Documente Academic
Documente Profesional
Documente Cultură
3159
AbstractThis paper proposes a novel approach for simulating pedestrian movement behavior based on artificial intelligence
technology. Within this approach, a large volume of microscopic
pedestrian movement behavior types were collected and encapsulated into an artificial neural network via network training. The
trained network was then fed back into a simulation environment
to predict the pedestrian movement. Two simulation experiments
were conducted to evaluate the performance of the approach.
First, a pedestrian-collision-avoidance test was conducted, and the
results showed that virtual pedestrians with learned pedestrian
behavior can move reasonably to avoid potential collisions with
other pedestrians. In addition, a critical parameter, i.e., defined
as reacting distance and determined to be 2.5 m, represented
the boundary of the collision buffer zone. Second, a pedestrian
counterflow in a road-crossing situation was simulated, and the
results were compared with the real-life scenario. The comparison
revealed that the pedestrian distributions, erratic trajectories,
and densityspeed fundamental diagram in the simulation are
reasonably consistent with the real-life scenario. Furthermore, a
quantitative indicator, i.e., the relative distance error, was calculated to evaluate the simulation error of pedestrians trajectories
between the simulation and the real-life scenario, the mean of
which was calculated to be 0.322. This revealed that the simulation
results were acceptable from an engineering perspective, and they
also showed that the approach could reproduce the lane-formation
phenomenon. We considered the proposed approach to be capable
of simulating human-like microscopic pedestrian flow.
Index TermsPedestrian flow, artificial intelligence, neural networks, lane formation.
I. I NTRODUCTION
Manuscript received June 7, 2015; revised December 24, 2015 and March 6,
2016; accepted March 11, 2016. Date of publication May 4, 2016; date of
current version October 28, 2016. The work described in this paper was fully
supported through a grant from the Research Grant Council of the Hong Kong
Special Administration Region, China (Project CityU 11206714). The Associate Editor for this paper was X. Cheng.
The authors are with the Department of Architecture and Civil Engineering,
City University of Hong Kong, Kowloon, Hong Kong (e-mail: yima23-c@my.
cityu.edu.hk; ericlee@cityu.edu.hk; bckkyuen@cityu.edu.hk).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TITS.2016.2542843
1524-9050 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
3160
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 17, NO. 11, NOVEMBER 2016
A. Input Layer
The input layer is used to accommodate various system
stimuli, the architecture of which is shaped by the system
itself. For the pedestrian traffic system, empirical study results
indicate that the movement behavior of pedestrians in crowds
reflects the perceived environmental stimuli and the properties
of the pedestrians [5], [9], [34][36]. Thus, the elements in
the input layer corresponding to the pedestrian traffic system
should consist of the following parts.
1) Part IInput Parameters Used to Represent the Pedestrians Current Movement Features: Pedestrians usually do not
change their current direction and speed abruptly or frequently
due to the inertia effect [37]. Thus, a pedestrians subsequent
moving direction and speed tend to be consistent with his/her
current moving direction and speed if there is no perturbance.
Likewise, a pedestrians current velocity should be selected as
an input parameter.
2) Part IIInput Parameters Used to Represent Interactions
With Other Pedestrians: As an element of group movement,
pedestrians direction and speed of movement are influenced
byand influencethe moving states of other individuals.
Such interactions between pedestrians are very complex and
difficult to quantify. Fortunately, empirical research results and
numerous simulation cases have indicated that the degree of
interaction is based on the relative positions of and velocity
between individual pedestrians. Some obvious facts in real life
also verify these conclusions. For example, the smaller the
distance between two pedestrians, the stronger the reciprocal
influence on the moving state. These facts and conclusions
indicate that it is reasonable to select the relative positions of
and velocity between pedestrians as input parameters.
3) Part IIIInput Parameters Used to Represent Interactions With Obstacles: Obstacles, such as boundaries and the
fixed obstacles placed in an environment may constrain or
obstruct the pedestrians movement. In fact, the interactions
between pedestrians and obstacles are similar to those between
pedestrians, particularly if we view an obstacle as being similar
to a pedestrian standing still. Thus, the interactions between
pedestrians and obstacles can be represented in a similar way
to the interactions between pedestrians. However, it must be
noted that obstacles are typically still, and the relative velocity
between pedestrians and obstacles equals that of the pedestrians, as introduced in Part I. Thus, to avoid repetition, we only
select the relative position as an input parameter to represent
interactions between pedestrians and obstacles in this part.
