VAN WAGENINGEN-KESSELS, Femke, LECLERCQ, Ludovic, DAAMEN, Winnie, HOOGENDOORN, Serge P.

, 2016, The Lagrangian

coordinate system and what it means for two-dimensional crowd flow models, Physica A: Statistical Mechanics and its Applications, 443,
ELSEVIER, pp. 272-285, DOI: 10.1016/j.physa.2015.09.048

The Lagrangian coordinate system and what it

means for two-dimensional crowd flow models
Femke van Wageningen-Kessels, Ludovic Leclercq,
Winnie Daamen and Serge P. Hoogendoorn

April 21, 2015

Abstract

A continuum crowd flow model is solved using the Lagrangian coordinate

system. The system has proven to give computational advantages over the
traditional Eulerian coordinate system for (one-dimensional) road traffic flow.
Our extension of the model and simulation method to (two-dimensional) crowd
flow paves the way to explore the advantages for crowd flow simulation. This
paper provides a first exploration and shows that a model and simulation method
for two-dimensional crowd flow can be developed using Lagrangian numerical
techniques and that it leads to accurate simulation results.

keywords Crowd flow, Model, Simulation, Lagrangian coordinates

1 Introduction
Crowd flow models are used to describe, understand and predict collective behaviour
of crowds. Roughly, three types of models exist [Bellomo et al., 2012, Duives et al.,
2013]: microscopic models in which the movements of individual pedestrians are de-
scribed and traced, mesoscopic models in which position and velocity of individual
pedestrians is described using probability distributions and macroscopic (or contin-
uum) models in which variables are aggregated and which are the focus of this study.
Continuum models describe the dynamics of crowds as a continuum flow, in terms
of average speed, velocity and density [Hughes, 2002]. Recent research shows that con-
tinuum models can reproduce self-organisation and certain important dynamic phe-
nomena such as lane formation and diagonal striping [Jiang et al., 2012, Hoogendoorn
et al., 2014, 2015], which is in contrast to previous claims. Furthermore, simulations
based on continuum models have the potential to significantly reduce computation
time whilst keeping high accuracy. This makes them useful for a larger range of
applications, including real time state estimation and prediction for crowd manage-
ment and optimisation of control strategies, under normal operating conditions and
for evacuations. For this, the solutions to the model equations need to be calculated
using efficient (i.e. both fast and accurate) computational methods.
The Lagrangian coordinate system has been applied to traffic flow models [Aw
et al., 2002, Leclercq et al., 2007]. Van Wageningen-Kessels et al. [2010], Yuan et al.

1
[2012] show that the system has many advantages over the traditional Eulerian coordi-
nate system, including more accurate reproduction of shock waves and more efficient
state estimation based on trajectory data. It is expected that the advantages also hold
in the case of two-dimensional crowd flow models resulting in more efficient simulation
results than with methods based on the Eulerian coordinate system such as proposed
by Dogbé [2008], Huang et al. [2009], Jiang et al. [2012], Hänseler et al. [2014]. In
this contribution we focus on showing how crowd simulations can be done using the
Lagrangian coordinate system. We also give some insight into the advantages of La-
grangian crowd simulations. Future studies will have to provide a definitive answer to
the question under which circumstances the Lagrangian coordinate system is superior
to the Eulerian coordinate system for crowd simulation.
Our main contribution is the proposal of an extension of Lagrangian simulation
methods previously applied for one-dimensional traffic flow models to two-dimensional
crowd flow models. We introduce the reformulation of the two-dimensional contin-
uum crowd flow model into Lagrangian coordinates (Section 2), develop a numerical
simulation method (Section 3) and show that it leads to meaningful simulation results
(Section 4). The contribution is summarised and open questions and future research
directions are discussed in Section 5.

2 Crowd flow model and its Lagrangian formulation

⃗ and time
In a traditional, Eulerian (fixed) coordinate system the density at location x
t is calculated:

ρ = ρ(⃗
x, t) (1)

In contrast, the main idea of the Lagrangian (or moving) coordinate system is that
⃗ of the n-th particle (i.e. a vehicle or a pedestrian) is calculated at time
the location x
t:

⃗=x
x ⃗(n, t) (2)

In vehicular traffic the n-th vehicle usually follows the (n−1)-th vehicle. Such a simple
ordering of pedestrians is usually not possible and limits the model formulation. We
show how the Lagrangian model can be formulated nevertheless, but without a closed-
form conservation equation.

