SELECTED APPLICATIONS OF A DYNAMIC ASSIGNMENT METHOD FOR

MICROSCOPIC SIMULATION OF PEDESTRIANS

Tobias Kretz, Karsten Lehmann, Thomas Friderich
PTV Group, Haid-und-Neu-Strae 15, D-76131 Karlsruhe, Germany
{Firstname.Lastname}@ptvgroup.com

1. INTRODUCTION
In micro-simulation models of pedestrian dynamics where pedestrians move
freely in two spatial dimensions (i.e. they are not restricted to links as vehicles
are) it has so far not been possible to apply methods of iterated (dynamic)
assignment which were developed for application with simulations of vehicular
traffic. The obvious reason is that the network structure of the infrastructure of
vehicular traffic a priori allows only a limited number of routes (provided loops
are excluded), while the areas of pedestrian dynamics in principle allow an
unlimited (even uncountable) number of routes.
If one wants to apply assignment methods which have been developed for
vehicular simulation to pedestrian simulation one therefore has to find a
limited subset out of the infinitely many routes a pedestrian can walk between
origin and destination. The subset needs to be meaningful in the sense that it
should represent route choices which are actually perceived as such.
In this conference contribution such a general method will be summarized.
However, rather than the details of the method its application will be in the
focus of this contribution. The full definition of the method will be given
elsewhere (Kretz, Lehmann, & Hofs, submitted) together with an extensive
overview of related work.
2. SHORT SUMMARY OF AN ASSIGNMENT METHOD FOR PEDESTRIAN
SIMULATION
Compared to vehicle simulations the main challenge for applying assignment
methods to pedestrian simulations is that pedestrians move on areas instead
of networks, i.e. they have not a limited set of discrete routes, but even if
one excludes loops an infinite number of possible paths which may differ
from each other only by infinitesimally small differences.
The task therefore is to find a limited, but meaningful set of routes.
Meaningful means that the routes should in fact model the main route
choices around obstacles and through rooms and corridors.
The basic idea of a method to compute such routes is to decompose the
pedestrian walking area into regions which are within a certain range of
AET 2013 and contributors

distance to destination, i.e. there is then a region which encloses all points
which are between 0 and 5 m from destination, another region which is 5 to
10 m to destination, and so on.
Obstacles and walls can cause that there are two or more separate regions
with the same distance to destination. This means that looking back from the
destination the path splits up (in upstream perspective).
If such a split up path combines again (in upstream perspective, i.e. in
downstream perspective this would be the split up point), the two regions
immediately downstream of the first combined upstream area are made
intermediate destination areas of two newly created separate routes. This is
visualized for a very basic example in FIGURE 1.

FIGURE 1: The destination is shown as green and black checkerboard to the right. In the middle
there is an obstacle shown in red and white. The light and dark cyan, as well as the magenta and
orange areas are region of equal distance to destination (modulo some value). The two orange
regions are split up and distinct and immediately neighboured to a region which is not split up
(the magenta region). Therefore they become intermediate destinations for two routes which
guide pedestrians around the obstacle.

FIGURE 1 shows that it is crucial how large the modulo value is, i.e. if it is 5 or
10 or 15 m which are combined to one region. For a value of 15 m in the
example of FIGURE 1 there would be no distinct regions with identical
distance to destination and consequently no route choice created. The modulo
value therefore is the key to the number of route choices and routes which are
being created in a model.
In more complex scenarios one would now continue recursively applying the
method starting at each of the newly created intermediate destination areas
and in this way create further route choices upstream, if required.

AET 2013 and contributors

A big advantage of the proposed method is that from the trajectories of

pedestrians locally one cannot see that the created intermediate destinations
exist at all. This is because in most if not all pedestrian simulations models
pedestrians basically follow the direction of the shortest path (fewest meters)
and because the edges of the intermediate destination areas are aligned with
lines of equal distance to the next downstream destination.
To understand the following examples it remains to be pointed out that the
method can produce two or more routes which for a certain origin area
imply the same navigation advice. As an example imagine in FIGURE 1 the
origin area was in the lower left corner. Pedestrians normally would walk the
same way if they follow a route without intermediate destination area (directly
to the final destination) or a route with the lower orange area as intermediate
destination area.
3. THE ASSIGNMENT METHOD
In each example a demand has been set, which is over the capacity of the
shortest path, but which can be accommodated by the full infrastructure.
There is a relevant time interval in which the travel times of all pedestrians
arriving are measured and contribute to the average travel time for the path
they have used. In each iteration only one simulation run is carried out. The
iteration process continues until the difference between the longest and the
shortest average route travel time is 0.5 seconds or less or until the shifted
route choice ratio was smaller than 0.0005. The computation of the route
choice ratios from the travel times is deliberately done rather simple. Between
the pair with the longest and the shortest travel times (tMax and tMin) the
following probability p is shifted (obviously from the route with the longer to
the route with the shorter travel time):
p=

