SWARMTRACK: A PARTICLE SWARM APPROACH TO
VISUAL TRACKING
Luis Antón-Canalís
1
, Elena Sánchez-Nielsen
2
, Mario Hernández-Tejera
1
1
Instituto de Sistemas Inteligentes y Aplicaciones Numéricas en Ingeniería. Campus Universitario de Tafira,
35017 Gran Canaria, Spain.
2
Departamento de E.I.O. y Computación, 38271 Universidad de La Laguna, Spain
Keywords: Computer Vision, Real Time Object Tracking, Swarm Intelligence.
Abstract: A new approach to solve the object tracking problem is proposed using a Swarm Intelligence metaphor. It is
based on a prey-predator scheme with a swarm of predator particles defined to track a herd of prey pixels
using the intensity of its flavours. The method is described, including the definition of predator particles’
behaviour as a set of rules in a Boids fashion. Object tracking behaviour emerges from the interaction of
individual particles. The paper includes experimental evaluations with video streams that illustrate the
robustness and efficiency for real-time vision based tasks using a general purpose computer.
1 INTRODUCTION
Tracking moving objects is a critical task in
computer vision, with many practical applications
such as vision based interface tasks (Turk, 2004),
visual surveillance (Sánchez-Nielsen, 2005a) or
perceptual intelligence applications (Pentland,
2000).
Template based approaches track a target through
a video by following one o more exemplars
(templates) of the visual appearance of the object.
The template tracking problem has been
classically formulated as a search problem of a
pattern in the current frame of the video stream that
matches the exemplars as closely as possible.
Several solutions have been proposed in this sense to
deal with the problem. At present, there are still
obstacles in achieving all-purpose and robust
tracking systems. Different issues must be addressed
in order to carry out an effective tracking approach:
(1) Dynamic appearance of deformable or
articulated targets, (2) Dynamic backgrounds, (3)
Following different target motions without
restriction, (4) Changing lighting conditions, (5)
Camera motion and (5) Real-time performance.
In this paper, a new approach is proposed. The
solution is based on a Swarm Intelligence paradigm
and, particularly, on focusing the tracking problem
under the eyes of a predator-prey metaphor. In our
tracking context the template is a sample of prey
pixels that supply the scent of preys to be tracked to
a swarm of predator particles. Then, using a prey
scent similarity principle, each predator particle will
track its prey. As a result, the tracking of the object
will be an emergent property of the Swarm of
Particles, where tracking behaviour appears thanks
to a set of individual and group behaviour rules.
In the next section, a review of the tracking
problem is included. In section 3, a presentation of
Swarm Intelligence is detailed. Sections 4 and 5
describe the proposed method. Section 6 includes
experimental evaluations of the proposal with video
streams in different contexts and finally, section 7
discusses the conclusions of this work.
2 PREVIOUS WORK
Traditional tracking approaches are based on the use
of models or templates that represent the target
features in the spatial-temporal domain. These
templates can be explicitly constructed by “hand”,
learned from example sequences or dynamically
acquired from the moving target. These template
based approaches are focused on the use of two main
processes: (i) matching and (ii) updating.
Template matching corresponds to the process in
which a reference template is searched for in an
221
Antón-Canalís L., Sánchez-Nielsen E. and Hernández-Tejera M. (2006).
SWARMTRACK: A PARTICLE SWARM APPROACH TO VISUAL TRACKING.
In Proceedings of the First International Conference on Computer Vision Theory and Applications, pages 221-228
DOI: 10.5220/0001372002210228
Copyright
c
SciTePress
input image to determine its location and
occurrence. Over the last decade, different
approaches based on searching the space of
transformations using a measurement similarity have
been proposed for template based matching. Some
of them establish point correspondences between
two shapes and subsequently find a transformation
that aligns these shapes. The iteration of these two
steps involves the use of algorithms such as iterated
closest points (ICP) (Besh et al., 1992), (Chen et al.,
1992) or shape context matching (Belongie et al.,
2002). However, these methods require a good
initial alignment in order to converge; particularly
whether the image contains a cluttered background.
