Hide and Seek

An Active Binocular Object Tracking System

Pramod Chandrashekhariah and Jochen Triesch

Frankfurt Institute for Advanced Studies (FIAS),

Johann Wolfgang Goethe University, Frankfurt am Main, Germany

Keywords:

Object Tracking, Active Vision, Stereo Vision, Segmentation, Object Recognition, Humanoid Robot, iCub.

Abstract:

We introduce a novel active stereo vison-based object tracking system for a humanoid robot. The system

tracks a moving object that is dynamically changing its appearance and scale. The system features an in-

built learning process that simultaneously learns short term models for the object and potential distractors.

These models evolve over time, rectifying the inaccuracies of the tracking in a cluttered scene and allowing

the system to identify unusual events such as sudden displacement, hiding behind or being masked by an

occluder, and sudden disappearance from the scene. The system deals with these through different response

modes such as active search when the object is lost, intentional waiting for reappearance when the object is

hidden, and reinitialization of the track when the object is suddenly displaced by the user. We demonstrate

our system on the iCub robot in an indoor environment and evaluate its performance. Our experiments show a

performance enhancement for long occlusions through the learning of distractor models.

1 INTRODUCTION

Tracking objects in cluttered dynamic environments

remains a major challenge. In fact, it can be argued

that tracking problems can be made arbitrarily difﬁ-

cult if objects move erratically, are occluded for long

stretches of time and can have a changed appearance

when they return into view. In such cases even human

observers will have difﬁculties identifying the correct

target, see, e.g., (Triesch et al., 2002). In such dif-

ﬁcult situations it is important to maintain and adapt

models for the object’s appearance and potential dis-

tractors and to explicitly detect problematic situations

and respond to them in an intelligent way.

Here we present the Hide and Seek system

(Fig. 1), which is an active tracking system imple-

mented on a humanoid robot that utilizes stereo vi-

sion and segmentation and learns appearance models

of objects and potential distractors. It uses these mod-

els to explicitly represent difﬁcult situations such as

occlusions resulting from the object hiding behind an

occluder or being masked by it, sudden disappearance

or displacement of the object, or other distractions and

responds to them appropriately.

There is a signiﬁcant amount of work on object

tracking pertaining to tasks such as video surveil-

lance, vehicle navigation, video editing, human-

Figure 1: Hide and Seek running on the iCub robot.

computer interaction, etc. However, most methods

consider only prerecorded video streams or a pas-

sive/static camera. There are comparitively few at-

tempts of tracking on a robotic platform which is es-

pecially challenging as the system should perform in

real time using active cameras while the object as well

as the robot move in a dynamic environment (for ex-

ample, Falotico and Laschi 2009; Ginhoux and Gut-

mann 2001).

In this work we develop an active tracking sys-

tem that primarily involves frame-to-frame matching

modulated by a segmentation scheme. We develop a

feature based approach that has proven to be robust

to occlusions and local deformations (for example, Ta

et al. 2009). Their discriminative ability is also suit-

able for better object recognition. However, in the

tracking scenario, they are sensitive to drastic changes

in appearance and pose of the moving object thus los-

584

Chandrashekhariah P. and Triesch J..

Hide and Seek - An Active Binocular Object Tracking System.

DOI: 10.5220/0004690705840591

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 584-591

ISBN: 978-989-758-009-3

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

ing global information of the tracked object over time.

Feature extraction is also computationally expensive

which would make the algorithm slow. To deal with

this, our system utilizes a single class of image fea-

tures that are used for a number of different objectives

including stereo matching, frame-to-frame matching,

and object and distractor modeling. To achieve real

time performance we use GPU acceleration for fea-

ture processing

Learning is an important part of our system that

makes it more reliable and versatile. It has been

shown that recognition and model based methods pro-

vide higher resistance to discontinuities and irregu-

larities of the track (for example, Nelson and Green

2002). A hybrid model of frame to frame tracking

along with object learning has also been demonstrated

on a passive monocular camera (Kalal et al., 2012). In

contrast to these, we learn models for both object and

distractors in the scene for an active stereo vision sce-

nario that not only rectiﬁes tracker inaccuracies but

also identiﬁes unusual events and takes necessary ac-

tion.