t Max t Min
t Max +t Min

where is a general sensitivity factor which was chosen to be =0.1 in all

computations and is a dynamic adaptation factor which usually was =1, but
was decreased when the routes with the longest and the shortest travel time
were the same in subsequent iterations and which was increased when they
exchanged roles in subsequent iterations.
The operational simulation of pedestrians has been done with PTV Viswalk
5.40 (Kretz, Hengst, & Vortisch, 2008).
AET 2013 and contributors

4. EXAMPLES
1.1. Example 1: A Room with two Obstacles

shows a simple example. There are two large obstacles which have just an
intermediate size that they can be seen as obstacles in one large room or as
walls separating two different rooms. In addition there are four smaller
obstacles which we want to have handled by the operational simulation model
of pedestrian dynamics.

FIGURE 2: Obstacles are shown red with white edge, walking areas dark grey, the origin to the
left is coloured orange and the destination to the right green; light grey areas are not accessible.
The left figure shows the basic geometry, the right figure has the computed intermediate
destinations and routes added.

It can easily be seen in FIGURE 2 that for the origin area as existing in this
scenario routes 1, 3, and 6 as well as routes 2 and 5 in general imply the
same routing for pedestrians. (However, note that if the origin area was
located somewhere else, for example in the upper left corner this would be
different.) Transferring a concept from road traffic network assignment we can
say they have a commonality factor of 1 (Cascetta, Nuzzolo, Russo, &
Vitetta), although a strict and general definition of the commonality factor for
pedestrian routes will not be given here. Whereas the problem of routes with a
high communality factor is known from and relevant in vehicular traffic
assignment, routes with a communality factor of 1 do not exist there,
respectively the problem is irrelevant. Such routes could easily be spotted and
eliminated as the data defining them would typically be identical (the same set
of links).
The normal way forward now would be to first simulate including all routes and
later try if and how results change when some of the routes with identical
routing are deleted change. However, as the results are simpler and converge
much quicker in the latter case we only present the results of assignments
iterations with only the routes 1, 2, and 4 respectively 4, 5, and 6. In these two
cases no two routes do identical routing. The case with all six routes and the
problems following will be discussed somewhere else. Here instead we
AET 2013 and contributors

present an additional example and a comparison with the results when a

dynamic potential is applied instead of the iterated assignment.
The former case (1, 2, 4) includes routes with as few intermediate destinations
as possible (i.e. much freedom' for pedestrians or better said for the
operational simulation model of pedestrian dynamics), the latter case includes
routes with a maximum of intermediate destinations (i.e. as few freedom as
possible).
and show the results for two different initial distributions when only routes 1,
2, and 4 are utilized. When nearly all pedestrians are assigned to route 1 (i.e.
the original route without intermediate destinations) in the first iteration the
process terminates after iteration 19 with a weighted average travel time of
61.4 seconds. Beginning with equal distribution results in a travel time of 62.2
seconds after 15 iterations.

FIGURE 3: Weighted average travel time (left) and weighted standard deviations of travel times
(right) -- both in [s] -- in the course of iterations when only routes 1, 2, and 4 are utilized. The two
lines show the results when (blue line) in the first iteration 98% of pedestrians choose route 1
(and the other two routes are used by 1% each) and (red line) when the initial distribution is
equal.

FIGURE 4: Travel times on each of the routes (left) and route choice ratios (right) vs. iteration
number when only routes 1, 2, and 4 are utilized. The case with 98% on route 1 in the first
iteration is shown blue, equal distribution in the first iteration red.

