flows evaluated.

41
 10

0
0 2 4 6 8 10 12 14 16 18 20

Figure 15: An example of pedestrian flows in the test room (Not to Scale)

Figure 15 shows one of the test runs for a given set of parameters, each line in the figure

represents the movement of a pedestrian over time. The test room was simulated with 50

pedestrians and qualitative observations were made. An example of such an observation

is that for large values of Bα B in (2.6.12) the pedestrians appeared to be too far from the

boundaries given the amount of traffic in the room. To test the interaction strengths a

simple test was to set the desired velocity of a pedestrian to 0 and place them next to a

boundary. The repulsive force of the boundary for different parameter values could then

be more accurately observed providing a better reference for deciding on realistic

parameters. A similar process was done for pedestrian interactions. Whilst there was only

a qualitative analysis at this stage it allowed for a set of realistic parameters to be chosen,

where the conventions for pedestrian behaviour come from the authors’ personal

42
experience and the experiments done by Helbing [1]. The following table gives the

chosen parameters

Parameter Formula Value

Initial Desired Velocity -
(2.6.6) 1.34 ms-1
vα0 ( 0 )
Relaxation Time - τ α (2.6.4) 0.5 s
Maximum Speed - vα max
(2.6.6) 1.3* vα0 ( 0 )
Territorial Sphere Pedestrian
(2.6.8) 0
Interaction Strength - Aα1
Territorial Sphere Pedestrian
(2.6.8) 0.3 m
Interaction Range - Bα1
Anisotropic Character - λα (2.6.10) 0.75
Physical Pedestrian
(2.6.8) 2
Interaction Strength - Aα2
Physical Pedestrian
(2.6.8) 0.2 m
Interaction Range - Bα2
Boundary Interaction Strength
(2.6.12) 5
- Aα B
Boundary Interaction Range
(2.6.12) 0.1 m
- Bα B
Radius of Pedestrians – rα (2.6.8), (2.6.12) 0.3 m

Table 1: Simulation Parameters

In the simulation it was chosen to take Aα1 = 0 as suggested by Helbing [1]. The purpose

of this is to speed up the simulation in dense crowds, which is appropriate for the

situations being considered. In all the cases that will be discussed in this thesis (in

particular the Scramble Crossing) the emergent behaviour is most evident in densely

populated situations. Since this term describes the tendency to respect each pedestrians

territorial sphere these considerations can be ignored in dense crowds and emergency

situations.

43
One final consideration in the startup phase is the size of the so-called Verlet-Sphere

implemented in (2.6.8) discussed in chapter 4. The radius was chosen to be 10 m,

anything further away than this has an overall contribution to the social force felt by α to

be ≥ O (10−15 ) which is negligible since the cumulative social force of all pedestrians

upon α is of O (1) .

5.2 Pedestrian Counterflows

In this section the code developed will be used to simulate pedestrian counterflows both

with and without a bottleneck in the walkway. The purpose of modelling these situations

is to assess the validity of the simulation in a qualitative sense. In chapter 3 certain

emergent phenomena were discussed; in particular lane formation in a counterflow and

oscillatory flow at bottleneck. For these particular geometries a more detailed analysis of

the pedestrian movements will not be considered, rather this section aims to confirm that

the model simulates well experimentally observed phenomena ([1], [2]).

In a counterflow expected emergent behaviour is lane formation as discussed in chapter

3. The simulation was run for 250 pedestrians on a 200 meter walkway which is 6 meters

wide. Half the pedestrians were randomly placed at each end of the walkway with a

bivariate uniform distribution using the rand function in MATLAB. The wavefront of

each the pedestrian bodies (in this case the group at either end) meet at the middle of the

walkway. The simulation results were used to produce a movie (submitted on cd with the

thesis – the file ‘250conterflow.avi’) of the pedestrian movement.

44
 20
20

18
18

16
16

14 14

12 12

10 10

8 8

6 6

4 4

2 2

0 0
25 30 35 40 45 50 55 60 65 70 75 25 30 35 40 45 50 55 60 65 70 75

(a) The initial pedestrian distribution (b) The final pedestrian distribution

Figure 16: The initial and final pedestrian positions created for a counterflow

The results of the simulation are better than expected, reproducing the self organisation

phenomena perfectly. However these results underline the idealised nature of the

implemented model. In a real counterflow the pedestrians would not be homogeneous,

that is the desired speed of the walkers would not be uniform and the pedestrians radii

would not be the same. These inhomogeneities would lead to behaviour that isn’t

observed in this simulation, a particular example of this would be overtaking manoeuvres

performed by faster moving pedestrians. In walkways with large pedestrian densities

these manoeuvres can lead to a breakdown in the lane formation; this suggests that these

idealisations may not always be valid assumptions.