In the following, we will describe the architecture

of our system, explain the methods used for model-

ing object and distractors, and present experiments

demonstrating the advantage of including adaptive

object and distractor models.

2 SYSTEM ARCHITECTURE

The overall system architecture is shown in Fig. 2.

Images from both cameras are fed to the central track-

ing module for processing. The central tracking mod-

ule is inﬂuenced by inputs from the object and dis-

tractor learning models and vice versa. The output

from these modules are fed to the decision and mo-

tor control module that decides the course of action

and sends commands to the robot. The output of this

module is also fed back to the tracking and learning

modules where this information is used to update rel-

evant parameters.

2.1 Tracking Module Overview

The principle behind our system is to combine efﬁ-

ciently the temporal and spatial information of the

moving object at every time frame. We use feature de-

scriptors (local cue) that add resistance to perturbation

and occlusions while stereo segmentation (global cue)

updates object information by adding new features

Speed of 20 fps is achieved on GPU which was ob-

served to be 4-5 times faster than that on CPU.

Object Learning

Distractor Learning

Decision

and

Motor control

Robot

Tracking

Figure 2: System Architecture.

that are revealed as the object is turning or changing

its shape while moving. A block diagram of the track-

ing module is given in Fig. 3. It comprises feature

extraction in left and right images, stereo matching,

frame-to-frame matching, estimation of object center,

bounding box, and segmentation. An internal mem-

ory aggregates necessary information over time. The

estimated location of the object is fed to the decision

and motor control. Inputs from various parts are fed

to object and distractor learning modules (shown by

dotted lines in Fig. 3).

2.2 Feature Extraction

We calculate local feature descriptors around identi-

ﬁed interest points on the images. Interest points are

detected using the FAST corner detection (Rosten and

Drummond, 2006). We use Gabor wavelets as fea-

tures, which have the shape of plane waves restricted

by a Gaussian envelope function. At each interest

point we extract a 40-dimensional feature vector (also

referred to as a Gabor-jet or Gabor-descriptor) result-

ing from ﬁltering the image with Gabor wavelets of

5 scales and 8 orientations, see (Wiskott et al., 1997)

for details. These descriptors are highly discrimina-

tive and repeatable from frame to frame.

2.3 Stereo Matching and Segmentation

We ﬁnd correspondences between features on left and

right images by performing a brute-force search for

best matches. The similarity between any two de-

scriptors d

(i)

and d

( j)

is calculated as their normalised

inner product and denoted S(d

(i)

( j)

). Each feature

on the left image is associated with the best matching

feature on the right image if the similarity measure S

between the two descriptors exceeds a preset thresh-

old (0.9 in our case). The advantage of this brute

force approach is that it does not require calibration

of the moving cameras. Figure 4 shows an example of

stereo matching. For better visibility, matching lines

are shown for only few objects in the scene and the

saturation of the line colors are made proportional to

their corresponding stereo disparities.

HideandSeek-AnActiveBinocularObjectTrackingSystem

585

Feature extraction

Stereo matching

Temporal matching

Object centroid

estimation

Segmentation

Object centroid

Correction

Bounding box

estimation

Internal

Memory

Decision and Motor control

Object and Distractor Learning

Feature extraction

Figure 3: Block diagram of the tracking module. Rectangular boxes indicate processing units. Boxes with same color indicate

identical processes operating on different inputs. The symbol resembling a cylinder indicates a time delay of 1 frame.

We use the information about matched features for

also segmenting the objects in the scene. This is done

by clustering the matched interest points from the pri-

mary camera (left in our case) into different groups.

A greedy clustering scheme is used that starts with

a single interest point and adds new ones if their x-

position, y-position, and disparity are all within 5 pix-

els of any existing cluster member. Figure 5 illustrates

an example wherein the objects in the scene have been

clustered into different segments.

2.4 Temporal Matching

Tracking in an active vision system has to cope with

image movement due to camera motion as well as

object movement. Matching local descriptors from

frame to frame is a robust approach in such a scenario.

An image at time t is represented by a feature set

consisting of N

elements. Every feature element f

is represented by a tuple containing its Gabor descrip-

tor d

and its location x

on the image.