2.1 Conservation model

Our approach includes a physical explanation of the conservation of particles. In
Eulerian coordinates, the one-dimensional conservation equation is well-known:

∂ρ ∂q
+ =0 (3)
∂t ∂x

2
with ρ the density (vehicles per meter), and q the flow (vehicles per second unit). The
Lagrangian equivalent of the one-dimensional conservation equation (3) is as follows:

∂(1/ρ) ∂v
+ =0 (4)
∂t ∂n

with v = q/ρ the vehicle speed (meter per second). To derive the Lagrangian equiv-
alent (4) of the one-dimensional conservation (3) equation in Eulerian coordinates,
a Moskowitz function can be defined [Newell, 1993]. Therefore, a numbering of the
vehicles n = 1, . . . , N is used and their position as a function of vehicle number x(n)
is defined [Aw et al., 2002, Leclercq et al., 2007, van Wageningen-Kessels et al., 2010].
However, for crowd flow, such a numbering of the pedestrians is not possible and
pedestrians can not be numbered consistently. This prevents a continuous mapping
of the pedestrians with respect to space and time. Therefore, an alternative approach
is proposed.

2.1.1 Two-dimensional conservation model

The two-dimensional extension of the Eulerian conservation equation (3) is well-

known:

∂ρ ∂qx ∂qy
+ + =0 (5)
∂t ∂x ∂y

with ρ the density (pedestrians per square meter), and qx and qy the flow in x- and
y-direction, respectively (pedestrians times meter per second).

2.1.2 Two-dimensional Lagrangian conservation model

We use the Lagrangian approach to solve the conservation equation in two dimensions
(5). Therefore, fixed portions of particles (i.e. groups of vehicles or pedestrians) are
defined and their position and speed is traced over time, see Figure 1. Note that the
portion does not necessarily include an integer number of particles but instead the
number of vehicles or pedestrians can take any positive value. In one-dimensional
road traffic flow, the portion of vehicles occupies a certain length of road L(t) at time
t. In two-dimensional crowd flow, the portion of pedestrians occupies a certain surface
A(t) at time t. The evolution of the size of such portions (i.e. the evolution of L(t)
or A(t) over time) and their location only depends on the movement of the particles
located at the boundary. Therefore, we have to assume that the boundary particles
remain at the boundary and that there is no switching of positions, i.e. vehicles or
pedestrians do not overtake each other. This assumption is true if the speed profile
is smooth enough and the evolution is followed over a small time period. The inflow
into the portion and the outflow out of the portion are by definition zero. Therefore,
we can apply the conservation principle and we find for one-dimensional flow:
t1
L(t1 ) = L(t0 ) + ∫ (vfirst (t) − vlast (t)) dt (6)
t0

with vfirst and vlast the speed of the first and the last vehicle in the portion, respec-
tively.

3
 space y
space x 5
6

A(t1 )

L(t1 )
A(t0 )
L(t0 )

x time t
ce
spa t0
t1
t0 t1 time t
(a) One-dimensional road traffic flow. (b) Two-dimensional crowd flow.

Figure 1: Following of the evolution of a fixed portion of particles (vehicles or pedestrians)

over time. Gray lines are trajectories.

The same principles can be applied to find the area occupied by a portion of the
pedestrian in two-dimensional crowd flows. Therefore, we first consider the movement
⃗j (t0 ) at time t = t0 and moves with velocity
of one pedestrian j, which is at location x
v⃗j (⃗
x, t). At time t = t1 > t0 its location will be:

t1
⃗j (t1 ) = x
x ⃗j (t0 ) + ∫ v⃗j (τ ) dτ (7)
τ =t0

Secondly, we consider a region Ωi with Ni (t) = ∫x⃗∈Ωi (t) ρ(⃗

x, t) d⃗
x pedestrians in it.
The area of this region equals:

Ai (t) = ∫ x=∮
1 d⃗ ⃗ d⃗
x x (8)
x
⃗∈Ωi (t) x
⃗∈Γi (t)

with Γ the boundary of the region. In the Lagrangian coordinate system, the region
moves with the pedestrians, the pedestrians on the boundary, are assumed to stay on
the boundary, the flux over the boundary is zero, and thus the number of pedestrians
does not change (N (t1 ) = N (t0 )). However, the region and its boundary move if the
speed of the pedestrians on the boundary is nonzero (Γi (t1 ) ≠ Γi (t0 )). Using both
(7) and (8), we can compute the area A(t1 ) at time t = t1 > t0 , see also Figure 1(b):

t1
Ai (t1 ) = ∮ ⃗ d⃗
x x=∮ x⃗ + ∫ v(⃗ x, τ ) dτ d⃗ x
x
⃗∈Γi (t1 ) x⃗∈Γi (t0 ) τ =t
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶ ´¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
new area old area change in area
t1
= Ai (t0 ) + ∮ ∫ x, τ ) dτ d⃗
v(⃗ x (9)
x
⃗∈Γi (t0 ) τ =t0

The above concept of change in area occupied by a group of pedestrians according

to the velocity of the pedestrians at the boundary, forms the basis of the numerical
computation method, as outlined in Section 3.

4
 qmax = ρcrit vfree vfree

ρcrit ρjam ρcrit ρjam

(a) Density-speed. (b) Density-flow.

Figure 2: Two representations of the bi-linear fundamental diagram.