Other approaches, in order to compute the
transformation that best matches the template into
the image, are based on searching the space of
transformations using exhaustive search based
methods (Rucklidge, 1996). A reduction of the
computational cost has been introduced by means of
the use of heuristic algorithms (Sánchez-Nielsen,
2005b).
Template updating is related to the process of
update of the template that represents the target. The
underlying assumption behind several template
tracking approaches is that the appearance of the
object remains the same through the entire video
(Tyng-Luh, 2004), (Comaniciu, 2000). This
assumption is generally reasonable for a certain
period of time and a naïve solution to this problem is
updating the template every frame (Parra et al.,
1999) or every n frames (Reynolds, 1998) with a
new template extracted from the current image.
However, small errors can be introduced in the
location of the template each time the template is
updated and this situation entails that the template
gradually drifts away from the object (Matthews et
al., 2004). Matthews, Ishikawa and Baker in
(Matthews et al., 2004) propose a solution to this
problem. However, their approach only addresses
the issue related to objects whose visibility does not
change while they are being tracked. An
improvement of the update problem for this situation
using a second order isomorphism based method has
been recently proposed by (Guerra, 2005).
Other approaches based on deformable templates
(Yuille et al., 1992) minimize, for each frame, an
energy function which is specific to the geometry of
the tracked object. Elastic snakes (Kass et al., 1987)
minimize a more general energy function, which has
terms representing elastic and tensile energy to
ensure that the snake is smooth, and an image-
dependent term that pushes the snake towards the
feature of interest. The Kalman tracker (Blake et al.,
1993) requires a learned linear stochastic dynamical
model which describes the evolution of the contour
to be tracked, assuming that the observation of the
contour has been corrupted by Gaussian noise. The
condensation tracker (Isard, 1998) also assumes a
dynamical model describing contour motion, which
requires training using the object moving over an
uncluttered background to learn the motion model
parameters before it can be applied to the real scene.
Currently, computing all the possible
transformations that best match a template into an
image and updating the new appearance of the target
without drifting the tracked object for tracking
arbitrary shapes with fast and vast movements under
unrestricted environments for real-time tasks are
open problems.
On the other hand, the main issue of deformable
template based approaches is that for any given
application, hand- crafting is required; that is, if it is
desired to track the motion of lips, a specific energy
function that is appropriate for lips must be
designed. Kalman trackers solve this problem, but
are not adequate to track moving objects with the
presence of clutter. This problem is addressed by the
condensation tracker. However, this tracker requires
a dynamical model of the object to be tracked.
In this paper, a new approach is proposed to solve
the problem of visual tracking of objects with
arbitrary shapes in cluttered moving scenes for
different visual applications under unrestricted
environments. As a result, instead of using region
template tracking or using salient features in the
image, or minimizing energy functions, we propose
to use a Swarm Intelligence metaphor based on a
prey-predator scheme with a particle swarm of
predators defined to track a herd of prey pixels using
the intensity of its scent. Neither complete aspect
based-templates of the visual target nor dynamical
model of the motion of the object are required.
3 SWARM INTELLIGENCE
Swarm intelligence (SI) (Bonabeau, 2000) is an
innovative computational and behavioral metaphor
that takes its inspiration from biological examples
provided by social insects and by swarming,
flocking, herding and shoaling phenomena in
vertebrates (Parrish et al., 1997). SI is an artificial
intelligence technique based on the study of
collective behaviour in decentralized, self-organized
systems. SI systems are typically made up of a
population of simple individuals interacting locally
with one another and with their environment.
VISAPP 2006 - MOTION, TRACKING AND STEREO VISION
222
Although there is no centralized control structure
dictating how individuals should behave, the main
characteristic of this approach is that the collective
behaviour is an emergent phenomen resulting from
the interaction of the local behaviour of each
independent individual. Thus, the abilities of such
natural systems transcend those of individuals. The
advantages of this metaphor are related, on one
hand, by the robustness and sophistication of the
obtained group behaviour and, on the other hand,
with the simplicity and low computational costs of
the individual computational elements.
