TOWARDS CONTEXTUAL GOAL-ORIENTED PERCEPTION FOR

PEDESTRIAN SIMULATION

Laure Bourgois, Julien Saunier and Jean-Michel Auberlet

Universit

e Paris Est, ISFTTAR-IM-LEPSIS, 58 Bld Lef

ebvre, 75015 Paris, France

Keywords:

Simulation, Evaluation, Perception.

Abstract:

Perception is often seen in multiagent systems and in robotics from a passive point of view. The sensors of the

agent collect information on its environment; however the potentially important number of percepts is not real-

istic and may decrease the agents efﬁciency. In this article, we introduce a contextual goal-oriented perception

ﬁltering. Besides the lack of plausibility of omniscient agents, it addresses the problem of transmitting too

much information to the agents. This goal-oriented perception module is evaluated model in terms of validity

of the resulting behavior and of time complexity.

1 INTRODUCTION

The use of trafﬁc simulation tools to design and as-

sess the infrastructure is widespread for highways or

interurban networks. However, several issues remain

to apply it realistically, in urban areas, because the

environment is more complex. In particular, pedes-

trians are difﬁcult to take into account, since their

behavior is less constrained and normative than the

behavior of the vehicles. Current pedestrian simula-

tion tools do not consider heterogeneity and dynam-

ics of the environment sufﬁciently, especially for road

crossing or sidewalk following. This paper describes

a pedestrian model that takes into account the interac-

tion with other pedestrians (on sidewalks) and vehi-

cles (for road crossing). This model is based on per-

ception operations which are inputs of decision and

action components.

Perception is often seen in multiagent systems

from a passive point of view. The sensors of the agent

collect information on its environment; however the

potentially important number of percepts is not realis-

tic and may decrease the agents efﬁciency. Moreover,

an agent does not always need a complete information

about its environment. The complexity of simulation

models and the amount of information clearly leads to

a bottleneck for the simulation of thousands of pedes-

trians.

The perception can be seen as a three-part pro-

cess (Weyns et al., 2004): an agent senses the state of

the environment to get a raw perception, then it inter-

prets the raw perception to transform the data into per-

cepts (consistent with its internal representation of the

world), and ﬁnally it ﬁlters these percepts to keep only

relevant percepts (ﬁltered perception). In this frame-

work, we propose a reﬁnement of the classical percep-

tion model that improves the ﬁltering to keep only the

relevant information. The ﬁlters are goal-driven and

depend on the current task of the agent. They also

take into account the context of the agent. Besides the

lack of plausibility of omniscient agents, the problem

of transmitting too many information to the agents is

that of scalability. Accessing and treating this infor-

mation is combinatorial in function of the number of

agents. Henceforth, many simulation tools include ad

hoc ways to limit the perception, but to the best of our

knowledge they are rarely evaluated.

In section 2, a non exhaustive state of the art on

microscopic pedestrian models is presented, stressing

on the computational features of perception and nav-

igation space representation. Section 3 presents our

pedestrian model, which enable the agents to move in

urban areas thanks to two different models (one for

moving and the other for decision making).Section 4

deals with the navigation model. In this case, the ﬁeld

of vision depends on the agent dynamic context. We

present our model and assess both the validity of the

model and its computation cost. Section 5 presents

the road-crossing decision model. Here, the percept

ﬁltering depends both on the context (trafﬁc ﬂuidity)

and on the agent’s goal (crossing decision). We show

some results of our simulations in terms of treatment

complexity gains. To conclude, we present our works

outlook.

197

Bourgois L., Saunier J. and Auberlet J..

TOWARDS CONTEXTUAL GOAL-ORIENTED PERCEPTION FOR PEDESTRIAN SIMULATION.

DOI: 10.5220/0003743601970202

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2012), pages 197-202

ISBN: 978-989-8425-96-6

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

2 RELATED WORKS

The Social Force Model (SFM) (Helbing and Molnar,

1995) is the most used in crowd simulation commu-

nity. A pedestrian agent acts as if it was subject to at-

tractive forces (its destination ...) and repulsive forces

(pedestrians ...). It tries to trail straight ahead in order

to reach its destination following a shortest path route,

and avoiding obstacles. Helbing interprets the private

sphere notion as ”repulsive sustained ﬁeld” sum com-

puted in function of environment obstacles and an at-

tractive ﬁeld function of a wanted walking speed and

a given destination (and possible ”attractive” pedes-

trians to form groups (Moussa

ıd et al., 2010)). This

model has several limitations; due to its reactive fea-

ture, a pedestrian agent avoids obstacles at the last

moment, moreover it has an unlimited depth of the

ﬁeld of vision that leads to a perception complexity in

O(n

) (with n the number of pedestrians).