2.2 Velocity: speed and route choice

To complete the model, we also need to describe the velocity v⃗(⃗
x, t). We propose
a simple velocity model, which may be applied either in Lagrangian or in Eulerian
coordinates. The velocity is calculated using a simplified version of the approach
proposed by Hoogendoorn et al. [2014, 2015]:

e⃗stat (⃗
x, t) + e⃗dyn (ρ(⃗
x, t))
v⃗(⃗ x, t)) e⃗(⃗
x, t) = V (ρ(⃗ x, t) = V (ρ(⃗
x, t)) (10)
∥⃗
estat (⃗x, t) + e⃗dyn (ρ(⃗
x, t))∥

with V (ρ) the fundamental relation and e⃗ the route choice (i.e. walking direction).
The route choice e⃗ is composed of the static route choice component e⃗stat and the
dynamic (local, state dependent) route choice component e⃗dyn .
We are well aware that other, possibly more realistic, speed and route choice
models exist. However, all models include key variables such as speed, density and
density gradient, just like our simple model. This allows us to develop a numerical
method, which could also be applied to more refined models. We emphasise that the
simple version of the velocity model will suffice to illustrate the numerical method,
which is the focus of our current study.

2.2.1 Speed

In our simplified approach, the speed only depends on density, following a bi-linear
fundamental relation, see also Figure 2:

⎧
⎪ if ρ ∈ [0, ρcrit ]
⎪ vfree
V (ρ) = ⎨ ρcrit vfree ρjam (11)
⎪
⎪ ( ρ − 1) if ρ ∈ (ρcrit , ρjam ]
⎩ ρjam −ρcrit

with vfree free flow speed, ρcrit critical density and ρjam jam density.

2.2.2 Route choice: static and dynamic

The static route choice component e⃗stat depends on the layout of the facility and
possibly other factors. It can, for example reflect the shortest route to the destination.
In the simulations (Section 4) we apply three different static route choices, which are
described in more detail there.
For the dynamic route choice we apply a simple model reflecting that pedestrians
avoid high density regions:

e⃗dyn = −βdyn ∇ρ (12)

5
 ○ ○ ○ ○ ○
▵ ▵ ▵ ▵ ▵ ▵
▵ ▵ ▵ ▵ ▵
○ ○ ○ ○ ○ ○
▵ ▵ ▵ ▵ ▵
▵ ▵ ▵ ▵ ▵ ▵
○ ○ ○ ○ ○
▵ ▵ ▵ ▵ ▵ ▵
▵ ▵ ▵ ▵ ▵
○ ○ ○ ○ ○ ○
▵ ▵ ▵ ▵ ▵
Figure 3: A staggered triangular grid. Density ρ▵ is calculated at the cell centres, velocity
▵ at the vertices.
v⃗○ is calculated ▵ ▵ ▵ ▵ ▵
○ ○ ○ ○ ○
with βdyn ≥ 0 the weight of the dynamic route choice. The dynamic route choice com-
ponent steers pedestrians in the direction with the steepest negative density gradient.,
i.e. away from other pedestrians.

3 Simulation approach
The main idea of the simulation approach using the Lagrangian coordinate system is
that a moving grid is applied. In particular, Lagrangian cells correspond to a fixed
portion of the crowd and the grid moves with the pedestrians in that portion. In the
following we discuss how the grid is constructed, how the velocity of the pedestrians
and the vertices of the grid is computed and how the grid is moved each time step
∆t. Furthermore, we discuss when and how the grid is reconstructed (remeshing).

3.1 The grid

We assume triangular grid cells for simplicity, but the methods can be extended to
other polygon-shaped grid cells. The grid is staggered, meaning that some variables
are calculated at the cell centres, others at the vertices, see Figure 3. We do not
consider any variables calculated at the edges. The cell centres are defined as the
centroid of the cells. Variables calculated at the cell centres are indicated with a
subscripted triangle (▵). Variables calculated at the vertex are indicated with a
subscripted circle (○). The main variable to be calculated at the cells centres is
density (ρ▵ ). The main variable to be calculated at the vertices is velocity (⃗
v○ ). This
is consistent with the cell definition: calculating the velocity at the vertices allows
the cells to be moved, calculating the density at the interior is natural as each cell
contains a fixed number of pedestrians.

3.2 Velocity of grid cell vertices

At the start of each time step, the location of the grid cells and their vertices and the
number of pedestrians within each cell is known from the previous time step. Recall
that the number of pedestrians in the portion can take any positive value and is not
integer. From the location of the vertices and the number of pedestrians in a cell, we

6
approximate the velocity of the grid vertices, as outlined below. We first discuss the
method for an inner vertex, later the adaptations for a boundary vertex are discussed.