(i)

i=1:N

, f

(i)

, x

(i)

, where x = (x,y).

(1)

We ﬁnd best matches between frames t − 1 and t by

brute force exhaustive search and associate every fea-

ture in the previous frame f

(i)

t−1

with a description q

(i)

t−1

that contains the index m of the best matched feature

in F

, the similarity s of this best match, the associated

image ﬂow vector ∆x, and a counter n indicating the

number of times a given feature was tracked success-

Figure 4: Example of stereo matching.

fully in the past. As explained in the next section, the

features with a high count n are given higher weights

for estimating the object location. The cyan lines in

Fig. 6 illustrate the matching process from time t − 1

to t. The thinkness of the lines is proportional to the

base 2 logarithm of the count n of the correspond-

ing feature to indicate that relevant features are given

higher weights. Black lines indicate wrong matches

that have been ﬁltered out as explained in the next

section.

2.5 Track Estimation

In this section we explain how the system efﬁciently

uses temporal and stereo matching information to es-

timate the object location at every frame. The part of

the system responsible for this is shaded with gray in

Fig. 3 and an example has been provided in Fig. 6 de-

picting outputs of different processing units (note the

color correspondence between the respective parts).

During tracking we estimate the object’s location

as well as its bounding box at every frame. The

bounding box of the object, which is a rectangle

placed around its centroid, is calculated using the ob-

ject segmentation information at the current frame.

The width and height of the bounding box are ob-

tained by ﬁnding the mean of absolute distances of

Figure 5: Example of segmentation based on feature prox-

imity in the image plane and in depth.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

586

t-1 t

tt-1

Estimation

Correction

Figure 6: Illustration of an instance of tracking that is pro-

cessed within the gray portion of the blockdiagram in Fig 3.

Colors in this image match their corresponding processing

units (boxes) in the blockdiagram.

the segmented feature points from the current object

centroid. The size of the bounding box is averaged

over a few frames which helps in rejecting bad seg-

ments caused by momentary occlusions.

In our system, bounding box and feature matching

count provide global consistency and long term sta-

bility to the track. On the other hand, object segment

and new features provide adaptibility by adding new

information about the object.

Centroid Estimation. As illustrated in Fig. 6 an ob-

ject changes its location, scale and appearance from

time t −1 to t. We estimate the object centroid at time

t using the matched features from time t − 1 to t. We

form a set U

t−1

from the locations x

( j)

t−1

in F

t−1

that

fall on the object (see purple dots in Fig. 6) such that

t−1

( j)

t−1

∈



segment(F

t−1

)

BB(F

t−1

)

o

where BB is the bounding box placed around the ob-

ject centroid at t − 1.

Using the matched information we ﬁnd another set U

in F

such that

(m( j))

: x

( j)

t−1

∈ U

t−1

where m( j) is the index of the matched feature at t for

every j at t − 1.

We now calculate the weighted median of the

matched feature locations in both x and y axis to ob-

tain the initial estimate of the centroid of the object at

time t.

c = Median

(m(1))

 log



(1)

t−1



,..., x

(m( j))

 log



( j)

t−1

i

where  indicates that left hand factor is weighed by

the count equal to right hand factor. The parameters

m and n are as described in section 2.4.

The estimated centroid of the object may not be its

true centroid at time t since the current segment of the

object may contain new features due to object rotation

or change in shape (if the object is not rigid). In order

to compensate for this, we ﬁrst pick the right segment

of the object, segment(F

), at time t by choosing a

segment from the scene on which the estimated cen-

troid falls. We then estimate the true centroid of the

object as follows:

c = Median

(1)

(2)

,...,x

(i)

∀x

(i)

∈ [segment(F

)

BB(F

)]

where BB(F

) is placed around

The calculations above are done only for the pri-

mary camera. The location corresponding to c on the

other camera is obtained by using the stereo matching

information (see section 2.3). The resultant stereo lo-

cation is sent to the decision and motor control mod-

ule

3 LEARNING SYSTEM

As tracking proceeds the system learns models for

both the object as well as potential distractors. The

object model helps in the case of swift movements

and occlusions. The distractor model helps to build a

clean model for the object, to avoid distraction of the

track by other objects and to identify unusual events.