Hoogendoorn’s model (Hoogendoorn and Bovy,

2000) partially takes up SFM and adds control theory

concepts to it. A pedestrian agent optimizes an utility

function with a set of different costs for acceleration,

private space sphere and path diversion.

In graphical computer community, Reciprocal Ve-

locity Obstacle model (Guy et al., 2010) is widely

used and shows the advantage of avoiding the clas-

sical phenomenon of oscillation between pedestrian

(major shortcoming of SFM model). This model takes

into account for a pedestrian agent the reaction time

and the altering pedestrian speed time. Authors code

pedestrian agents trajectory as a linear programming

problem. As the SFM, these models use a continuous

space representation with at least a O(n

) complex-

ity (n the number of pedestrians) (Hoogendoorn and

Bovy, 2000).

However, some other environment modelings ex-

ist with different shape cells division (Blue and Adler,

2001) or data structures. These cells enable to geo-

metrically characterize a set of zone perceived by the

pedestrian agents. Their major advantage lies in the

decreasing of the perception complexity. Indeed, the

complexity of Reynolds initial method is O(n

) for

boids modeling (n the number of boids); it falls to

O(kn) with a discrete space (k the number of cells)

(Reynolds, 2000). Nevertheless, this approach has a

certain time computation cost, merely, it is executed

off line. The method has the major shortcoming of a

constant perception modeling, whereas a pedestrian

does not necessarily pay attention to every nearby

zone with the same intensity.

Collision avoidance is generally based on the per-

ception in order to detect moving and/or ﬁxed ob-

stacles (other pedestrians, vehicles, walls). Although

perception has long been considered as an undergone

state, it has emerged recently that awareness is an ac-

tive state, and not only the result of stimuli. The per-

ception is directed towards a subject via a receptor.

The detection of the obstacles is executed at each

time step. Therefore, in order to simulate several

thousands of agents, the perception process should be

implemented with a low complexity algorithm. We

choose to characterize the complexity of the interac-

tion between agents as the number of perceptual op-

erations that an agent must perform to interact with

another agent, whereas we consider that the agent’s

decision time and the agent’s decision execution time

are constant.

3 PEDESTRIAN AGENT MODEL

As our model does not depend on a space represen-

tation, we consider in this article a continuous space

(the worst case in term of complexity). Besides, a

continuous-space model is well suited for simulating

real pedestrians and allows to know for each step the

exact position of each agent. Moreover, we focus

on the inputs of perception and decision components,

hence we do not go into details regarding action algo-

rithms (more details in (Bourgois et al., 2010)).

We use the simplifying assumption that the pedes-

trians do not try to form groups. We do not focus on

the lane formation self organization.

Our model consists of two models, one for mov-

ing and one for road-crossing decision. In this section,

we will ﬁrst present the general characteristics of an

agent then we will detail each model.

3.1 Pedestrian Agent Characteristics

Following Helbing recent model (Johansson et al.,

2007), a pedestrian is represented by a circle with

a variable radius. Each agent has a set of variables

and parameters P for the pedestrian interactions that

mainly are : its obstacle reaction intensity A and a

reactivity threshold B, over which the agent does not

react anymore to environmental elements.

We add here I the set of characteristics necessary

to model pedestrian-vehicle interactions: a vehicle

perceptual ﬁeld, that is to say the depth of the ﬁeld

of vision on the infrastructure (e.g. the street) and a

crossing willingness, which depends on the environ-

mental context perceived by the agent.

3.2 Models Involved

We follow Hoogendoorn’s model concepts (Hoogen-

ICAART 2012 - International Conference on Agents and Artificial Intelligence

198

doorn and Bovy, 2000) that distinguish 3 decision lev-

els: strategic, tactical, and operational. We assume

that the route of an agent (its origin and destination)

is already known (strategic level). The model is there-

fore composed of: a moving model for obstacle avoid-

ance (operational level) and a decision model for road

crossing (tactical level).

The moving model takes place within the frame-

work of perception-action, and the decision model

lies in the perception-decision framework. To ensure

the continuity of pedestrian movement (along a side-

walk, pedestrian crossing), we must articulate these

two models. If an agent always tries to avoid all the

obstacles, the priority of an agent crossing a street is

above all to avoid vehicles. When the agent decides

to stop at the curb to allow vehicles to pass, it does

not apply the moving model.