3.2.1 Density in grid cells

The average density in the grid cell is calculated:

N▵
ρ▵ = (13)
A▵

and serves as an approximation of the density at the grid cell centre. To simplify
notation, we leave out the time index in this section wherever possible. N▵ is the
number of pedestrians in the grid cell (remains constant over time). A▵ is the area of
the grid cell (may change each time step). In general, the area is computed as in (8),
but because the grid cell is triangular we can use:

A△ = 1
2
∣(xj1 − xj3 )(yj2 − yj1 ) − (xj1 − xj2 )(yj3 − yj1 )∣ (14)

with the locations of the vertices {j1 , j2 , j3 } ∈ J of the triangle (see also Figure 4(a)):

⎛xj ⎞
⃗○,j =
x (15)
⎝ yj ⎠

3.2.2 Speed at vertices

To obtain the speeds from the fundamental relation, the densities at the vertices are
needed. The density at the vertex is an interpolation of the densities at the centroids
⃗▵,i are calculated
of the surrounding cells, see Figure 4. The centroids (cell centres) x
as the centre of mass of the cells:

⃗▵,i =
x 1
3
⃗○,j
∑x (16)
j∈Ji

with j ∈ Ji the indices of the vertices of cell i. The density at a vertex ρ○ (x○ , y○ ) is
calculated as a linear interpolation of the densities in the surrounding cells ρ▵ (x▵ , y▵ ).
This results in:

ρ○ (x○ , y○ ) = c1 x○ + c2 y○ + c3 . (17)

with parameters c1 , c2 and c3 determined by the linear interpolation function, which

uses a truncated Taylor series [D’Errico, 2012]. Subsequently, the speed of the vertices
are approximated using the fundamental relation v○ = V (ρ○ ), see equation (11). In
this simple algorithm we do not take into account anisotropy (the fact that pedestrians
react differently to pedestrians in front than to those to the side and to the back).
The method could possibly be improved by including anisotropy using weights for
each of the surrounding cells.

7
3.2.3 Route choice at vertices

For the dynamic route choice component we need the density gradient at the vertices.
These can be obtained directly from the parameters of the linear interpolation (17),
see also Figure 4:

⎛c1 ⎞
∇ρ○ = (18)
⎝c2 ⎠

We use e⃗dyn,○ = −βdyn ∇ρ○ (see equation (12)) to calculate the dynamic route choice
component.
The static route choice component e⃗stat is supposed to be given at any location
and time.

2 1
3
3 ∇ρ○ , ρ○ 6
⃗▵
A▵ , ρ ▵ , x
4
2 5
1
(a) Vertex numbering to compute (b) Cell numbering to compute density and density
area, density and centre of a cell. gradient at vertex.

Figure 4: Examples of vertex and cell numbering. This is used to compute the area, density
and centre of a cell (a) and to compute the density and density gradient at a vertex (b).

3.2.4 Velocity of vertices

We can now approximate the velocity of the vertex (see equation (10)):

e⃗stat,○ − βdyn ∇ρ○

v⃗○ = v○ (19)
∥⃗
estat,○ − βdyn ∇ρ○ ∥

3.2.5 Boundary conditions

At the boundaries, vertices are only partly surrounded by cells and density and den-
sity gradient can not be computed straightforwardly as in (17) and (18). A simple
approach to deal with boundaries is introduced. Alternative, more accurate numerical
boundary conditions may be developed in future research. In the current approach,
ghost cells are introduced, see Figure 5. They have their centroids opposite of the
centroids of the surrounding cells of the boundary vertex. Furthermore, since we are
only dealing with free boundaries in the test cases in the next section, the ghost cells
have zero densities. To calculate the density gradient, the ghost cells are included in
the linear regression model as in (17) and the density gradient is as in (18). However,
when calculating density, ghost cells are only included in the regression model (17) if
they are located in the walking direction according to the static route choice.

8
 static route choice

G3
C1 G2

V0 G1
C2
C3

Figure 5: Example of boundary with ghost cells. C1 , C2 and C3 are the centroids of
the cells neighbouring boundary vertex V0 . The centroids of their respective ghost cells are
located at G1 , G2 and G3 . Only ghost centroids G2 and G3 are in the static route choice
walking direction and they are taken into account when computing the density at boundary
vertex V0 . All ghost cells are taken into account when calculating the density gradient.

Figure 6: Example of movement of the grid during one time step. The old grid is shown in
thin blue lines. The new grid in thick black lines. One grid cell is highlighted and the figure
shows how it moves and deforms, following the velocities of its corners (red arrows).

3.3 Moving of grid cells

Once the velocity of the vertices is known, they can be moved. Therefore, we assume
that the velocity remains as computed in (19) during the time step. The vertices are
moved as follows:

⃗new
x ○ ⃗old
=x ○ + ∆t⃗
v○ (20)

This is illustrated in Figure 6. We assume that the velocity of the edges is a linear
interpolation of the velocities of the neighbouring vertices. Therefore, edges remain
straight and the cell remains triangular.