The next geometry to consider is a counterflow with a bottleneck. In this case the

bottleneck refers to a doorway although in practice it could be a longer bottleneck (c.f.

the test room geometry). The bottleneck, a 1 meter wide doorway, is positioned in the

middle of the walkway which is 3 meters wide. 30 pedestrians were placed randomly,

again with a bivariate uniform distribution, with half at each end of the corridor. Again a

movie (on the cd – ‘100corridor.avi’) of the simulation was produce – although it must be

45
stressed it is not in real time. This is due to the fact that ode45 is a variable timestep

solver, which means that the time between each frame varies. However the movie

creating tool in MATLAB has a fixed frame rate leading to the movies not being in real

time.

10 10

9 9

8 8

7 7

6 6

5 5

4 4

3 3

2 2

1 1

0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20

(a) Initial Pedestrian Positions (b) Pedestrians meet at Bottleneck

with right moving (red) pedestrians
flowing through
10 10

9 9

8 8

7 7

6 6

5 5

4 4

3 3

2 2

1 1

0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20

(c) Switch in flow – now the left (d) The flow has switched several
moving (blue) pedestrians are times, now the right moving
flowing through pedestrians are passing bottleneck

Figure 17: These figures show how flow oscillates at a bottleneck

The figure illustrates the oscillatory flow through a bottleneck is reproduced by the

simulation. This oscillatory behaviour is much clearer in the associated movie. Again the

results of the simulation are consistent with the experimental data of Helbing [1].

46
5.3 A Ninety Degree Corner
Having shown in two separate cases that the model is qualitatively accurate, the next

stage is to do some quantitative analysis of the results for different geometries. The first

situation to be modelled is a crowded corner in a corridor. A particular case where this

geometry is of specific interest would be a modern sports stadium. There are often very

large crowds which may need to evacuate the facility quickly and ninety degree corners

are commonly found along escape routes. Thus the effect that they have upon pedestrian

flow is of great interest.

The first thing to explore is how the corner effects the movement and speed of a single

pedestrian. Obviously if a ninety degree corner has an effect on the large scale behaviour

of pedestrian flows then it is reasonable to expect that there will be an observable change

in the individual walkers movement at the corner.

The speed of a single pedestrian through a ninety degree corner

1.5

1
Speed

0.5

0
0 2 4 6 8 10 12 14 16 18 20
Time

Figure 18: The Speed of a single Pedestrian around a ninety degree corner

47
Figure 18 clearly shows that the change of direction associated with moving around a

corner leads to a sudden drop in speed as the walker reaches it. This highlights the

problem with ninety degree corners in pedestrian facilities, that is that rather than

allowing for a smooth change of direction and thus maintaining a steady speed it forces

the pedestrians to change direction abruptly which has concertina effects on the following

traffic. The figure also shows how the pedestrian speeds up after passing the corner to

compensate for the obstruction. This effect results in the pedestrian density (pedestrians

per unit area) being lower after the corner if the crowd is dense enough.

The next step is to model a large flow and analyse these concertina effects. As previously

mentioned, the average pedestrian density is reduced after the corner and is greatest just

before the corner, corresponding to the shockwave phenomena described in section 3.4.

10

0
0 1 2 3 4 5 6 7 8 9 10

Figure 19: The simulation of a ninety degree corner

48
Figure 19 shows the simulation for only 200 pedestrians moving around a 3 meter wide

corner, not a particularly large number when compared to what may be expected in a

large sports stadium and already the corner is slowing traffic. This is highlighted in

Figure 20, which shows that the overall mean speed of the pedestrians is reduced as the

first walkers come to the corner. However, as more of the pedestrians pass the corner the

overall mean speed starts to increase, reflecting the compensatory increase in speed of

each individual as they pass the corner due to the change in their desired velocity

described by (2.6.6).