We brieﬂy detail the behavior of an agent accord-

ing to its position (sidewalk / road) and its state. For

each situation, we distinguish two main functions: ﬁ-

nal destination computation and navigation. The fol-

lowing Navigation algorithm is given here:

IF local destination is not reached

use navigation model

ELSE IF next destination is a crosswalk

Find a gap to cross the road

IF Gap is found

use navigation model

ELSE

calculate number of steps n while crossing

is predictably unfeasible

wait n steps

Once an agent reaches a sidewalk, it computes its lo-

cal destination. To avoid artiﬁcial pedestrian conver-

gence on the same geometrical destination, an agent

randomly calculates its own local destination corre-

sponding to an area surrounding the given local desti-

nation.

4 SHORT-TERM NAVIGATION

4.1 Moving Model (Operational Level)

Perception is based on the depth of the ﬁeld of vi-

sion and the sensibility parameter (that determines

the anisotropy of the agent). In the initial perception

model, an agent α may be inﬂuenced by the agent β,

if is in the ﬁeld of vision of agent α. This inﬂuence is

represented by a force exerted on α by β. This force

depends on the sensibility of α to β and on g(d

αβ

) the

distance dependant reaction of agent α to β which is

deﬁned as :

g(d

αβ

) = A

−d

αβ

(1)

where R

, R

are the radii of the agents and d

αβ

is the

distance between their centers.

We see that the force decreases when the distance

between the pedestrians increases. We remark here

that perception does not take into account the speed

nor the course of the agents α and β, but only their re-

spective locations. Helbing did not mention the depth

of the ﬁeld of vision, we thus added a variable in or-

der to perceive only the neighbors. We calibrated A

and B with the same value as recent SFM.

In the initial moving model, each agent has a de-

sired speed, and tries to walk in a (desired) direction,

which depends on its current position and its desti-

nation. It also tries to maintain its speed to its de-

sired speed, which is therefore the target for the agent

to reach. Agent acceleration is then deﬁned by: its

actual speed, its impatience and the repulsive forces

to avoid obstacles (pedestrians, walls) and to keep its

private sphere.

In previous works, we enhanced this model (Bour-

gois et al., 2010) with sensitivity measures and by in-

troducing a depth of the ﬁeld of vision for pedestrian-

pedestrian interactions. We have proved that the depth

of the ﬁeld of vision inﬂuences the pedestrians’ ﬂows.

In those research, we tested a depth that is a good

trade-off collision avoidance and speed preservation.

The depth of the ﬁeld decreases the number of

perception operations. This perception efﬁciency im-

provement comes from the ﬁltering of the perception.

We now add a dynamic adaptation of the depth

of the ﬁeld of perception to acquire only the relevant

data depending on the agent context.

Relevance distance We compute the depth of the

ﬁeld of vision depending on the density of agents in

the neighborhood of the agent. Indeed, we observed

in our simulations that the high density of agents tends

to reduce the agent acceleration to zero. Based on the

intuitive idea that in congestion situations, a pedes-

trian perceives only pedestrians nearby, we propose

to modify the depth of the ﬁeld of vision in function

of the density of agents surrounding the agent.

To reproduce this phenomenon, we introduce a ﬁl-

tering function that allows relevant agent selection ac-

cording to the density of agents. This function con-

sists in a distance computation from which the neigh-

bors are taken into account for the acceleration com-

putation. In this way, if the number of neighbors is

low, the agent takes into account distant neighbors,

while if the density is large, it only considers the near-

est neighbors. The following ﬁltering function has

been determined empirically:

= (e

neighbor

− 1) ∗ α (2)

with D

the relevance distance, nb

neighbor

the raw

number of neighbors perceived by the agent and α a

TOWARDS CONTEXTUAL GOAL-ORIENTED PERCEPTION FOR PEDESTRIAN SIMULATION

199

factor which has been calibrated to 10 × e.

Thanks to this formula, the agent relevance dis-

tance variates according to the number of pedestrians

present in its ﬁeld of perception. We initialize the pa-

rameters thanks to the results found in our previous

study : the depth of the ﬁeld of vision ranges from

1,65m to 9m.