3.4 Remeshing the grid

The vertices of an individual grid cell, will usually have different velocities. Therefore,
the grid cell will change its shape. If this deforms the grid cells ‘too much’ and the
grid becomes too irregular, it is be necessary to remesh. Furthermore, densities may
become too large (i.e. above jam density) and remeshing becomes necessary for this
reason. Finally, if vertices move in opposite directions or if they have large speed

9
 (a) Opposite di- (b) Different speeds.
rections.

Figure 7: Cells ‘flip’ if the velocities of their vertices (red arrows) are very different. Blue
cell with thin lines: old, gray cell with thick lines: new.

(a) Enlarging grid cell. (b) Moving vertex in inner domain.

Figure 8: Illustrations of the remeshing approaches. Thin black line indicates the old mesh,
thick blue ones the new mesh, arrows indicate the movement of the vertex or vertices. In
Figure 8(a), the dot indicates the centroid of the cell that is to be enlarged, in Figure 8(b),
the dot is the centroid of the polygon consisting of all triangles shown in the figure.

differences, the cell may ‘flip’ (see Figure 7).

In the remeshing process a new mesh is formed and the pedestrians in the old
grid cells, are assigned to new grid cells in such a way that the density at each
location x⃗ remains (approximately) the same and the total number of pedestrians
in the computational domain does not change (conservation of pedestrians). Too
achieve a computationally efficient remeshing process, only vertices and cells that
satisfy certain remeshing criteria are remeshed, while others are left unchanged.
The objective of the remeshing is twofold: the density in each cell is to be kept
below (or at) jam density and the grid is to be kept regular with the shape of the cells
‘close enough’ to an equilateral triangle. To achieve the first objective, grid cells that
have densities close to jam density are enlarged, see Figure 8(a). To achieve the second
objective vertices are moved if the are ‘too close’ to other vertices, see Figure 8(b).
The combination of both the cell enlargement and the moving of vertices, prevents
the cells from flipping. The remeshing procedures are detailed below.

3.4.1 Enlarging grid cells

A grid cell is enlarged if its density becomes too large:

ρ > αρ ρjam (21)

10
with αρ ∈ ( ρρjam
crit
, 1] a parameter. Too enlarge the grid cells, all vertices of the cell are
moved outward. They follow the line from the centroid (⃗
xc ) of the grid cell through
the original position of the vertex (⃗
xi ), see Figure 8(a). The new position is:

⃗∗i = (1 + βρ )⃗
x ⃗c
x i − βρ x (22)

with βρ > 0 a parameter. Currently, an appropriate value for βρ is determined by

trial and error for the simulations. A more rigorous approach, possibly taking into
account the actual traffic state can be developed in the future.

3.4.2 Moving vertices to keep grid regular

⃗i is moved if its position is far away from the centroid of the

Furthermore, a vertex x
surrounding polygon x⃗c , compared to its distance to any of its neighbouring vertices
⃗j :
x

∥⃗ ⃗c ∥ > αc max ∥⃗
xi − x ⃗j ∥
xi − x (23)
j

with αc > 0 a parameter. The vertex is then moved towards the centroid of its
surrounding polygon, see Figure 8(b). Its new position is:

⃗∗i = (1 − βc )⃗
x ⃗c
x i + βc x (24)

with βc ∈ (0, 1) a parameter. Just like with βρ , the value of βc is currently determined
by trial and error for the simulations, leaving room for more efficient approaches.

A2
A1

A3 A∗2 ∗
A∗3 A1 A6

A∗6 A7
A∗4
A4
A∗5
A∗7
A5

Figure 9: Example of recalculating densities after remeshing. The old mesh consists of
(lightly coloured) triangles Ai . One triangle of the new mesh is drawn with thick black lines.
It overlaps with 7 triangles of the old mesh. The intersections of the old and new triangles
are brightly coloured and are indicated with A∗i .

3.4.3 Recalculation the densities

After the new grid has been determined, the density in each of the reshaped grid
cells is computed. Figure 9 shows an example. For the remeshing, density in the j-th

11
 y y R2 y
R
R1
x x Cy
x
Lx Cx
(a) Case 1: initial jam will dis- (b) Case 2: pedestrians are ini- (c) Case 3: pedestrian are ini-
solve as pedestrians walk to the tially positioned in circle, static tially positioned in ellipse, will
right. route choice attracts pedestri- walk to right and then turn right
ans to its centre. around a corner.

Figure 10: Initial state and static route choice. Gray area indicates original position of
pedestrians. Thick blue arrows indicate static route choice walking direction.

triangle of the new mesh ρremeshed

j is calculated as follows:

A∗i
Njremeshed ∑iAi i
N
ρremeshed
j = = (25)
Aremeshed
j
remeshed
Aj

with Ai the area of the i-th triangle of the old mesh, Ni the number of pedestrians
in that triangle and A∗i the area of the intersection between the i-th triangle of the
old mesh and the j-th triangle in the new mesh.