1.4

1.2

1
Overall Mean Speed

0.8

0.6

0.4

0.2

0
0 2 4 6 8 10 12 14 16 18 20
Time

Figure 20: The overall mean speed of pedestrians around a ninety degree corner

The average densities were calculated for the 3 distinct phases of the corridor – the

unidirectional flow before the corner, the corner itself and the corridor after it. The

average pedestrian density before and after the corner are similar here, at approximately

49
1.5 pedestrians per square meter, however on the corner this increases to approximately

2.7 pedestrians per square meter, if the 3 empty cells on thee outside of the corner are

excluded. It is not unexpected that these 3 cells are empty; this is due to the fact that the

natural instinct of a walker will be to head for the apex of the corner, as this will provide

the shortest route to their desired destination. This reasoning also goes some way toward

explaining why the corner itself is such a hindrance to traffic flow. Clearly all the

pedestrians are seeking the shortest route and so all will want to pass as close to the apex

of the corner as they can. This in turn will lead to increased densities in the cells next to

the apex and explain the zero pedestrian density at the far edge of the corner.

In this case the densities before and after the corner is essentially the same; but, as figures

18 and 20 show, the speed of the walkers has increased after the corner. This quantitative

effect is a good foundation for explaining why corners can cause clogging in dense

crowds. Unfortunately, due to the n2 complexity of the code, it wasn’t possible to run the

simulation for higher numbers of pedestrians. As discussed, for larger crowds, it would

be expected that the corner would clog more noticeably leading to a significant reduction

in pedestrian density after it. It should also be noted that although some shockwaving has

occurred at the corner, this effect is not as prominent as the experimental data suggests it

should be and again this may simply be because the current crowd isn’t large enough. At

this point it is worth mentioning again the homogeneity of the pedestrians in the current

simulations, in particular the uniform desired speed. This uniformity is both unrealistic

and means that no overtaking manoeuvres will occur. The lack of certain expected

behaviours in the simulation therefore may, in part, be explained by this.

50
5.4 The Scramble Crossing
The final pedestrian flow to consider is the Scramble Crossing. Scramble Crossing is the

term used to describe the meeting of two unidirectional flows with a non-zero angle of

intersection (if the angle of intersection is zero this simply describes a bidirectional

counterflow). Such intersecting flows are commonplace in pedestrian walkways and so

the dynamics of them is of great interest. Only the simplest case, where there are no

boundaries for the pedestrians to interact with, will be simulated in this project as the

effect of particular interest is the stripping phenomena described in section 3.1.

A Scramble Crossing with 500 individuals was simulated with a ninety degree angle of

intersection. The first quantitative data to consider is the velocities of the pedestrians in

the simulation. It is expected that the pedestrians will interact in a manner that leads to

the least deviation from the desired velocities.

600 1.5

500
Overall Mean Speed

400 1
Frequency

300

200 0.5

100

0 0
0.6 0.8 1 1.2 1.4 1.6 1.8 0 5 10 15 20 25 30 35 40 45
Speed m/s Time

(a) A Histogram of the speed of 10 (b) The overall mean speed of

random walkers pedestrians in a Scramble Crossing
Figure 21: Plots of the speed characteristics of a Scramble crossing

The plots in figure 21 illustrate well that the simulated interactions within a scramble

crossing are optimal. In Figure 21(b) the drop in overall mean speed at about 15 seconds

is when the two pedestrian wavefronts meet. This drop in the overall mean speed is

51
approximately 0.1 ms-1 which would is clearly not particularly significant. Obviously the

overall mean speed may hide the fact that movement in the area where the 2 flows meet is

very slow. Clearly only a small proportion of the total number of pedestrians will be in

the intersection at any one time and so the speed of the pedestrians outside of the point of

intersection could mask the fact that movement where the flows meet is slow. Figure

21(a) however dispels this; it is a histogram of the speeds at each timestep of a random

sample of 10 pedestrians where each pedestrian is represented by a different colour bar. It

clearly shows that the pedestrian speeds are tightly distributed around the initial desired

speed of 1.34 ms-1 indicating that even when a pedestrian is moving through the area of

intersection their actual speed does not deviate too far from their desired speed.

0.45
Variance
Standard Deviation
0.4

0.35

0.3

0.25
m/s

0.2

0.15

0.1

0.05

0
0 5 10 15 20 25 30 35 40 45
Time

Figure 22: Variance and Standard Deviation of Pedestrian in a Scramble Crossing

52
Figure 22 shows both the variance and standard deviation of the pedestrian speeds in the

simulated Scramble Crossing. These are perhaps better statistical measures of the

pedestrian behaviour and are consistent with the previous conclusions. The mean

standard deviation, after the two flows meet (at 16 seconds into the simulation), is 0.1888

which shows that the pedestrian speeds do not vary wildly when the flows meet

supporting the conclusions drawn from the plots in Figure 21.