Many successful SI techniques have been
developed during last years, including Ant Colony
Optimization (ACO), (Dorigo, 1996), or Particle
Swarm Optimization (PSO) (Eberhart, 1995) as
metaheuristic optimization techniques. SI simulation
techniques of animal group behaviour have been
used in artificial life, computer graphics and picture
animation.
Among artificial life simulations, Boids
(Reynolds, 1987) is an example of emergent
behaviour; the complexity of Boids arises from the
interaction of individual agents (boids, in this case)
adhering to a set of simple rules. The rules applied in
the simplest Boids world are: (1) separation, (2)
alignment and (3) cohesion. This framework, related
to Steering Behaviours, is often used in computer
graphics, providing realistic-looking representations
of flocks, shoals, herds or crowds.
In the following two sections, the proposed
method, using the Swarm Intelligence paradigm, is
described.
4 PREDATOR SWARM BASED
MODEL
The tracking process is formulated in terms of a
predator-prey scheme where pixels in a video
sequence are considered preys and a particle swarm
cooperates to hunt them.
A set of prey samples (pixels) is selected in an
initial image of the video sequence. Preys are
characterized by their scent intensity, which is an
abstraction of their pixel image information: colour
and gradient magnitude. In order to follow them, a
swarm with the same number of predator particles is
generated. Each predator particle will be fed with a
single sample, and it will adapt its taste preference to
that prey’s features. During the video sequence, each
predator will try to satisfy its taste hunting similar
preys, following their scent. However, as preys may
disappear due to pixel attributes changing over time,
predators will be able to adapt their sense of smell in
order to hunt different preys. This way, each image
in the video sequence may be understood as a herd
where each pixel is a potential prey for the swarm,
depending on its colour and gradient value.
Predators are designed as described in the
following subsections.
4.1 Swarm Structure
In order to be able to hunt its favourite preys, each
particle stores the following information:
1) Position in the search space.
2) Unitary velocity in the previous iteration, initially
zero.
3) Speed, the amount of pixels that a particle is able
to travel between two iterations. Speed varies in time
depending on a particle’s comfortness (Pcf) (see
below).
4) Colour bank list (Pcbl), a list of recently seen
colours that is a representative subset of the colours
that are similar to the colour of the presented prey
pixel at initial time. Bank colours are represented by
CIE L*a*b colour space. Thus, a certain light
intensity independence may be obtained weighting
each L*a*b colour vector when two colours (0.1*L,
1.0*a, 1.0*b) are compared.
A particle’s comfortness (Pcf) is a measurement of
its similarity with its neighbour image pixel’s
colour, given by:
PcblmNnmnsimPcf ∈∀∈∀
=
,)),(min(
(1)
Where N is the particle’s neighbourhood, Pcbl is the
particle’s colour bank list and sim is a similarity
measurement given by:
e
cc
mn
mnsim
σ
||
),(
−
=
(2)
Where |n
c
– m
c
| is the Euclidean distance between
the two colours in CIE L*a*b* colour space. This
coefficient measures the quality of prey tracking as
it is carried out by the predator particle.
In order for each particle to keep contact with the
swarm, three global values are computed using a
weighted average of each particle’s information.
SWARMTRACK: A PARTICLE SWARM APPROACH TO VISUAL TRACKING
223
1) Swarm centroid:
)(
)(
0
0
∑
∑
=
<
=
<
⋅
=
i
Si
i
i
Si
ii
c
Pcf
PcfP
S
(3)
Where P
i
corresponds to the particle’s position and
Pcf
i
represents the particle’s comfortness.
2) Swarm velocity:
)(
)(
0
0
∑
∑
=
<
=
<
⋅
=
i
Si
i
i
Si
ii
v
Pcf
PcfPv
S
(4)
Where Pv
i
is the particle’s velocity.