4 Simulation and results

We perform simulations with the newly developed method to show that crowds can
indeed be simulated using Lagrangian techniques and that this approach gives good
solutions. We consider three test cases. In the first and second case we show that
the numerical solution using the newly developed Lagrangian simulation method ap-
proximates the analytical solution well. Furthermore, the second test case shows the
influence of the weight of the dynamic route choice component. The third test case
shows a more realistic setting where a group of pedestrians rounds a corner.

4.1 Test set up

In all three test cases we use remeshing settings and a bi-linear fundamental diagram
with parameters as in Table 1 and initial state and static route choices as in Figure
10. The other settings are different for each of the cases, as Table 1 shows, and they
are described in more detail below.

4.1.1 Case 1 (‘straight’)

Test case 1 is used to show the numerical accuracy, that is: the ability of the La-
grangian simulation method to approximate the exact solution of the model equations.
In this simple test case, initially, there is a high density area, which could resemble
a jam in a wide corridor. At the front of the jam, pedestrians start walking and the
jam dissolves. We ignore the influence of the dynamic route choice (i.e. we set its
weight to zero: βdyn = 0), which enables us to calculate the analytical solution using

12
 Table 1: Parameters of test cases

Fundamental diagram parameters

free flow speed vfree 1.3 m/s
critical density ρcrit 1.35 peds/m2
jam density ρjam 5.4 peds/m2

Remeshing settings
threshold enlarge cell αρ 1.0
factor enlarge cell βρ 0.1
threshold move vertex towards centroid αc 0.33
factor move vertex towards centroid βc 1.0

Case dependent settings: route choice

Case 1 Case 2 Case 3
1 x
static e⃗stat ( ) √ −1 ( ) see Figure
0 x2 +y 2 y
10(c)
dynamic βdyn 0 [5, 10, 20, 40, 80] 20 m2 /ped

Case dependent settings: initial condition

Case 1 Case 2 Case 3
density ρ(x, 0) 0.75ρjam = 4.05 0.4ρcrit = 0.54 0.6ρcrit = 0.81 peds/m2
Rx =−30 Ry =35
position Lx = 80 R = 25 m
Cx =10 Cy =15
nr of peds Ntot 3637 1058 380

Case dependent settings: mesh and numerical settings

Case 1 Case 2 Case 3
initial cell area A 2.1 [10.9, 14.3] [7.3, 9.6] m2
peds per cell N 8.4 [5.9, 7.7] [5.9, 7.8]
nr of cells I 434 150 54
nr of vertices P 240 91 37
time step size ∆t [0.2, 0.5, 1, 2] 0.1 0.5

13
 initial mesh initial mesh initial mesh
40 50
20
35 45

10
30 40

25 0 35
y

y
20 30
−10

15 25
−20
10 20
−30 −20 −10 0 −20 −10 0 10 20 −40 −30 −20
x x x

(a) Case 1 (b) Case 2 (c) Case 3

Figure 11: Initial meshes.

kinematic wave theory and compare it with the simulation result. Furthermore, we
compare the results for different time step sizes to get more insight into the accuracy.

4.1.2 Case 2 (‘circular’)

Test case 2 is also used to show the numerical accuracy, now with taking the dynamic
route choice term into account. It also shows the influence of the weight of the
dynamic route choice on the outcome of the model. In this test case, initially, there
is a circular high density area. The static route choice represents an attractor at the
centre of this area: that is where the pedestrians would want to go. The static route
choice component is normalised such that its length always equals one: ∥⃗
estat ∥ = 1.
However, the pedestrians will also walk away from the centre because it is too crowded,
which is represented by the dynamic route choice term. For different weights of the
dynamic route choice term, the stationary state in which the crowd does not move
anymore is computed analytically and compared to the simulation results.

4.1.3 Case 3 (‘corner’)

Test case 3 is used to show that the numerical method also works for (slightly) more
realistic cases in which a corner is involved. The static route choice is chosen according
to equidistant trajectories around the corner. And again, the static route choice
component is normalised such that its length always equals one: ∥⃗
estat ∥ = 1. The
analytical solution is not computed, but we will discuss the plausibility of the results.

4.2 Initialization
For the initialisation of the test cases, a mesh is created manually, depending on the
initial positioning of the pedestrian. The initial meshes are shown in Figure 11. A
certain number of pedestrians is assigned to each cell according to the initial density.
The values are shown in Table 1.

4.3 Test results

Test results are shown in Figures 12–15. The results for Case 1 (Figures 12 and
13) show that the numerical method is able to reproduce how the jam dissolves and

14
 y (m)
30 5
25
20 0
−20 0 20 40 60 80 100
x (m)
(a) Time t = 10 s

y (m)
30 5
25
20 0
−20 0 20 40 60 80 100
x (m)
(b) Time t = 20 s
y (m)
30 5
25
20 0
−20 0 20 40 60 80 100
x (m)
(c) Time t = 40 s
y (m)

30 5
25
20 0
−20 0 20 40 60 80 100
x (m)
(d) Time t = 80 s

Figure 12: Case 1, densities at different times, with time step size ∆t = 1 s.