The most interesting aspect of the Scramble Crossing is the observed stripping

phenomena. The simulation does reproduce this effect although it is not as obvious the

lane formation of a counterflow shown in Figure 16 or the illustration of the effect in

Figure 7. As this is the case there is little point showing still frames of the simulation. The

effect is much more evident from the movies (submitted on CD with the thesis –

‘500scramblecrossing.avi’) and certainly the movie shows that the simulation recreates

the expect phenomena qualitatively. One statistical measure to show this stripping

phenomena is occurring is the computation of Lattice Constants, l1 and l2 . These are

described by Ripley [14] in chapter 2. In the simplest case, where the standard basis is

used, l1 is the average distance between

C D

pedestrians in the x direction and l2 the y

E
direction. These Lattice Constants were
A B
calculated for 5 different regions, the 2 flows

Figure 23: The regions the Lattice before the intersection (A and B) and the 2
Constants are calculated in
flows after intersection (C and D) and the area

of intersection itself (E). The Lattice constants were taken at t = 40 so a sufficient number

53
of pedestrian had managed to pass the intersection of the flows. The Lattice constants are

summarised in Table 2:

Region l1 l2
A 0.5486 0.5491
B 0.5516 0.5521
C 0.6132 1.0236
D 0.5962 0.9742
E 0.3212 0.3195

Table 2: Lattice Constants

Initially the pedestrians were uniformly distributed and given the homogeneity of the

walkers in this simulation it is to be expected that the lattice constants in regions A and B

are very close to each other. In region E both the Lattice Constants are close to 0.3 which

is the pedestrian radius defined for this simulation (See Table 1) describing the high

density in the area of intersection. The interesting values are produced in regions C and

D. In these area, l1 , the average separation in the x direction, is close to the initial values

but l2 , the average separation in the y direction has doubled. This change, whilst visually

obvious, provides a quantitative value highlighting the stripping effect.

5.5 Model Shortcomings

Having simulated a variety of pedestrian flows, the results obtained have highlighted a

number of areas in which the model could be improved. The most obvious of these,

which has already being touched upon, is the homogeneity of the system. Currently the

pedestrians are modelled as uniform, which may not be a reasonable assumption and may

lead to important effects not being simulated. The solution would be to obtain further

information on the makeup of pedestrian crowds from experimental data. In the case of

54
the desired velocity it would be a case of setting it a randomly distributed value, probably

a Normal distribution with a mean value the same as the current constant desired velocity

and a variance of 0.3 say, that is vα0 ( 0 ) ~ N (1.34,0.3) . A similar approach could be

adopted with the pedestrian radii.

Another problem that arises in the current implementation are shadowing effects. These

occur when there is a solid wall between two pedestrian but the distance between them is

less than the radius of the Verlet-Sphere. Even though the pedestrians can’t see each other

the current code still calculates the social force between them and adds it to the

acceleration. Clearly this is unrealistic given the fact that there is no line of sight between

them and thus no way either can know the other is there. One way to solve this problem

would be to adjust the Verlet-Sphere for each pedestrian. This re-evaluation is shown

below

Figure 24: The New Verlet-Sphere

Only Pedestrians in the shaded area would be considered in the model, eliminating the

shadowing effects.

55
Chapter 6

Summary and Future Directions

The aim of this project was to implement a version of the Social Force Model and use it

to simulate various pedestrian flows. The results produced by the simulation with this

implementation provided some excellent qualitative results, reproducing the expected

self-organisation phenomena very well. These included oscillating flows at bottle necks,

lane formation in a counter flow and stripping in a Scramble Crossing. However in the

final chapter several flaws in the current implementation were identified though solution

to each of them was suggested.

If this project were to be continued in future there are several areas that should be looked

at. The first of these would be to include the additional pedestrian behaviours set out in

section 2.6 such as the random fluctuations described by (2.6.13) and the joining

behaviour discussed in section 2.6.4.

Currently the simulation requires order n2 calculations at each iteration of the ODE

solver. There are techniques to reduce this to an order n algorithm which would allow the

simulation to be run for much larger systems on a single processor. Even if an order n

algorithm is achieved it would be useful to parallelise the implementation so that the

simulation could be extended to huge scale problems such as airport terminals or world

sporting events, where there may be as many as 150,000 pedestrians present.

56
With the order n algorithm in place it would then be possible to explore the effect of

vastly increased density upon the systems already investigated. Certainly in the case of

the ninety degree corner it would be valuable to find a quantitative value for the critical

density of the geometry. The effect of the flow reaching its critical density is of

paramount importance in emergency situations as this is when such a density may be

reached.