3) Predicted centroid: given the current swarm
centroid and velocity, the swarm predicts where its
centroid may lay in the following iteration.
Using Pcf as a weighting factor, we assure that those
predator particles that are closer to their objective
prey are much more relevant to the swarm’s global
behaviour than those particles that may have lost
their target.
4.2 Swarm Behaviour
The swarm follows a Boid-like movement
(Reynolds, 1987), preying those high gradient areas
that best suit its particles P
cbl
colours. Each particle
follows four movement rules, each of which returns
a velocity vector, where the weighted sum of them
will characterize the final particle velocity and
speed.
4.3 Particle Movement Rules
Swarm movement and preying behaviour emerges
from the interaction of each particle’s movement,
which is defined by the following rules:
Rule 1) Colour & Topography: A particle analyzes
its closest preys (image pixels in the neighbourhood
of its initial location) obtaining a vector towards the
area with higher gradient magnitude and colour
similarity with the particle’s colour bank.
∑
∑
=
<
=
<
∇⋅
∇⋅⋅−
=
0
0
1
)),(min(
)),(min()(
i
Ni
ii
i
Ni
iii
IPcblPcsim
IPcblPcsimPP
V
(5)
where P
i
represents a prey’s position, P corresponds
to the current particle’s position, sim(Pc
i,
Pcbl) is
given by expression (2) for each value stored in the
particle’s Pcbl, and
∇
I
i
is the gradient magnitude at
pixel i. This element introduces a topographical-
related weight in the equation, giving priority to
significant image points (high gradient magnitude
pixels) in the particle movement.
The particle’s speed is computed by the following
expression:
MAXS
N
IPcblPcsim
MINSP
i
Ni
ii
s
⋅
∇⋅
+=
∑
=
<
0
)),(min(
(6)
Where MINS and MAXS are predefined minimum
and maximum speeds for a given particle. The sum
is related to a measurement of how well a particle’s
colour fits in its neighbourhood. The higher the
value (worse fitting) the faster it will move.
Increasing its speed, a particle will likely escape
faster that part of the image, hopefully finding better
preys guided by the rest of rules.
Rule 2) Grouping: Computes a vector from the
particle’s position towards the current swarm
centroid. This rule will avoid scattering, keeping the
swarm together. A particle uses the swarm centroid
instead of it closest neighbours positions like Boids
do, because group splitting is not desirable. It is
obtained as follow:
(
)
PS
PS
V
c
c
−
−
=
2
(7)
Where S
c
comes from (3).
Rule 3) Alignment: Computes the sum of the
particle’s current velocity and the swarm velocity.
With this rule a particle will adapt its movement to
head towards where the rest of the swarm is heading
to. Once again, instead of its closest neighbours the
whole swarm is considered. This rule acts like a
voting system where the majority decides where the
swarm will move.
VISAPP 2006 - MOTION, TRACKING AND STEREO VISION
224
()
vv
vv
PS
PS
V
−
−
=
3
(8)
where S
v
comes from (4)
Rule 4) Prediction: This rule will direct the
particle’s movement towards the position where the
swarm’s centroid will most probably be at the
following iteration.
(
)
PS
PS
V
pc
pc
−
−
=
4
(9)
Where S
pc
corresponds to the swarm predicted
centroid position. This way, a particle is able to
guess the group position in future iterations.
The classic Boids separation behaviour (Reynolds,
1987) was not included in our swarm model because
each particle has its own colour information, its own
prey, so even if two particles share the same spatial
position they do not have to necessarily move
towards the same point.
Finally, the four resultant velocities are weighted
and added to a portion of the previous iteration
velocity for each particle, Ps
t-1
and multiplied by the
current particle speed.
()
tt
PsPswvwvwvwvV .....
144332211 −
+
+
++=
(10)
5 TRACKING
The cooperative social interaction leads the swarm
towards those areas in the image which are similar to
that where the swarm was created, emerging a non-
structural pattern tracking behaviour where the
swarm centroid, velocity and speed will respectively
indicate the tracked object position and relative
movement information, as seen in figure 1.