FIGURE 5 and FIGURE 6 show the corresponding results when only routes 4,
5, and 6 are utilized. When nearly all pedestrians are assigned to route 6 in
the first iteration the process terminates after iteration 23 with a weighted

AET 2013 and contributors

average travel time of 62.0 seconds. Beginning with equal distribution results
in a travel time of 61.2 seconds after 8 iterations.
TABLE 1 summarizes the results of the cases when only three routes are
utilized. It can be seen that under different conditions the results of the
assignment process agree closely. This is a good sign for the quality of the
algorithm that computes the intermediate destination areas: obviously it does
not make a difference if pedestrians are sent via route 1 or 6 and respectively
route 2 or 5. This does not mean that in early iteration steps there is much of a
similarity. FIGURE 7 shows a comparison of the 1, 2, 4 and the 4, 5, 6 routes
computations during the first iteration step when nearly all pedestrians are
sent via route 1 respectively 6.

FIGURE 5: Weighted average travel time and weighted standard deviations of travel times -- both
in [s] -- in the course of iterations when only routes 4, 5, and 6 are utilized. The two lines show
the results when (blue line) in the first iteration 98\% of pedestrians choose route 1 (and the
other two routes are used by 1\% each) and (red line) when the initial distribution is equal.

FIGURE 6: Travel times on each of the routes and route choice ratios vs. iteration number when
only routes 4, 5, and 6 are utilized. The case with 98\% on route 1 in the first iteration is shown
blue, equal distribution in the first iteration red.

To get an estimation what the remaining differences for mean travel times in
TABLE 1 imply we have done four more computations with different
initialization values for the random number generator of the pedestrian
simulation for the 4, 5, 6, equal case. The minimum average travel time of
the overall ten computations was 61.2 s, the maximum 65.6 s and the
average 63.7 s with a standard deviation of 1.3 s. This shows that the results
given in TABLE 1 are not more disperse than for one and the same case with
only different random numbers.
AET 2013 and contributors

11

Initial distribution
98% route 1
Equal dist
98% route 1
Equal dist

Route 1
r.c.r.
t.t.
0.480 61.5
0.482 62.2
Route 6
r.c.r.
t.t.
0.480 62.0
0.480 61.2

Route 2
r.c.r.
t.t.
0.399 61..4
0.402 62.1
Route 5
r.c.r.
t.t.
0.411 62.0
0.394 61.2

Route 4
r.c.r.
t.t.
0.121 61.1
0.117 62.4
Route 4
r.c.r.
t.t.
0.109 62.4
0.126 61.3

Mean
t.t.
61.4
62.2
Mean
t.t.
62.0
61.2

TABLE 1: Summary of results when only three routes are utilized (1, 2, and 4 or 4, 5, and 6): the
route choice ratios (r.c.r.) and travel times (t.t.) per route and in the last column the weighted
average travel time. As route 6 corresponds to route 1 and route 5 corresponds to route 2 in the
lower half of the table the sequence of columns does not follow the sequence of route IDs.

FIGURE 7: Still images from the first iteration step of the case with routes 1, 2, and 4 (upper
figure) and 4, 5, and 6 (lower figure). Nearly all pedestrians are sent along route 1 or 6
respectively. As the demand clearly exceeds the capacity of these two routes a jam forms. In the
former case at some point it is shorter (in terms of meters) for pedestrians to pass the largest
obstacle on the left side. I.e. although they are sent on route 1, they move on route 2, disturbing
the assignment algorithm. In the latter case the second intermediate destination of route 6
which is located between the two larger obstacles prevents this from happening for a long time
and when it happens pedestrians do not walk directly to the destination, but first have to
approach the back side of the second intermediate destination area of route 6 also disturbing the
assignment process.

In both cases (utilizing routes 1, 2, and 4 as well as routes 4, 5, and 6) the

operational model of pedestrian dynamics at some point handles the large
obstacle on the operational level in a way that interferes with the tactical level.
This is a phenomenon that is new in pedestrian simulation compared to
vehicular simulation, where it either does not occur or can be prevented rather
easily. Nevertheless in subsequent iterations the phenomenon disappears as
the load on route 1 is decreased, i.e. the assignment process is able to cope
with the problem.
It is interesting to compare these results with the results that emerge from a
simulation utilizing a dynamic potential. Dynamic potential means that there
are no intermediate destinations, but that in each simulation time step the
basic (desired) direction of movement for the pedestrians is calculated as
direction along the shortest (fewest meters) path, but the estimated earliest
arrival (fewest seconds). I.e. the dynamic potential is a one-shot assignment
approach (Kretz, et al., 2011) (Kretz, et al., 2011) (Kretz & Groe, 2012)
(Kretz, 2013) (Kretz, 2014). We have simulated the scenario with three
different values for the parameter pair (g, h) which controls the dynamic
AET 2013 and contributors