Further investigation should focus primarily on the effect of boundaries on pedestrian

flows and try to develop solutions to flaws in current building design. This is simply

because the most interesting pedestrian flows occur when boundaries are present and that

in emergency situations it is the boundaries that become the biggest hazards when they

restrict pedestrian flow.

57
Bibliography
[1] D. Helbing et al. Self-Organized Pedestrian Crowd Dynamics: Experiments,
Simulations, and Design Solutions. Transportation Science, 39(1):1-24, February 2005

[2] D. Helbing and P. Molnár. Social force Model for Pedestrian Dynamics. Physical
Review E, 51(5):4282-4286, May 1995

[3] V. J. Blue and J. L. Adler. Bi-directional Emergent Fundamental Pedestrian Flows

from Cellular Automata Microsimulation. Proceedings of the 14thInternational
Symposium of Transportation and Traffic Theory, 1999

[4] L. F. Henderson. The Statistics of Crowd Fluids, Nature, 229:381-383, February 1971

[5] S. Hoogendoorn and P. H. L. Bovy. Gas-Kinetic Modeling and Simulation of

Pedestrian Flows. Transportation Research Record, 1710:28-36, 2000.

[6] S. Okazaki and S. Matsushita. A Study of Simulation Model for Pedestrian Movement
with Evacuation and Queuing. Proceedings of the International Conference on
Engineering for Crowd Safety, pages 271-280, 1993

[7] S. J. Yuhaski, Jr and J. Macgregor Smith. Modeling Circulation Systems in Buildings

Using State Dependent Queuing Models. Queueing Systems, pages 319-338,1989

[8] G. G. Løvas. Modeling and Simulation of Pedestrian Traffic flow. Transportation

Research B, 28(6):429-443, 1994

[9] V. J. Blue and J. L. Adler. Cellular Automata Microsimulation of Bi-Directional

Pedestrian Flows, Transportation Research Record, 1678:135-141, 2000

[10] J. Dijkstra, J. Jessurun and H. Timmermans. A Multi-Agent Cellular Automata

Model of Pedestrian Movement. Pedestrian and Evacuation Dynamics, pages 173-181,
2001

[11] D. Helbing and P. Molnár. Self-Organization Phenomena In Pedestrian Crowds.

Self-Organization of Complex Structures: From Individual to Collective Dynamics, pages
569-577, 1997

[12] J. Dzubiella and H. Löwen. Pattern Formation in Driven Colloidal Mixtures: Tilted
Driving Forces and Re-Entrant Crystal Freezing. J. Physics: Condensed Matter 14:9383-
9395, 2002

[13] L. Verlet. Computer “Experiments” on Classical Fluids. I. Thermodynamical

Properties of Lennard-Jones Molecules. Physical Review, 159(1):98-103

[14] B. D. Ripley. Stochastic Simulation John Wiley & Sons, Inc, 1987

58
Appendix A

MATLAB Codes

A.1 SFMDFPIBI.m
% Driving Force, Pedestrian interactions and boundary interactions %
% dv/dt = f0a + fab + faB %
%------------------------------------------------------------------------%

clear all
global m bdry p p1 p2 D way BDRY
global tau v0a0 vmax r A1a A2a B1a B2a lambda A1 B1 VS dist

% Number of Pedestrians%
m = input('Enter the number of Pedestrians in the system (must be even):');
% Timespan of simulation %
T = input('Enter the desired length of the simulation (in seconds): ');
tspan = [0 T];
% Select the geometry to be simulated %
disp('The code has 5 geometries to choose from, they are:')
disp('1. A counterflow')
disp('2. A Counterflow with a doorway (bottleneck)')
disp('3. The test room')
disp('4. A Scramble Crossing')
disp('5. A ninety degree corner')
K = input('Enter the number of the room you wish to simulate: ');
%----------------------------------------------------------%
% Vector used in waypoint calculation - their value tells %
% pedestrian alpha which waypoint to head for %
%----------------------------------------------------------%
D = zeros(1,m);
%-----------------%
% model constants %
%-----------------%

% Relaxation time %
tau = 0.5;
% Initial desired velocity %
v0a0 = 1.34;
% Maximum desired velocity %
vmax = 1.3*v0a0;
% Pedestrian radii %
r = 0.3*ones(1,m);
% Pedestrian Interaction constants from (2.6.8) %
A1a = 0;
A2a = 2;
B1a = 0.3;
B2a = 0.2;
lambda = 0.75;
% Boundary Interaction constants from (2.6.12) %
A1 = 5;