Figure 1: Object being tracked (preyed) by a swarm.
White dots represent particles, while the white line shows
current swarm velocity.
Tracking is enhanced using two key ideas: (i)
individual comfortness optimization and (ii) swarm
adaptation.
Individual comfortness optimization is related to
a direct application of Particle Swarm optimization
theory (Eberhardt, 1995); where each particle tries to
minimize a certain error using local and global
information based on colour matching and gradient’s
magnitude. As a result, each particle will move
towards those prey pixels that best match the tracked
scent (colour). Note that prey pixel scent intensities
are proportional to image gradient magnitudes, so
predators will be attracted to interest points in
images that match their scent track.
Figure 2 shows a detail of those points that seem
to be most interesting to a swarm that is tracking a
white road line in an automated vehicle based
context.
Figure 2: Swarm perception. A swarm created in a white
region will be attracted by white colours on high gradient
magnitude pixels, shown brighter on the image on the
right.
In order to avoid the introduction of small errors
in the location of the swarm, the swarm is updated
using a colour bank for each particle. This colour
bank will allocate a list of similar prey scents,
avoiding any kind of false averaged values when a
particle is comparing itself with its neighbourhood
as seen in section 4.1, using rule 1.
6 RESULTS
In order to test the proposed tracking approach,
different indoor and outdoor video streams related to
different visual tasks have been used for
experimental evaluations. Each one of these
sequences contains frames of 320 x 200 pixels that
were acquired at 25 fps. All experimental results
were computed on a P-IV 1.4 Ghz.
Prey samples are initialized defining a rectangular
area on the first image of the sequence. This process
can be automated, e.g. using cascade classifiers for
face or hand detection (Anonymous). The swarm,
SWARMTRACK: A PARTICLE SWARM APPROACH TO VISUAL TRACKING
225
once created and fed with sample prey pixels, is able
to follow them on a varied number of non-cluttered
backgrounds and light conditions, as seen in Figure
3, 4, 5, 6 and 7.
The parameters used for the swarm (100
particles) have been initialized with the following
values: W
1
(colour & topography) = 1.0, W
2
(grouping) = 0.3, W
3
(alignment) = 0.5, W
4
(prediction) = 0.2, W
5
(Particle’s velocity at time t-1)
= 0.1, δ= 10.0, neighbourhood size = 15, minspeed
= 5.0, maxspeed = 10.0 and colour bank list size = 3.
The achieved processing rate is around 15fps.
Note no optimizations have been implemented.
In order to evaluate the robustness of the proposed
approach, we manually annotate the centroid point
to be tracked and then we measure the Euclidean
distance from the annotated hand-tracked point and
the swarm’s centroid to the origin (0, 0) through
time. Values were measured every ten frames.
Graphics in figure 3, 4, 5, 6 and 7 illustrate the
results obtained.
The dotted line represents the hand-tracked point
and the continuous line corresponds to the swarm’s
centroid. It is important to point out that the swarm
floats freely over tracked objects, so both lines will
not necessarily coincide. However, they evolve
similarly when the swarm follows successfully the
tracked object.
0
50
100
150
200
250
300
350
1 3 5 7 9 1113151719212325272931333537394143454749
Object Swarm centroid
Figure 3: The swarm is created over a face, and follows it
while it moves around
.
0
50
100
150
200
250
300
350
1 5 9 13172125293337414549535761656973778185
Object Swarm centroid
Figure 4: The swarm is created over a girl’s face, and
follows it while she makes faces and moves around. The
swarm loses its target when it is hidden almost at the end
of the sequence
.
0
50
100
150
200
250
300
350
1 3 5 7 9 11131517192123252729313335373941434547495153555759
Object Swarm centroid
Figure 5: This time, our swarm follows a continuously
gesture changing hand. It has no problems even when the
hand meets the face on its movement
.