13

potential method. The results are shown in TABLE 2. It shows three

remarkable things: first, the dynamic assignment method for this particular
scenario is able to generate route choice behaviour close to user equilibrium.
Second, the average travel times are slightly smaller than for iterated
assignment. This is because pedestrians on route 2/5 walk more efficiently
around the corner of the large obstacle as can be seen in FIGURE 8. This in
consequence causes that there is no need to walk along route 4. And third:
user equilibrium and system optimum obviously do not coincide.

g
1.5
2.0
2.5

h
0.7
1.0
0.0

Route (1, 3, 6)
t.t.
r.c.r.
60,6
0,421
62,9
0,414
60,0
0,389

Route (2, 5)
t.t.
r.c.r.
55,0
0,579
57,4
0,586
61,2
0,611

Mean
t.t.
57,4
59,7
60,7

TABLE 2: Results for simulations with dynamic potential (one-shot assignment) instead of
iterated assignment. Route 4 was not utilized.

FIGURE 8: Still image from a simulation run with dynamic potential (g=2.5).

That with the dynamic potential realistically many pedestrians choose route 2 /
5 and that the simulation is even close to user equilibrium shows that with the
dynamic potential the obstacles can be handled on the operational level. This
is different in the next example as there the obstacles are too large.
1.2. Example 2: a Ring Road
FIGURE 9 shows a scenario which is rather corridor- than room-dominated.
As initial distribution we have set for once a heavy load on the routes along
the globally shortest connection (routes 1, 3, and 4) and in a second run an
equal distribution on all routes.
FIGURE 10 shows the results for an iterated assignment in the geometry of
FIGURE 9 with a demand of 16,000 pedestrians per hour. Travel times for all
pedestrians arriving at the destination in the time interval 300..900 sec were
considered. It can be seen that the total average travel time at equilibrium is
the same (88.3 sec) independent of the route choice ratios at the first
iteration. Similar holds for the route choice ratios of the three detouring routes
AET 2013 and contributors

15

2 (9.9% resp. 10.3%), 5 (14.2% resp. 14.4%) and 6 (0%). However the three
routes 1, 3, and 4, which all lead along the globally shortest path, end up with
largely different route choice ratios. This is a sign that the method works as
intended as obviously the intermediate destinations are shaped such that they
as intended locally have no impact on the travel time. To show that clearer
in FIGURE 11 the sum of route choice ratios and the average of travel times
of route 1, 3, and 4 is shown next to the ones of routes 2, 5, and 6. In this way
it can be better seen that the two starting states lead to the same result.

FIGURE 9: Example geometry with calculated intermediate destinations (blue) and routes with
their IDs (yellow). Note the bottlenecks along the straight connection between origin (red) and
destination (green). From left to right the model extends over 106 m. The corridors are 4 m wide,
respectively 2 m at the two bottlenecks. Note that route 6 implies counterflow with routes 1, 3,
and 4 on the short middle piece.

FIGURE 10: The two figures at the top show how (on the left side) the weighted average travel
time and (on the right side) the weighted standard deviations of travel times evolve in the course
of iterations. The blue line shows the evolution when the initial distribution is mainly on the
globally shortest path and only one percent is assigned to each of the three detouring routes.
The red line shows the evolution when the initial distribution is equal on all routes. The lower left
figure shows the evolution of route choice ratios for each route and for both cases of initial
distribution (lines for equal distribution are marked with a cross). The figure on the lower right
shows the evolution of travel times for each route over the course of iterations.

AET 2013 and contributors

17

FIGURE 11: Route choice ratio (left) and travel times (right) with routes 1, 3, and 4 averaged or
added, respectively.

Finally we have started the iterations with a vast majority of pedestrians

walking route 6, to see if there is another stable equilibrium where the short
part in the center is used in opposite walking direction, but this was not the
case. The iteration process let to very similar results as the ones presented.
1.3. Example 3: Two Times Two Doors
FIGURE 12 shows a scenario where pedestrians have to enter a room
through one of two doors and leave it again through one of two doors on the
opposite site.