59
B1 = 0.1;
% Radius of Verlet Sphere %
VS = 10;
% Distance to waypoint at which it deem to have been reached %
dist = 1.2;
%--------------------------------------%
% Initial position and velocity vector %
%--------------------------------------%

if (K == 1)
way = 1;
Vx = [1.34*rand(1,m/2),-1.34*rand(1,m/2)];
Vy = zeros(1,m);
X = [9*rand(1,m/2),9*rand(1,m/2)+11];
Y = [2*rand(1,m)+3.5];
elseif (K == 2)
way = 0;
Vx = zeros(1,m);
Vy = zeros(1,m);
X = [100*rand(1,m/2)-50,100*rand(1,m/2)+50];
Y = [6*rand(1,m)+7.5];
elseif (K == 3)
way = 1;
Vx = zeros(1,m);
Vy = zeros(1,m);
X = [5*rand(1,m/2),5*rand(1,m/2)+15];
Y = [10*rand(1,m)];
elseif (K == 4)
Vx = zeros(1,m);
Vy = zeros(1,m);
X = [45*rand(1,m/2)-40,45*rand(1,m/2)+35];
for i = 1:m
if (i <= m/2)
Y(i) = X(i) - 10*rand(1,1) + 5;
elseif (i > m/2)
Y(i) = (35 - X(i)) + 10*rand(1,1);
end
end
elseif (K == 5)
way = 1;
Vx = zeros(1,m);
Vy = zeros(1,m);
X = [35*rand(1,m)-30];
Y = [2*rand(1,m)+0.5];
end
% Setting u - the initial position and velocity vector %
for i = 1:m
u((2*i)-1) = Vx(i);
u((2*i)) = Vy(i);
end
for i = 1:m
u((2*m)+(2*i)-1) = X(i);
u((2*m)+(2*i)) = Y(i);
end
start = u;
%----------------------------------------------------------------------%
% Desired destination - p is initially set as a waypoint (if required) %
%----------------------------------------------------------------------%

if (K == 1)

60
 p = [[10.5*ones(1,m/2),9.5*ones(1,m/2)]',[4.5*ones(1,m)]'];
p1 = [[50*ones(1,(2/0.1)+1)]',[3.5:0.1:5.5]'];
p2 = [[-30*ones(1,(2/0.1)+1)]',[3.5:0.1:5.5]'];
elseif (K == 2)
p = zeros(m,2);
p1 = [[100*ones(1,(6/0.1)+1)]',[7:0.1:13]'];
p2 = [[zeros(1,(6/0.1)+1)]',[7:0.1:13]'];
elseif (K == 3)
p = [[5*ones(1,m/2),15*ones(1,m/2)]',[5*ones(1,m)]'];
p1 = [[20*ones(1,(4/0.1)+1)]',[3:0.1:7]'];
p2 = [[zeros(1,(4/0.1)+1)]',[3:0.1:7]'];
elseif (K == 4)
p = zeros(m,2);
p1 = [[50:0.1:55]',[55:-0.1:50]'];
p2 = [[-15:0.1:-10]',[50:0.1:55]'];
elseif (K == 5)
p = [[8.5*ones(1,m)]',[2*ones(1,m)]'];
p1 = [[7.5:0.1:9.5]',[40*ones(1,(2/0.1)+1)]'];
p2 = [[7.5:0.1:9.5]',[40*ones(1,(2/0.1)+1)]'];
end
%----------------%
% Boundary Array %
%----------------%

if (K == 1)
bdry = [[0:0.1:20,0:0.1:20,10*ones(1,(1/0.1)+1),10*ones(1,(1/0.1)+1)...
]',[3*ones(1,(20/0.1)+1),6*ones(1,(20/0.1)+1),3:0.1:4,5:0.1:6]',];
BDRY = 1;
elseif (K == 2)
bdry = [[-50:0.1:150,-50:0.1:150]'...
,[6*ones(1,(200/0.1)+1),14*ones(1,(200/0.1)+1)]'];
BDRY = 1;
elseif (K == 3)
bdry = [[5:0.1:15,5:0.1:15,5*ones(1,(3/0.1)+1),5*ones(1,(3/0.1)+1)...
,15*ones(1,(3/0.1)+1),15*ones(1,(3/0.1)+1)]'...
,[3*ones(1,(10/0.1)+1),7*ones(1,(10/0.1)+1),0:0.1:3,7:0.1:10,...
0:0.1:3,7:0.1:10]'];
BDRY = 1;
elseif (K == 4)
BDRY = 0;
elseif (K == 5)
bdry = [[-30:0.1:10,-30:0.1:7,10*ones(1,(40/0.1)+1),...
7*ones(1,(37/0.1)+1)]',[zeros(1,(40/0.1)+1),3*ones(1,(37/0.1)+1)...
,0:0.1:40,3:0.1:40]'];
BDRY = 1;
end
%----------------------------------------%
% The solver for the system - here ode45 %
%----------------------------------------%