VISAPP 2006 - MOTION, TRACKING AND STEREO VISION
226
-50
50
150
250
350
123456789101112131415
Object Swarm centroid
Figure 6: The swarm follows a skier, who moves in a fast
wavy course. Sudden changes in speed (acceleration) and
direction confuses the swarm, but it is able to follow the
skier
.
0
50
100
150
200
250
300
350
12345678910111213141516171819202122232425
Object Swarm centroid
Figure 7: A car is followed by the swarm while it drives
away, until it becomes too small for the swarm to follow
it.
On Figure 7 the car is lost when the shape that
characterizes the car is too small and the colour and
gradient magnitude are not significant for the
swarm. Background areas with high magnitude
gradient and significant colours for the swarm may
also attract it.
A swarm, however, may deal with occlusion as
long as it has tracked an object for some frames and
it does not alter its movement during occlusion.
When this happens, rule 1’s resultant vector will not
be significant. The swarm’s acquired velocity (rule
3) will allow it to surpass the occlusion. However, if
the occluding object’s features satisfy the swarm’s
taste, it may decide to follow it and lose its original
target. In general, swarms may be confused by those
areas with high gradient magnitudes and colours
similar to what the swarm expects. This could be
solved creating leaders in the swarm, able to follow
feature points in the tracked object, which would
have a higher influence over the swarm’s movement.
The amount of weights and parameters could be
seen as a drawback of the proposed method.
However, once a good set of parameters have been
computed, the proposal works for a wide range of
visual applications and arbitrary shape with a vast
range of movements such those illustrated in Figure
3, 4, 5, 6 and 7.
7 CONCLUSIONS AND FUTURE
WORK
In this paper, a new tracking method based on a
Swarm Intelligence Metaphor has been described.
The main idea of the proposal consists on a prey-
predator scheme, where a swarm of predator
particles follows pixel scents (colours) similar to
those that where presented to predators at initial
tracking time. Image gradient is used as a feature
regulator, defining the scent intensity, which is
proportional to the value of the gradient. Thus,
matching colours located in high interest pixels are
much more interesting for a given predator. Each
predator particle’s movement is governed by four
basic rules. Tracking behaviour emerges from the
interaction of each particle, where the tracked
object’s position is defined by the swarms’ centroid.
Because our swarms do not follow shapes but
light intensity independent colours, the resulting
tracking method is robust under deformations of the
tracked object, cluttered images and ligh changes,
being computationally a low cost solution.
Experimental results show that, with unrestricted
images, and using general purpose hardware, almost
real time tracking is obtained (~20fps, tracking with
100 particles, using 320*200 pixels images in a P-IV
1.4 Ghz). Due to its computational simplicity the
proposed solution is very efficient and highly
parallelizable.
The method’s accuracy is based on the size of the
tracked object. With a good area to track, as e.g,
sequences in Figures 3 and 4, accuracy is maximum,
decreasing proportionally to the size of the region to
be tracked, such as in the last frames of the sequence
corresponding to Figure 7. Future work will include
comparisons with classic tracking methods.
SWARMTRACK: A PARTICLE SWARM APPROACH TO VISUAL TRACKING
227
ACKNOWLEDGMENTS
This work has been supported by the Spanish
Government and the Canary Islands Autonomous
Government under projects TIN2004-07087 and
PI2003/165.
REFERENCES
Aloimonos, Y., 1993. Active Perception. Lawrence
Erlbaum Assoc., Pub., N.J.
Belongie, S., Malik, J., and Puzicha J., 2002. Shape
matching and object recognition using shape context.
In IEEE Trans. on Pattern Analysis and Machine
Intelligence, 24(4):509-522.
Besh, P. J., and McKay N., 1992. A method for
registration of 3D shapes. In IEEE Trans. on Pattern
Analysis and Machine Intelligence, 14(2):239-256.
Blake, A., Curwen R., and Zisserman A., 1993. A
framework for spatio-temporal control in the tracking
of visual contours. In International Journal of
Computer Vision, 11(2):127-145.
Bonabeau, E., Dorigo, M., and Theraulaz, G., 2000.