FIGURE 12: Walls are shown red with white edge, walking areas dark grey, the origin is to the left
and the destination to the right; light grey areas are not accessible. Note that the doors have
different width: in the first row of walls the (in pedestrian walking direction) left door has a width
of 3 m and the right door a width of 1 m, in the second row it is vice versa. The method
computing the intermediate destinations and the routes gives five more routes which have nor
(one route) or only one (four routes) intermediate destinations. We have not considered them in
the assignment process for this example as they bring no new navigation variants. Note that for
reasons of clarity we have edited the backside of some of the intermediate destinations to have
less of them overlap. Modifying the backside does not change the routing behaviour.

In principle there can be crossing flows in this scenario. However, from basic
considerations one would expect that on route 7 there will be no pedestrians,
as it is one of the two long routes and it runs through both narrow doors. This
is confirmed in the assignment process as the results in FIGURE 13 and
FIGURE 14. In this case the assignment was started with an equal distribution
on all routes.

AET 2013 and contributors

19

A visual comparison of the assignment result with a simulation with dynamic

potential as it is shown in FIGURE 15 shows that in this case the dynamic
potential is not able to produce a balance as precise as the assignment
process does, although it obviously shifts route choice behaviour strongly
toward equilibrium (compared to the shortest path movement paradigm, resp.
static potential approach).

FIGURE 13: Travel times on routes (left) and route choice ratios.

FIGURE 14: Mean travel times of all routes and standard deviations in seconds.

AET 2013 and contributors

21

FIGURE 15: Still image from the simulation at t= 600 s. The upper figure shows the final iteration
of the assignment process, the middle figure a simulation with dynamic potential (g=2.5) and the
lower figure shows what would happen, if one simply let all pedestrians walk the shortest path
(fewest meters) towards destination (just one route for all, leading directly to the final
destination; static potential approach).

5. SUMMARY
We have presented three pilot studies of dynamic assignment for a
microscopic pedestrian simulation which utilizes the outcomes of a novel
algorithm to compute relevant routing alternatives in the 2D pedestrian
movement environment. The results promise a general applicability of the
method. The major task for the future will be the development of a refined,
assignment method working efficiently with pedestrian microsimulation.
6. LITERATURVERZEICHNIS
Cascetta, E., Nuzzolo, A., Russo, F., & Vitetta, A. (kein Datum). A modified logit route choice
model overcoming path overlapping problems: Specification and some calibration
results for interurban networks. Proceedings of the 13th International Symposium on
Transportation and Traffic Theory, (S. 697--711).
Kretz, T. (2013). Multi-Directional Flow as Touch-Stone to Assess Models of Pedestrian
Dynamics. Annual Meeting of the TransportationResearch Board 2013, (S. 13-1160).
Kretz, T. (2014). The Effect of Integrating Travel Time. In U. Kirsch, U. Weidmann, & M.
Schreckenberg (Hrsg.), Pedestrian and Evacuation Dynamics 2012. Zrich: Springer
Berlin / Heidelberg.
Kretz, T., & Groe, A. (2012). From Unbalanced Initial Occupant Distribution to Balanced Exit
Usage in a Simulation Model of Pedestrian Dynamics. Human Behaviour in Fire
Symposium, (S. 536--540).
Kretz, T., Groe, A., Hengst, S., Kautzsch, L., Pohlmann, A., & Vortisch, P. (2011). Quickest
Paths in Simulations of Pedestrians. Singapore: World Scientific Publishing.
Kretz, T., Hengst, S., & Vortisch, P. (2008). Pedestrian Flow at Bottlenecks-Validation and
Calibration of Vissim's Social Force Model of Pedestrian Traffic and its Empirical
Foundations. 8th International Symposium on Transport Simulation (S. on CD).
Surfer's Paradise, Queensland, Australia: Monash University.
Kretz, T., Hengst, S., Roca, V., Prez Arias, A., Friedberger, S., & Hanebeck, U. (2011).
Calibrating Dynamic Pedestrian Route Choice with an Extended Range Telepresence
System. 2011 IEEE International Conference on Computer Vision Workshops, (S.
166-172).
Kretz, T., Lehmann, K., & Hofs, I. (submitted). User Equilibrium Route Assignment for
Microscopic Pedestrian Simulation. submitted.

AET 2013 and contributors

23