options1 = odeset('AbsTol',1d-3,'RelTol',1d-4);
[t1,u1] = ode45(@fun5,tspan,start,options1);
pause
%---------------------------------------%
% Plotting the paths of the pedestrians %
%---------------------------------------%

% Plotting the boundaries %

if (K == 4)
hold on

61
else
plot(bdry(:,1),bdry(:,2),'rx')
if (K == 1)||(K==3)
axis([0,20,0,10]);
elseif (K == 2)
axis([25,75,0,20]);
elseif (K == 5)
axis([0,10,0,10]);
end
hold on
end
% Plotting the pedestrian paths %
for i = (2*m)+1:2:(4*m)
if (i<=3*m)
plot(u1(:,i),u1(:,i+1),'b-')
if (K == 1)||(K == 3)
axis([0,20,0,10]);
elseif (K == 2)
axis([25,75,0,20])
elseif K == 4
axis([0,40,0,40]);
elseif K == 5
axis([0,10,0,10]);
end
hold on
else
plot(u1(:,i),u1(:,i+1),'r-')
if (K == 3)||(K ==1)
axis([0,20,0,10]);
elseif (K == 2)
axis([25,75,0,20])
elseif K == 4
axis([0,40,0,40]);
elseif K == 5
axis([0,10,0,10]);
end
hold on
end
end
pause
%-----------------------------------------------------------%
% Making the Movie - plots the pedestrian positions at each %
% timestep and then uses each plot as a movie frame %
%-----------------------------------------------------------%

for j = 1:length(t1)
% Plotting the boundaries %
if (K == 4)
hold off
else
hold off
plot(bdry(:,1),bdry(:,2),'rx')
if (K == 1)||(K == 3)
axis([0,20,0,10]);
elseif (K == 2)
axis([25,75,0,20]);
elseif (K == 5)
axis([0,10,0,10]);
end
hold on

62
 end
% Plotting the pedestrians %
for i = (2*m)+1:2:(4*m)
if (i<=3*m)
plot(u1(j,i),u1(j,i+1),'bo')
if (K == 1)||(K == 3)
axis([0,20,0,10]);
elseif (K == 2)
axis([25,75,0,20])
elseif K == 4
axis([0,40,0,40]);
elseif K == 5
axis([0,10,0,10]);
end
hold on
else
plot(u1(j,i),u1(j,i+1),'ro')
if (K == 3)||(K ==1)
axis([0,20,0,10]);
elseif (K == 2)
axis([25,75,20])
elseif K == 4
axis([0,40,0,40]);
elseif K == 5
axis([0,10,0,10]);
end
hold on
end
end
F(j) = getframe;
end

clf

A.2 fun5.m
% Driving force, pedestrian interaction and boundary interaction function %
%-------------------------------------------------------------------------%

function f0a = fun5(t,u)

global m bdry p p1 p2 D way BDRY x0
global tau v0a0 vmax r A1a A2a B1a B2a lambda A1 B1 VS dist

%---------------------------------------------------%
% Putting pedestrian positions in a separate vector %
%---------------------------------------------------%

for i = 1:m
x(i,1) = u((2*i)-1+(2*m));
x(i,2) = u((2*i)+(2*m));
end
%-------------------------------------------------------%
% Setting inital position for average speed calculation %
%-------------------------------------------------------%