Swarm Intelligence: From Natural to Artificial
Systems, Oxford University Press.
Chen Y. amd Medioni G., 1992. Object modelling by
registration of multiple range images. In Image and
Vision Computing, 10(3):145-155.
Comaniciu, D., Ramesh V., and Meer P., 2000. Real-time
tracking of non-rigid objects using mean shift. In IEEE
Conf. on Computer Vision and Pattern Recognition,
volume II, pp. 142-149, Hilton Head, SC.
Dorigo M., V. Maniezzo & A. Colorni, 1996. Ant System:
Optimization by a colony of cooperating agents. IEEE
Transactions on Systems, Man, and Cybernetics-Part
B, 26(1):29-41.
Eberhart, R. C. and Kennedy, J., 1995. A new optimizer
using particle swarm theory. Proceedings of the Sixth
International Symposium on Micromachine and
Human Science, Nagoya, Japan. pp. 39-43.
Guerra Cayetano, Hernández Mario, Domínguez Antonio,
Hernández Daniel, 2005. A new approach to the
template update problem. In Lecture Notes in
Computer Science LNCS 3522, pp. 217-224.
Isard, M., and Blake A., 1998. Condensation-conditional
density propagation for visual tracking. In International
Journal of Computer Vision, 29(1):5-28.
Kass, M., Witkin A., and Terzopoulos D. Snakes: active
contour models. In Proc. 1
st
International Conference
on Computer Vision.
Kennedy, J. and Eberhart, R. C., 1995. Particle swarm
optimization. In Proceedings of IEEE International
Conference on Neural Networks, Piscataway, NJ. pp.
1942-1948.
Matthews, I., Ishikawa, T., and Baker S., 2004. The
template update problem. In IEEE Transactions on
Pattern Analysis and Machine Intelligence, 26(6):810-
815.
Parra, C., Murrieta-Cid, Devy, M., and Briot, M., 1999.
3-D Modelling and Robot Localization from Visual
and Range Data in Natural Scenes. In Lecture Notes in
Computer Science 1542, Springer Verlag, pp. 450-468.
Parrish, Julia K. (Editor), William M. Hamner (Editor),
1997. Animal Groups in Three Dimensions: How
Species Aggregate. Cambridge University Press.
Pentland, A., 2000. Perceptual Intelligence. In
Communications of ACM, 43(3):35-44.
Reynolds, C. W., 1987. Flocks, Herds, and Schools: A
Distributed Behavioural Model. In Computer Graphics,
21(4) (SIGGRAPH '87 Conference Proceedings) pp.
25-34.
Reynolds, J., 1998. Autonomous underwater vehicle:
vision system. PhD thesis, Robotic Systems Lab.
Department of Engineering. Australian National
University Canberra, Australia.
Rucklidge W. J., 1996. Efficient Visual Recognition Using
the Hausdorff Distance. In Lecture Notes in Computer
Science, nº 1173, Springer-Verlag, NY.
Sánchez-Nielsen Elena, Hernández-Tejera Mario, 2005a.
A fast and accurate tracking approach for automated
visual surveillance. In 39
th
IEEE International
Carnahan Conference on Security Technology, pp.
113-116.
Sánchez-Nielsen Elena, Hernández-Tejera Mario, 2005b.
A heuristic search based approach for moving objects
tracking. In 19
th
International Joint Conference on
Artificial Intelligence (IJCAI-05), pp. 1736-1737.
Turk, M., 2004. Computer Vision in the Interface. In
Communications of the ACM, 47(1): 61-67.
Tyng-Luh Liu, Hwan-Tzong Chen, 2004. Real-Time
tracking using trust-region methods. In IEEE
Transactions on Pattern Analysis and Machine
Intelligence, 26(3):397-401.
Yuille, A., Hallinan, P., and Cohen D., 1992. Feature
extraction from faces using deformable templates. In
International Journal of Computer Vision 8(2):99-112.
VISAPP 2006 - MOTION, TRACKING AND STEREO VISION
228