150
150

100
100
x−location (m)

x−location (m)

50
50

0
0

−50 −50
0 20 40 60 80 100 0 20 40 60 80 100
time (s) time (s)

(a) ∆t = 0.2 s (b) ∆t = 0.5 s

150 150

100 100
x−location (m)

x−location (m)

50 50

0 0

−50 −50
0 20 40 60 80 100 0 20 40 60 80 100
time (s) time (s)

(c) ∆t = 1 s (d) ∆t = 2 s

Figure 13: Case 1, development of densities at cell centres over time, with different time
step sizes ∆t. Colored dots indicate locations of cell centres each time step. Red indicates
high density, gradually changing into orange for lower densities, until critical density. Green
indicates density below critical. Blue lines indicate analytical solution.

15
 50
5
3.5

3 4

2.5
density (ped/m2)

y (m)
2 0

1.5 2

1
1
0.5

0 −50 0
0 5 10 15 20 25 −50 0 50
radius (m) x (m)

(a) Weight dynamic route choice βdyn = 5.

50
5
1.4

1.2 4

1
density (ped/m2)

3
y (m)

0.8 0

0.6 2

0.4
1
0.2

0 −50 0
0 5 10 15 20 25 30 −50 0 50
radius (m) x (m)

(b) Weight dynamic route choice βdyn = 20.

50
5
0.7

0.6 4

0.5
density (ped/m2)

3
y (m)

0.4 0

0.3 2

0.2
1
0.1

0 −50 0
0 10 20 30 40 50 −50 0 50
radius (m) x (m)

(c) Weight dynamic route choice βdyn = 80.

Figure 14: Densities for Case 2, for various weights of the dynamic route choice. Left:
average density at certain radii. Shown every second: light colors for first time steps, gradu-
ally getting darker. Black thick line: density in steady state according to analytical solution.
Blue: density after 15 seconds. The plot shows that the density barely changes anymore and
this is considered the numerically computed steady state. Right: space-density plots. Only
density after 15 seconds (numerical steady state) is shown. Black circle indicates radius of
group in steady state according to analytical solution.

16
 80 80 80
5 5 5
60 60 60
4 4 4
40 40 40
3 3 3
y (m)

y (m)

y (m)
20 20 20

2 2 2
0 0 0

−20 1 −20 1 −20 1

−40 0 −40 0 −40 0

−60 −40 −20 0 20 40 60 −60 −40 −20 0 20 40 60 −60 −40 −20 0 20 40 60
x (m) x (m) x (m)

(a) Time t = 10 s (b) Time t = 20 s (c) Time t = 30 s

80 80 80
5 5 5
60 60 60
4 4 4
40 40 40
3 3 3
y (m)

y (m)

y (m)
20 20 20

2 2 2
0 0 0

−20 1 −20 1 −20 1

−40 0 −40 0 −40 0

−60 −40 −20 0 20 40 60 −60 −40 −20 0 20 40 60 −60 −40 −20 0 20 40 60
x (m) x (m) x (m)

(d) Time t = 40 s (e) Time t = 60 s (f) Time t = 80 s

Figure 15: Case 3: Densities at different times.

pedestrians start walking. Figure 13 also includes a comparison with the analytical
solution. This comparison shows that there is some delay in when the pedestrians start
to walk, resulting in the front of the congestion wave travelling backward slower than
expected. Comparison over different time step sizes, shows that this delay decreases
for small time step sizes and that more accurate results can be obtained using smaller
time steps, which indicates that the method is consistent.
The results for Case 2 (Figure 14) show that the numerical method is well able
to approximate the stationary solution. After 15 seconds the solution barely changes
anymore and this is considered to be the numerical steady state. The results on
the left show that the numerical steady state (blue lines) and analytical steady state
(black lines) are almost identical. Furthermore, the test results show that the solution
depends heavily on the weight of the dynamic route choice. This can be seen best on
the right hand side, where the radius of the circle depends on the weight. If it has a low
value (βdyn = 5, Figure 14(a)), the pedestrians stay closely together (in this example
in an area with radius of about 17 meter), while a high value (βdyn = 80, Figure 14(c))
makes the pedestrians spread out over a much larger area (in this example with radius
of about 43 meter). Therefore, in most applications, it will be important to calibrate
this parameter well.
The results for Case 3 (Figure 15) show that the method is also capable of sim-
ulations with slightly more complex layout. The pedestrians first walk to the right,
while spreading out a bit more than in the initial solution. They round the corner in
a way similar to observations with low densities. The results show that the numerical
method allows for changes of the shape of the pedestrian group. Furthermore, the
numerical scheme reproduces what we expect from the behavioural hypothesis in the
model.