3.1 Object Learning

Tracking is susceptible to occlusions as well as

changes in object appearance. In order to help the

system track an object for long durations we learn an

object model (see also Chandrashekhariah et al. 2013)

using the same feature descriptors and the same pro-

cessing blocks that are used for the tracking scheme.

The object model should capture the stable ap-

pearance of the object rather than picking up irregu-

larities that would arise due to object occlusions. This

demands carefully choosing features that are stable.

We mark a feature on the object as eligible for enter-

ing the object model only when it has been tracked

successfully for a certain time i.e. n > γ. All those

features that enter the object model will eventually

form a feature dictionary for the object.

Dictionary for Object Modeling. The feature dic-

tionary is a buffer that has a ﬁxed amount of memory

as shown in Fig. 7. The dictionary is initially empty

when the tracking begins. Every new feature that

enters the dictionary will enter its descriptor vector

d, location (x,y) relative to the object center and

counter k (initialized to 1). If the dictionary already

The motor control part uses the iKinGazeCtrl module

of the iCub to control the neck and eye coordination while

tracking (Pattacini, 2010)

HideandSeek-AnActiveBinocularObjectTrackingSystem

587

Object Dictionary

Descriptor

x1 y1

Count

Figure 7: Object dictionary.

contains a similar feature, then the associated count

k in the dictionary is incremented. A new feature

that enters the dictionary after it gets full will ﬂush

out the element that has the smallest value of k if

it does not match with any of the existing features

and its location (x,y). Eventually the developed

dictionary represents a model for the object. Having

a dictionary of limited size helps in continuously

evolving the model with the recent variations of the

object appearance. It also gives room for correcting

the model in case the tracking inaccuracies have

incorporated wrong features into the model.

Recognition. The recognition module is always run-

ning in the background to identify if the object has

been displaced while tracking. It uses the saved fea-

tures in the dictionary to ascertain the presence of the

tracked object in the scene. Each feature in the dictio-

nary is matched with all the features on the incoming

image to see if there is a valid match that exceeds the

preset similarity threshold (0.9 in our case) and de-

clare it as eligible for casting a vote. These eligible

features vote for the centroid location (using the (x, y)

information from the dictionary) with respect to the

location of the feature it matches on the image (see

Fig. 8). These votes are collected in a 2D space (a

matrix of image size) are smoothed using a box ﬁlter

or a gaussian ﬁlter and the global maximum is found.

The object is declared to be present in the scene if

this location contains a count higher than a suitable

threshold (10% of dictionary size in our case where

the dictionary contains 300 feature elements).

Figure 8: Object recognition.

3.2 Distractor Learning

In this work, we also develop a model for other ob-

jects in the scene considering the fact that they can

distract the tracking process and hence should be

identiﬁed. We learn short term models for only po-

tential distractors in the scene that would affect the

tracking process in the near future.

A dictionary for the distractor model is structured

in the same way as for the object dictionary but with-

out having the location information (x, y) of features.

This is because the contents of the scene as well as

their locations change continuously in an active vision

scenario making it futile to save feature locations. It

is also a challenge to selectively pick critical parts

of the scene for learning (using a dictionary of ﬁnite

size). We hence pick only those objects or parts of the

scene for learning that are likely to occlude the object

shortly.

Let c

and x

be object centroid and feature lo-

cation respectively (on the left camera). The vector

− x

indicates the distance vector from the fea-

ture to the object centroid whose euclidean distance

= kc

− x

k. Let ∆c

and ∆x

be object and fea-

ture ﬂow vectors respectively. The relative ﬂow vec-

tor of a feature with respect to the object will then be

∆x

− ∆c

which indicates motion in one time frame.

We calculate the distance component of the relative

feature ﬂow vector along the direction of distance

vector which is given by :

∆D



∆ x

− ∆ c

, c

− x



− x

The sign of the term ∆D

indicates the direction of

motion of the feature. The time taken by a given fea-

ture at x

to reach the object centroid in the image

plane is given by :

τ =

∆D

× f ps

where f ps is the speed of the algorithm in frames per

second.