4) Part IVInput Parameters Used to Represent Interactions With Desired Targets: Target positions determine the
pedestrians desired moving direction; that is, the pedestrians
desired moving direction always orients on the target if there is
no perturbance. Here, the relative position between the pedestrian and the desired target is selected as an input parameter to
represent the interaction between them.
5) Part VProperties of the Pedestrian: For the heterogeneous pedestrian flow, some of the individuals subjective
properties also influence his/her movement behavior. These
properties include physiological properties such as health level,
energy level and sensing range and emotional properties such
3161
as attraction intensity, panic intensity and social group property [38]. Pedestrians with different individual properties may
engage in different moving behavior under the same environmental stimuli.
Note that although the architecture of the input layer consists
of the abovementioned five input parameter types, it does not
mean that the number of input parameters is fixed. The number
of input parameters can be added or reduced according to
specific situations. For example, for the homogenous pedestrian
flow, the input parameters in Part V can be ruled out. Such
a flexible architecture can guarantee the adaptability of the
network model.
B. Hidden Layer
The number of neurons in the input and output layers is
fixed, determined by the system and equal to the number of
input and output parameters, respectively. However, there is no
unified approach to determining the number of hidden neurons.
Fortunately, many rules have been proposed to assist users in
estimating the number of hidden neurons. In this paper, the rule
developed by Ward System [39] is used to estimate the number
of hidden neurons. (See (1).)
(1)
Nh = (Nin + Nout )/2 + Ns
where Nin , Nh , and Nout represent the number of neurons
on the input, hidden and output layers, respectively, and Ns
represents the number of training samples.
C. Output Layer
Pedestrians microscopic movement behavior consists of a
series of continuous stimulus-response processes. Pedestrians
respond to subsequent moving direction and speed in terms of
the perceived environmental stimuli and to their own properties
in each moment. Thus, the subsequent response to velocity is
regarded as an output parameter in the output layer, which is
unique to this paper.
Thus far, input and output parameters have been specified,
but unlike most of the existing pedestrian flow simulation
models in which the functional relationships between input and
output parameters are provided explicitly in terms of empirical
formulas, we do not define any parameters or give any formulas
to express these functional relationships. Instead, we obtain
these functional relationships via training/learning from the
real-life data about pedestrian movement. This feature ensures
that the captured functional relationships between input and
output parameters are more objective and persuasive. In the
following sections, we show how we captured these functional
relationships and used the developed approach to examine case
studies.
III. S IMULATION OF P EDESTRIAN M OVEMENT B EHAVIOR
Based on the aforementioned introduction for the developed
simulation approach, we conducted the pedestrian movement
simulation on a crosswalk scenario to test it. The flowchart of
the simulation process is presented in Fig. 2.
3162
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 17, NO. 11, NOVEMBER 2016
3163
movement was recorded from a top-down view using a camcorder and then processed into video clips. The recording site
chosen was a crosswalk on a street in Hong Kong. The highrise buildings and dense pedestrian flow in Hong Kong made it
possible to capture sufficient pedestrian movement.
The video clips were used to extract the positions of pedestrians by computer and record their coordinates at every time step.
It is extremely time-consuming to extract such data manually,
but using a fully automatic pattern recognition approach does
not guarantee 100% accuracy. Thus, we used the semiautomatic
approach developed by Teknomo [45]. A simple computer
program was developed to complete this work. The front-end of
the program is as shown in Fig. 5. It played the video clip on a
screen in slow motion while we used a mouse pointer to indicate
the head of a pedestrian on-screen as he/she walked. In this
way, the time stamp and the 2D coordinates of the pedestrian
were recorded on the interface. The video clip was replayed
until the movements of all of the pedestrians in the clip were
captured. However, as the top view of a video is not 100%
horizontally projected, the collected data were geometrically
corrected, converting the 2D coordinates of the pedestrian on
the interface into 2D coordinates in realistic scenarios. Upon the
completion of this semi-automatic data collection process, the
positions of all of the pedestrians at every time step under realistic scenarios were recorded.