17
5 Summary and future work
A numerical simulation method for crowd flow based on the Lagrangian coordinate
system is introduced. The method uses a triangular mesh that moves according to
rules dictated by the traffic flow model. In order to prevent too high densities and
too much deformation of the mesh, a remeshing procedure is introduced. Test results
show that this leads to plausible simulation results and that high accuracy can be
obtained.
The Lagrangian simulation method can be applied to solve continuum crowd flow
models. It can be applied to a simple model as presented here, but also to more
advanced models. Further research is necessary to quantify the efficiency in terms
of accuracy and computational speed, and to compare the efficiency with that of
alternative methods such as those based on Eulerian grids fixed in space[Hänseler
et al., 2014]. Certain aspects can be studied further, such as the approach to free
boundaries and the remeshing procedure. This can include a sensitivity analysis of
the remeshing parameters. Furthermore, future research will focus on alternative
settings of the numerical method, including meshes consisting of generic polygons
and alternative remeshing procedures. Future extensions of the method include the
introduction of closed boundary conditions (including walls and obstacles), inflow and
outflow boundary conditions and multiple groups that have different walking or route
choice characteristics or destinations.

Acknowledgements
This research is performed as part of the The Netherlands Organisation for Scientific
Research (NWO) Aspasia grant of Daamen (Daamen, Van Wageningen-Kessels, grant
number 016.125.440) and as part of the NWO-VICI project ‘Travel Behavior And
Traffic Operations In Case Of Exceptional Events’ (Hoogendoorn, grant number 453-
08-006).

References
A. Aw, A. Klar, M. Rascle, and T. Materne. Derivation of continuum traffic flow
models from microscopic follow-the-leader models. SIAM Journal on Applied Math-
ematics, 63(1):259–278, 2002.

N. Bellomo, B. Piccoli, and A. Tosin. Modeling crowd dynamics from a complex

system viewpoint. Mathematical Models and Methods in Applied Sciences, 22(2),
2012.

J. D’Errico. Polyfitn matlab function, 2012. URL http://www.mathworks.com/

matlabcentral/fileexchange/34765-polyfitn. Updated in April 2014.

C. Dogbé. On the numerical solutions of second order macroscopic models of pedes-

trian flows. Computers & Mathematics with Applications, 56(7):1884–1898, 2008.

18
D. C. Duives, W. Daamen, and S. P. Hoogendoorn. State-of-the-art crowd motion
simulation models. Transportation Research Part C: Emerging Technologies, 37:
193–209, 2013.

F. S. Hänseler, M. Bierlaire, B. Farooq, and T. Mühlematter. A macroscopic loading

model for time-varying pedestrian flows in public walking areas. Transportation
Research Part B: Methodological, 69:60–80, 2014.

S. P. Hoogendoorn, F. L. M. van Wageningen-Kessels, W. Daamen, and D. C. Duives.

Continuum modelling of pedestrian flows: From microscopic principles to self-
organised macroscopic phenomena. Physica A: Statistical Mechanics and its Appli-
cations, 416:684–694, 2014.

S. P. Hoogendoorn, F. L. M. van Wageningen-Kessels, W. Daamen, and M. Sarvi.

Continuum theory for pedestrian traffic flow: Local route choice modelling and its
applications. 2015. Accepted for ISTTT21 and subsequent publication in Trans-
portation Research Series.

L. Huang, S.C. Wong, M. Zhang, C.-W. Shu, and W. H. K. Lam. Revisiting Hughes’
dynamic continuum model for pedestrian flow and the development of an efficient
solution algorithm. Transportation Research Part B: Methodological, 43(1):127–141,
2009.

R. L. Hughes. A continuum theory for the flow of pedestrians. Transportation Research

Part B: Methodological, 36(6):507–535, 2002.

Y. Jiang, S.C. Wong, P. Zhang, R. Liu, Y. Duan, and K. Choi. Numerical simulation
of a continuum model for bi-directional pedestrian flow. Applied Mathematics and
Computation, 218(10):6135–6143, 2012.

L. Leclercq, J.A. Laval, and E. Chevallier. The Lagrangian coordinates and what
it means for first order traffic flow models. In R. E. Allsop, M. G. H. Bell, and
B. G. Heydecker, editors, Transportation and Traffic Theory 2007, pages 735–753,
Oxford, 2007. Elsevier.

G. F. Newell. A simplified theory of kinematic waves in highway traffic (part I-III).

Transportation Research Part B: Methodological, 27(4):281–313, 1993.

F. L. M. van Wageningen-Kessels, J. W. C. van Lint, S. P. Hoogendoorn, and C. Vuik.

Lagrangian formulation of a multi-class kinematic wave model. Transportation
Research Record: Journal of the Transportation Research Board, 2188:29–36, 2010.

Y. Yuan, J. W. C. van Lint, R. E. Wilson, F. L. M. van Wageningen-Kessels, and

S. P. Hoogendoorn. Real-time Lagrangian traffic state estimator for freeways. IEEE
Transactions on Intelligent Transportation Systems, 13(1):59–70, 2012.

19