From the relevance distance, we compute theoret-

ically the maximum number of agents that could be

present (i.e. treated) surrounding the given agent for

that distance. It is the relevant-distance area divided

by the maximum surface occupied by one pedestrian

(even though the maximum possible density, reported

to be 5.4 ped/m

(Weidmann, 1993) is use, the number

of perceived pedestrian is bounded). After a threshold

(15), the ﬁeld of vision becomes too little to contain

more than 12 agents. This result can be explained by

the underlying formula: when there are few agents in

the system, the ﬁeld of vision is wide, hence enabling

the agents to perceive all the other agents. However,

the ﬁeld of vision diminishes when the number of

agents increases.

The operational complexity of perception is O(kn)

where n is the number of simulated agents and k the

maximum number of agents that can ﬁt in the neigh-

borhood of an agent. We now compare these theoret-

ical computations to empirical results.

4.2 Experimentation and Results

In order to validate and verify the consistency of our

model, we executed several series of simulations. For

these simulations, we have recreated the standard con-

ditions described in (Lobjois and Cavallo, 2009).

The next sections present our simulation results

in order to quantify the validity and efﬁciency of the

perception function. In the ﬁrst case, we simulated a

bidirectional corridor, which is similar to a sidewalk:

pedestrians enter the corridor at one end and leave it at

the other end. The size of the corridor is 10m × 60m.

The ﬂow (ped/h) is the number of agents created at

each end of the corridor.

Perceptual Improvements. For each simulation of

bidirectional corridor, we recorded the average inten-

sity difference between the force sustained with a raw

perception and the one with a ﬁltered perception. In

this way, we measure the difference in behavior in-

duced by the ﬁltering process. We see in Figure 1

that for high ﬂows, the pedestrian acceleration force

with a ﬁltered perception is weaker than the one with

a raw perception; although the maximum difference

is 0.4dm/s

(at worst a 3% average speed deviation).

Therefore, the behaviors are not or slightly modiﬁed

by the active perception. A second validation aims

Figure 1: Average intensity difference (2000, 4000 ped/h).

to compare our simulation with a global behavior of

pedestrian ﬂows. For that aim, we have aggregated

the individual data and compared with the existing

data like fundamental diagrams (FD) (which describe

the relation between mean speed, density and ﬂow of

pedestrians).

Unfortunately, literature concerning pedestrian

FD is poor, mainly because of the difﬁculty to ac-

quire data. Our experimental results shows that our

model approximates correctly other aggregate mea-

sures found in the literature (Weidmann, 1993).

Improvements in Bidirectional Corridors. Figures

2 present for each minute of simulation, the records

for all agents, of: the average and maximum number

of agents perceived (raw perception) and the average

and maximum number of agents treated (ﬁltered per-

ception).

For a low ﬂow, the average number of agents per-

ceived and treated are comparable, the difference be-

tween the maximum number of perceived and treated

agents is negligible. Indeed, with a low density, the

relevance distance is always close to the maximal

distance of perception (9m). The efﬁciency gain is

marginal but always positive.

The more the ﬂow increases (Figure 2) , the more

the difference between raw and ﬁltered perception in-

creases. For 2000 agents, the ﬁltered perception treats

only 30% of the raw perception.

Figure 2: Flow of 2000 ped/h.

In the previous section, we showed that the ﬁlter-

ing function did not deteriorate noticeably the behav-

ior of the agents. The simulation results show that the

ﬁltering function is very efﬁcient in mid to high ﬂows

and that the number of percepts an agent has to treat

decreases strongly (−70 to −90%).

ICAART 2012 - International Conference on Agents and Artificial Intelligence

200

5 ROAD CROSSING

5.1 Decision Model (Tactical Level)

This model involves two models: Tom’s model

(Auberlet et al., 2011) and a ”gap acceptance” model

(acceptance of time slot for road crossing).

Tom’s model is based on the environment percep-

tion. It establishes a set of Boolean variables, each

corresponding to a visual cue used cross. The model

allows to predict with a reliability close to 70% the

decision in these contexts. We interpret the output of

this model as a crossing willingness This value mod-

ulates a ”gap acceptance” model (the pedestrian de-

cides to cross if the time slot offered by the trafﬁc is

higher than its own estimate of the crossing time).

For the decision model, the agent gets one or

several vehicle queues moving along the road. The

agent selects the ﬁrst vehicle that can either let it pass

through or block it. Other vehicles are not treated.

This ﬁlter is thus based on a prediction : the time to

collision (TTC) between the pedestrian and the ve-

hicle, if all agents maintain their actual speeds. We

point out this ﬁlter is different from the relevance ﬁlter

(based on the actual situation, the density of agents).