if (t == 0)
x0 = x;
end

63
%-------------------------------%
% Desired Velocity from (2.6.6) %
%-------------------------------%
for i = 1:m
if t == 0
v0a(i) = v0a0;
else
V (i,:) = x(i,:) - x0(i,:);
Vbar = norm(V(i,:)/t);
na = 1 - (Vbar/v0a0);
v0a(i) = (1 - na)*v0a0 + na*vmax;
end
%-----------------------%
% Desired destination p %
%-----------------------%
if (way == 1)
% Pedestrians in area A (L-Hand corner for Scramble Crossing) %
if (i <= m/2)
nrmpx(i) = norm(p(i,:) - x(i,:));
if (nrmpx(i) < dist)
D(i) = 1;
end
if (D(i) == 1)
p(i,:) = p1(1,:);
nrmpx(i) = norm(p(i,:) - x(i,:));
for j = 1:length(p1)
if (norm(p1(j,:) - x(i,:)) < nrmpx(i))
p(i,:) = p1(j,:);
nrmpx(i) = norm(p(i,:) - x(i,:));
end
end
end
% Pedestrians in area B (R-Hand corner for Scramble Crossing) %
elseif (i > m/2)
nrmpx(i) = norm(p(i,:) - x(i,:));
if (nrmpx(i) < dist)
D(i) = 1;
end
if (D(i) == 1)
p(i,:) = p2(1,:);
nrmpx(i) = norm(p(i,:) - x(i,:));
for j = 1:length(p2)
if (norm(p2(j,:) - x(i,:)) < nrmpx(i))
p(i,:) = p2(j,:);
nrmpx(i) = norm(p(i,:) - x(i,:));
end
end
end
end
elseif (way == 0)
% Pedestrians in area A (L-Hand corner for Scramble Crossing) %
if (i <= m/2)
p(i,:) = p1(1,:);
nrmpx(i) = norm(p(i,:) - x(i,:));
for j = 1:length(p1)
if (norm(p1(j,:) - x(i,:)) < nrmpx(i))
p(i,:) = p1(j,:);
nrmpx(i) = norm(p(i,:) - x(i,:));
end

64
 end
% Pedestrians in area B (R-Hand corner for Scramble Crossing) %
elseif (i > m/2)
p(i,:) = p2(1,:);
nrmpx(i) = norm(p(i,:) - x(i,:));
for j = 1:length(p2)
if (norm(p2(j,:) - x(i,:)) < nrmpx(i))
p(i,:) = p2(j,:);
nrmpx(i) = norm(p(i,:) - x(i,:));
end
end
end
end
%---------------------------------------%
% Desired Direction vector from (2.6.5) %
%---------------------------------------%
if (nrmpx(i) == 0)
e(i,:) = [0,0];
else
e(i,:) = (p(i,:) - x(i,:))/(nrmpx(i));
end
%--------------------------------------%
% Driving Force component from (2.6.4) %
%--------------------------------------%
f((2*i)-1) = ((v0a(i)/tau) * e(i,1)) - (u((2*i)-1)/tau);
f((2*i)) = ((v0a(i)/tau) * e(i,2)) - (u((2*i))/tau);
%---------------------------------------------------------------------%
% Pedestrian Interactions from (2.6.8) - Summing all the interactions %
%---------------------------------------------------------------------%
fab = [0,0];
for j = 1:m
if (i ~= j)
dab = norm(x(i,:) - x(j,:));
if (dab < VS)
rab = r(i) + r(j);
nab = ((x(i,:) - x(j,:))/dab);
fab = fab + ((A1a*exp((rab - dab)/B1a))*nab)...
*(lambda + (1 - lambda)*((1+(-nab*e(i,:)'))/2))...
+ ((A2a*exp((rab - dab)/B2a))*nab);
end
end
end
f((2*i)-1) = f((2*i)-1) + fab(1,1);
f((2*i)) = f((2*i)) + fab(1,2);
%---------------------------------------------------------------------%
% Boundary Interaction component, this is analogous to the pedestrian %
% interactions - infact the formula only needs different constants %
%---------------------------------------------------------------------%

%-----------------------------------------------------%
% Summing all the boundary interactions from (2.6.12) %
%-----------------------------------------------------%
if BDRY == 1
for j = 1:length(bdry)
if (j == 1)
daB = norm(x(i,:) - bdry(j,:));
BDY = bdry(j,:);
rd = r(i) - daB;
elseif ((norm(x(i,:) - bdry(j,:))) < daB)
daB = norm(x(i,:) - bdry(j,:));

65
 BDY = bdry(j,:);
rd = r(i) - daB;
end
end
naB = (x(i,:) - BDY)/daB;
faB = ((A1*exp(rd/B1))*naB);
f((2*i)-1) = f((2*i)-1) + faB(1,1);
f((2*i)) = f((2*i)) + faB(1,2);
end
%----------------------------------%
% Displacement update from (2.6.1) %
%----------------------------------%
X((2*i)-1)= u((2*i)-1);
X(2*i) = u(2*i);
end

f0a = [f,X]';

66