The collected data must be processed in advance to cater to
the necessity of network training; that is, the collected data
must be converted to the training samples. To complete this
work, the time step was first determined to calculate the velocity
parameters of each pedestrian at each time step. The rational
time step is crucial to network training because a small time
step may generate a large amount of noise that may reduce the
performance of network training, and a large time step may
be disadvantageous in describing the variety of velocity, as
the velocity direction may become increasingly narrow as the
time step increases. As such, we chose a time step of 0.48 s to
balance the performance and variety of moving velocity. Based
on the chosen time step, 4496 training samples were generated.
The input end of each sample (25 input parameters) represented
3164
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 17, NO. 11, NOVEMBER 2016
C. Network Training
The collected samples were used to train the developed
approach. To complete this work, the Backpropagation [46]
(BP) training algorithm was used for its powerful adaptivity.
It calculated the errors between the training and target outputs
based on pre-set weights and activation functions, then adjusted
the weights of the links between the neurons to minimize the
errors. Next, a new round of training was conducted to calculate
the new training errors based on updated weights and adjust the
weights again. This procedure was repeated numerous times
until the training errors were reduced to an acceptable scope.
To prevent overfitting, the early-stop validation approach was
used to supervise the BP training. It splits off parts of the
samples as the validation sets from the samples, which are
trained using an intermediate-state trained network to obtain
validation errors. Such a validation error is used to detect
the overfitting. Specifically, the network training is terminated
when no reduction in validation errors is detected over a certain
number of epochs (15 in this paper). Eventually, the network
state with the minimum number of validation errors was taken
as the ultimate trained network.
The trained network was also applied to testing sets split
from the sample database to evaluate the training performance.
R-squared, as defined in (2) between target output T and
training output P was selected as the performance evaluation
indicator
R2 = 1
n
i=1
(Ti Pi )2 /
n
(Ti T )2
(2)
i=1
f (x|, ) =
(3)
3165
3166
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 17, NO. 11, NOVEMBER 2016
(4)
i
Pm
is subject to:
i
Pm P j Rmin for j(j = i)
i
P Li Dmax
m
(5)
(6)
3167
Fig. 11. Comparison of pedestrian distributions at the 11th, 21th and 31th time step, respectively. The black and white circles represent pedestrians moving to
the right and left, respectively. The number in each circle represents each individuals serial number. (a) Pedestrian distribution in the simulation. (b) Pedestrian
distribution in the corresponding real-life scenario.
be more urgent and have a higher speed in the real life for
crossing the road within limited green time.
Moreover, another quantitative evaluation indicator, namely,
the relative distance error [44], was calculated to better illustrate
the reliability of the developed approach. The relative distance
error at time step t for pedestrian is defined by:
Fig. 12. Comparison of pedestrian trajectories. The blue and red lines represent
pedestrians moving to the right and left, respectively. (a) Pedestrian trajectories
in the simulation. (b) Pedestrian trajectories in the corresponding real-life
scenario.
simulated
P
(t + t, ) P tracked (t + t, )
E(t, ) =
P tracked (t + t, ) P tracked (t, )
(7)
3168
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 17, NO. 11, NOVEMBER 2016
where P tracked (t, ) represents the actual coordinate of pedestrian a at time step t in the real-life scenario; P tracked (t +
t, a) represents the actual coordinate of pedestrian a after one time step t in the real-life scenario; the values of
P tracked (t, ) and P tracked (t + t, a) can be extracted from
the recorded video clip; P simulated (t + t, a) represents the
predicted position of pedestrian at time step t + t; and
the value of P simulated(t + t, a) can be calculated using the
developed approach in terms of actual input parameters at
time step t. The relative distance error reflects in some sense
the inconsistency between the real and simulation trajectories.
Here, the relative distance error was calculated over each time
step for each pedestrian. Fig. 14 shows the corresponding frequency distribution in histogram form. The mean of the relative
distance error was calculated to be 0.322. This value indicates
that the accuracy of the simulation trajectory was acceptable.
Thus far, the comparison on pedestrian distributions, erratic trajectories, and the density-speed fundamental diagram
and quantitative relative distance error calculation over the
pedestrian movement trajectory between the simulation and the
real-life scenario were conducted. To validate the developed
approach perfectly, we also investigated the possibility of reproducing pedestrian self-organized phenomena. Fortunately, the
typical self-organized phenomena (lane-formation) that existed
in real life were successfully reproduced in our simulation
experiments. Fig. 15(a) shows the lane-formation phenomenon
in the simulations. Several narrow pedestrian walking lanes
consist of a sequence of same directional pedestrians can
be observed. This phenomenon revealed that the pedestrians
tended to follow the pedestrian ahead in the same direction to
obtain free moving space and avoid the potential for collision
from pedestrians moving in the opposite direction. Fig. 15(b)
shows the corresponding video screenshot in the real-life scenario when the lane formed. This finding further indicates that
the developed approach is capable of simulating bidirectional
pedestrian flow.