This prediction is based on the assumption that one

can anticipate the movement of perceived vehicles

(a standard procedure in agent-oriented or distributed

simulation).

We distinguish two types of context for vehicles

perception: free trafﬁc ﬂow and congested trafﬁc ﬂow.

In case of congested trafﬁc ﬂow, the agent avoids

the vehicles on the roadway as ﬁxed obstacles. In

this situation, the agent only perceives the station-

ary vehicles between which it wants to get through.

Hence, the perception depends on the current goal of

the agent and of its context (free or congested trafﬁc).

In a free trafﬁc ﬂow context, if the agent raw percep-

tion is an upstream queue of vehicles, it selects only

the vehicle with which it might interact.

The perception and decision functions must take

into account: ”side time”: time taken by the pedes-

trian to reach the vehicle’s side (in function of the dis-

tance between the side of the vehicle and the pedes-

trian), ”lag before” (or front lag) (time for the vehicle

front bumper to reach the side of the pedestrian), ”lag

behind” (time for the vehicle rear bumper to be com-

pletely beyond the pedestrian).

We notice that if the agent is waiting on the curb,

it does not longer need to perceive. We can compute

the time it has to wait before deciding to cross again

with the no perception condition.

WHILE Lag behind < Side time

wait without perception

This function decrease the computation time because

the agent no longer has to perceive and to treat infor-

mation while it is stopped.

5.2 Tactical Perception Complexity

Each time an agent decides to cross, it must perceive

a set of lanes and for each a set of vehicles. The num-

ber of perceived vehicles depends on: the number of

lanes, m and the ﬁeld of perception, implying a maxi-

mum number of vehicles perceived, k.

Taking into account exceptional situations with a

very high trafﬁc density and assuming that there is no

perception ﬁlter, the maximum number of perceived

vehicles for a crossing decision is O(mk).

5.3 Results

The vehicles are modeled as agents which have a ba-

sic car following task behavior. We use Ketenci’s

work (Ketenci et al., 2010) to animate these agents.

However, here we do not focus on the vehicle coor-

dination task on roads intersections. Currently, the

vehicle agents only perceive each other.

In order to study plausible situations, we initial-

ized the parameters with values found in the literature

(Lobjois and Cavallo, 2009). The vehicle agents are

created with the following features: Time Headway

(THW) between the vehicles, distributed according

to the normal law on [1, 2, ..., 8]s, speed is distributed

uniformly on the intervals [38 − 42], [48 − 52], and

[58 − 62] km/h. We simulated the road crossing of

a 285m long unidirectional road (Staplin, 1995). We

simulate two different vehicle ﬂows: 400 and 600 ve-

hicles per hour.

Behavior Validity. For each simulation, we recorded

for each decision taken by an agent: the speed of the

selected interacting vehicle and the TTC (in seconds).

Indeed, referring authors take into account an initia-

tion crossing time (the pedestrian start to move on the

sidewalk toward the road whereas there is still a vehi-

cle in front of him). Actually, a pedestrian take also

into account the time of arrival of the upstream vehi-

cle, i.e., the TTC.

We observed that the mean values of TTCs belong

to the following range of values : [5.02, 7.8]. In our

simulations, the crossing behavior is consistent with

the time slots observed. Indeed, for the time slots

taken by the agents, more than 95% of pedestrians

would cross in the reality.

Improvements in Street Crossing Situations. We

recorded for each agent: the number of decision(s)

condition it has to make to cross the road, with the

possibility to block the perception (no perception con-

TOWARDS CONTEXTUAL GOAL-ORIENTED PERCEPTION FOR PEDESTRIAN SIMULATION

201

Table 1: Average gain of decision depending on the average

vehicle speed.

Gain

Vehicle speed

30 40 50

Flow of 400 veh/h 6,79 5,56 5,67

Flow of 600 veh/h 15,62 11,88 9,25

dition) when the agents knows it will not be able to

cross because of an incoming vehicle, the number of

decision(s), no condition it would have to make with-

out the no perception condition.

Table 1 presents the average gain of perception-

decision cycles with the no perception condition : it

is the difference between the average of condition and

no condition for all the agents. We see an impor-

tant gain, even for high speeds. It is not surprising

that this gain also increases with a higher ﬂow, since

the agent has less opportunities to cross. The road-

crossing model diminishes the number of treatments,

both in terms of road-crossing decisions and of num-

ber of vehicles perceived, while keeping acceptable

behavior results.