The features that are going to occlude the object in

less than 5 seconds are considered critical and up-

dated to the distractor model (see Fig. 9a). However,

features of the objects that are moving away from

the object (when ∆D

is negative) are not considered

harmful for tracking and hence they are neglected (see

Fig. 9b).

Distractors in the scene that are behind the object by

a considerable depth are neglected, i.e. the following

disparity condition should be satisﬁed for a positive

value of δ.

(x) − c

(x)) −



(x) − x

(x)



> δ ∀i

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

588

(a) (b)

Figure 9: Distractor model update example : (a) Features

that are included in the model. (b) Features that are not

included in the model.

However, distractors in the front are considered

a threat irrespective of their position between the

camera and the object.

In addition, self occlusion of the object is also

neglected by not considering the features belonging

to the object (i.e. ∈ segment(F

))

Distractor Recognition. A Distractor in the scene is

recognized using the distractor dictionary in the same

manner as that of object recognition. However, the

features will vote for their absolute locations rather

than for the centroid. The presence and location of

the distractor is decided by merely calculating the

density of the distractor’s features that are overlap-

ping/masking the tracked object (see Fig. 10).

Figure 10: Distractor recognition.

4 SYSTEM STATES

The system goes into different states as shown in

Fig. 11 depending on inputs received from tracker and

recognition (object and distraction). This will further

guide the decision module of the architecture (Fig. 2).

The system primarily contains four states, each of

which changes to the other based on events. An event

is a logical combination of outputs of different parts

of the system namely - tracking, object recognition

and distractor recognition (see Table 1). Each of these

parts is associated with two quantities - output status

(true/false) and output location.

1. Initialize. The system is initialized with an object

that is in the center and close to the stereo cam-

Initialize

Track

Pause

Hidden

Lost

In focus

Time out

Discovered

Revealed

Figure 11: States of the system.

eras. When an object is brought in the region of

focus the system triggers and starts tracking it.

2. Track. This is the normal state of the system un-

less the object is lost or hidden. When the user

displaces the object to a different location in the

scene with swift movements the object recogni-

tion helps in restoring the track.

3. Pause. The system goes to a pause mode when

the object is masked by a moving occluder or the

object moves behind an occluder. This is known

when the tracker location overlaps with a patch P

that is identiﬁed as an occluder by the distrac-

tor model i.e. when (c ∈ P), concurrently there

is no output from the object recognition (R = 0).

The tracking resumes when the object is revealed

again (R = 1).

4. Search. System starts searching when the object

is lost i.e. when the user moves it out of the scene.

When an object suddenly disappears or when a

distractor suddenly masks the object, the focus of

the cameras in depth shifts drastically from on ob-

ject to either behind or front, again followed by

no output from the object recognition within the

scene purview, i.e. when T = 0 and R = 0. A

random search is then initiated in the direction in

which the object was lost until it is discovered us-

ing recognition output. All the model updates are

stopped in this state as well as the pause state.

5 SYSTEM EVALUATION

Performance evaluation of object tracking often re-

quires ground truth information of the object as in the

case of prerecorded images/videos (Yin et al., 2007).

However, this information is not easily available in a

realtime active vision system. There are other algo-

rithms in the literature that use no ground truth infor-

mation but spatial and temporal differences of object

cues with respect to its surrounding as an alternative.

HideandSeek-AnActiveBinocularObjectTrackingSystem

589

Table 1: System states and events.

System Output

Tracker Status T(1/0), Centroid c

Object Model Recognition R(1/0), Location l

Distractor Model Occluded O(1/0), Patch P

Events Combination

In focus Object is in center and close

Displaced R=1 ∧ kl − ck > T hreshold

Masked/Hidden O=1 ∧ R=0 ∧ c∈P

Revealed O=1 ∧ R=1

Lost T=0 ∧ R=0

Discovered T=0 ∧ R=1

This is again used only in the case of self occlusions

but not for occlusions from other objects in the scene

(Erdem et al., 2004). In order to keep the evaluation

procedure simple for unusual active binocular track-

ing setting we are using, we evaluate our system in a

staged manner by comparing performance of the par-

tial system with the full system in different scenarios.