IV. C ONCLUSION
In this paper, we presented a novel microscopic pedestrian
flow simulation approach based on an ANN. The approach
3169
3170
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 17, NO. 11, NOVEMBER 2016
[31] J. G. Juang and K. C. Cheng, Application of neural networks to disturbances encountered landing control, IEEE Trans. Intell. Transp. Syst.,
vol. 7, no. 4, pp. 582588, Dec. 2006.
[32] P. C. M. Leung and E. W. M. Lee, Estimation of electrical power consumption in subway station design by intelligent approach, Appl. Energ.,
vol. 101, pp. 634643, Jan. 2013.
[33] K. Hornik, M. Stinchcombe, and H. White, Multilayer feedforward
networks are universal approximators, Neural Netw., vol. 2, no. 5,
pp. 359366, 1989.
[34] W. Wang, S. Lo, S. Liu, and H. Kuang, Microscopic modeling of
pedestrian movement behavior: Interacting with visual attractors in the
environment, Transp. Res. C, Emerging Technol., vol. 44, pp. 2133,
Jul. 2014.
[35] F. Martinez-Gil, M. Lozano, and F. Fernndez, MARL-Ped: A multiagent reinforcement learning based framework to simulate pedestrian groups, Simul. Modell. Pract. Theory, vol. 47, pp. 259275,
Sep. 2014.
[36] H. Gotoh, E. Harada, and E. Andoh, Simulation of pedestrian contra-flow
by multi-agent DEM model with self-evasive action model, Safety Sci.,
vol. 50, no. 2, pp. 326332, Feb. 2012.
[37] D. Yanagisawa et al., Introduction of frictional and turning function for
pedestrian outflow with an obstacle, Phys. Rev. E, Stat. Nonlinear Soft.
Matter Phys., vol. 80, 2009, Art. no. 036110.
[38] L. Luo et al., Agent-based human behavior modeling for crowd simulation, Comput. Animation Virtual Worlds, vol. 19, no. 3/4, pp. 271281,
2008.
[39] NeuroShell 2 Manual, Ward Syst. Group Inc., Frederick, MA, USA, 1996,
Ward.
[40] M. Ballerini et al., Interaction ruling animal collective behavior depends
on topological rather than metric distance: Evidence from a field study,
Proc. Nat. Acad. Sci., vol. 105, pp. 12321237, 2008.
[41] C. Saloma, G. J. Perez, G. Tapang, M. Lim, and C. Palmes-Saloma, Selforganized queuing and scale-free behavior in real escape panic, Proc.
Nat. Acad. Sci., vol. 100, no. 21, pp. 11 94711 952, 2003.
[42] J. L. Deneubourg, J. Halloy, J. M. Am, C. Rivault, and C. Detrain, Selforganised choice based on inter-attraction: The example of gregarious
animals, in Proc. Pedestrian Evac. Dyn., N. Waldau, P. Gattermann,
H. Knoflacher, and M. Schreckenberg, Eds., 2007, pp. 455463.
[43] J. Ma, W. G. Song, J. Zhang, S. M. Lo, and G. X. Liao, k-NearestNeighbor interaction induced self-organized pedestrian counter flow,
Physica A, vol. 389, pp. 21012117, 2010.
[44] A. Johansson, D. Helbing, and P. K. Shukla, Specification of the social force pedestrian model by evolutionary adjustment to video tracking
data, Adv. Complex Syst., vol. 10, pp. 271288, 2007.
[45] K. Teknomo, Microscopic pedestrian flow characteristics: Development of an image processing data collection and simulation model,
Ph.D. dissertation, Department of Human Social Information Sciences,
Tohoku Univ., Sendai, Japan, 2002.
[46] P. Werbos, Beyond regression: New tools for prediction and analysis in
the behavioral sciences, Ph.D. dissertation, Department of Mathematics,
Harvard Univ., Cambridge, MA, USA, 1974.
[47] S. L. zesmi and U. zesmi, An artificial neural network approach to
spatial habitat modelling with interspecific interaction, Ecol. Modell.,
vol. 116, pp. 1531, 1999.