6 CONCLUSIONS

We have presented a pedestrian agent model encom-

passing both action (locomotion) and decision (street

crossing). Both tasks are highly dependent on the per-

ception of the situation, but do not rely on the same

pieces of information. Hence, we have shown that the

reﬁnement of the perception, based on a goal-oriented

ﬁltering of the percepts, enables the decrease of the

number of objects treated by the agents while keep-

ing consistent results in term of behavior.

The difference between the raw perception and the

ﬁltered perception has been evaluated and depends

on the model: from 70% to 90% for the locomotion

model. The ﬁltering we have proposed is not an im-

plementation artefact to ameliorate the time complex-

ity of our agents, but a part of the human cognitive

process. We note that humans can only acknowledge

and handle a limited amount of information. Future

works needs to be done to establish which percepts

must be kept to ameliorate the perception process.

More research is needed to better understand and

reproduce a group behavior (Moussa

ıd et al., 2010),

on particular in the street crossing case. Donikian

notes that groups show bolder behavior by accepting

shorter time gaps and by forcing the vehicle drivers to

slow down (Donikian, 2004). Several problems arise

(group identiﬁcation and formation). Another inter-

esting possibility to model large areas would be to

switch from a microscopic to a macroscopic model

at run-time.

REFERENCES

Auberlet, J.-M., Bodard, V., Bourgois, L., Br

emond, R.,

esir

e, L., Maestracci, M., Grani

e, M.-A., Rabier, R.,

and Tom, A. (2011). Simulation de travers

ee de car-

refour par des pi

etons. Technical report, Fondation

ecurit

e Routi

ere, projet SiCAP. (to be published).

Blue, V. J. and Adler, J. (2001). Cellular automata mi-

crosimulation for modeling bi-directional pedestrian

walkways. Transportation Research Journal Part B.

Bourgois, L., Oulhaci, A., and Auberlet, J.-M. (2010). Sim-

ulation de d

eplacement de pi

etons : Vers un mod

ele

de perception et de pr

ediction d’action chez autrui. In

Journ

ees Francophones des Syst

emes Multi-Agent.

Donikian, S. (2004). Mod

elisation, contr

ole et animation

d’agents virtuels autonomes

evoluant dans des en-

vironnements inform

es et structur

es. Habilitation

Diriger des Recherches, Institut de Recherche en In-

formatique et Syst

emes Al

eatoires.

Guy, S. J., Lin, M. C., and Manocha, D. (2010). Modeling

collision avoidance behavior for virutal humans. In

9th International Conference on Autonomous Agents

and Multi-agent Systems.

Helbing, D. and Molnar, P. (1995). Social force model for

pedestrian dynamics. Physical Review E.

Hoogendoorn, S. and Bovy, P. (2000). Gas-kinetic model-

ing and simulation of pedestrian ﬂows. Transportation

Research Record.

Johansson, A., Helbing, D., and Shukla, P. (2007). Speciﬁ-

cation of the social force pedestrian model by evolu-

tionary adjustment to video tracking data. Advances

in Complex Systems.

Ketenci, U., Br

emond, R., Auberlet, J.-M., and Griselin,

E. (2010). Bounded active perception. In Proceedings

EUMAS, European Workshop on Multi-agent Systems.

Lobjois, R. and Cavallo, V. (2009). The effects of aging

on street-crossing behavior: From estimation to actual

crossing. Accident Analysis and Prevention.

Moussa

ıd, M., N.Perozo, Garnier, S., Helbing, D., and

Theraulaz, G. (2010). The walking behaviour of

pedestrian social groups and its impact on crowd dy-

namics. In PLoS ONE.

Reynolds, C. (2000). Interaction with groups of au-

tonomous characters. In Game Developers Confer-

ence.

Staplin, L. (1995). Simulator and ﬁeld measures of driver

age difference in left-turn gap judgments. Transport

Research Record.

Weidmann, U. (1993). Transporttechnik der fussg

anger,

transporttechnische eigenschaften des fuss-

gngerverkehrs (literturauswertung). Technical

report, Schriftenreihe des IVT Nr. 90, Zweite,

erg

anzte Auﬂage, Z

urich.

Weyns, D., Steegmans, E., and Holvoet, T. (2004). Towards

active perception in situated multiagent systems. Spe-

cial Issue of Journal on Applied Artiﬁcial Intelligence.

ICAART 2012 - International Conference on Agents and Artificial Intelligence

202