5.1 Experimental Setup

We conduct our experiment on the iCub robot head in

a regular ofﬁce room with cluttered background (see

Fig. 12). We collect various objects such as a coffee

cup, tea packet, coke tin, books etc that have sufﬁcient

texture on them. We ask a subject unfamiliar with

the system to test the algorithm while we count the

success and failure rates for various scenarios:

1. No Occlusion Scenario. In this case the object

moves in front of a cluttered background making

changes in the scale and in-plane/in-depth rotation

of the object. The algorithm is mainly tested for

three cases:

• Displacement. Swift and erratic movements of

the tracked object.

• Distraction. Bringing another object next to the

object in focus and push it or move them to-

gether

• Merging: Moving the object far away from the

robot and merging it with the background for a

signiﬁcant amount of time.

2. Severe Occlusion Scenario. In this case the object

is completely occluded by other objects. The ob-

ject is either occluded by another object (masking)

or the object goes behind an occluder (hiding). We

further consider two scenarios:

• Momentary Occlusions. We occlude the object

for a short time of the order of 2 to 5 seconds.

• Long-term Occlusions. We occlude the object

for a long time of the order of 10 to 30 seconds.

Figure 12: An instant of the experiment wherein an object

is hiding behind an occluder.

3. Object disappearance scenario: In this case the

object is suddenly moved out of the scene to see if

the system can search and discover the lost object.

5.2 Performance Quantiﬁcation

We conducted two day long experiment (comprising

345 trials in all) and performance was quantiﬁed for

various scenarios.

No Occlusion. We initially compare a partial system

having no learning modules with the full version

of the system, for the no occlusion scenario. Since

there are no complete occlusions the distractor model

would not add much improvement to the performance

and hence it is ignored. We ask the subject to

randomly select 3 objects and make 5 trials for each

of the above cases (90 trials in all). We count the

number of trials in which tracking was successful

throughout the sequence. The percentage of success

is listed in Table 3. We observe that the performance

of the system improves by incorporating learning that

rectiﬁes inaccuracies of the tracker. It particularly

helps in the case of merging. Since in our system the

object segment evolves over time, the system would

gradually fail when the object is merged with the

background for signiﬁcant amount of time. This can

be avoided to a good extent by introducing learning.

Occlusion. We then test the system in the occlusion

scenarios wherein we compare the system containing

object learning alone with the full system also

containing the distractor model. We ignore the

partial system without learning in this case since it is

bound to fail in case of complete occlusions. Table 3

lists the performance in terms of Precision that is

calculated as T P/(T P + FP) where T P and FP are

true and false positives respectively. FP originates if

the system mistakes the occluder for the object. The

numbers listed in the table are averaged over 30 trials

(5 for each of 6 objects with different occluders; 240

trials in all).

We observe that the precision gets better with the

disctractor model since it helps the system to stop

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

590

Table 2: Performance evaluation : No occlusion scenario.

No learning Full system

Displacement 53% 73%

Distraction 86% 93%

Merging 0% 80%

Table 3: Performance evaluation : Occlusion scenario.

No distractor

learning

Full system

Occlusions Mask Hide Mask Hide

Momentary 90% 66% 100% 96%

Long-term 63% 60% 90% 73%

updating the object model when it is occluded.

This is even more evident in the case of long-term

occlusions. We also observe that the performance

for masking is better than hiding because of object

perspective change, camera jitter etc. Any mistake in

the distractor identiﬁcation would harm the system

more in the case of long term occlusions.

Searching. In the end, we also test the searching per-

formance when the object is lost. We count how many

times the object was discovered in less than 30 sec-

onds. We observed that the searching was successful

in 73% of the trials (15 in all).

The tracker performance recorded as a video

can be viewed at https://ﬁas.uni-frankfurt.de/neuro/

triesch/videos/icub/tracking/.

6 CONCLUSIONS

We developed and demonstrated an active stereo-

vision based tracking system on the robot. In con-

trast to other systems in the literature, our system at-

tempts to identify unusual events in the scene and take

necessary actions. The system is also generic enough

to be applied for applications such as face/pedestrian

tracking on static/active cameras (an instance of face

tracking is shown in Fig. 13). There is a scope for fur-

ther improvement in every part of our system. How-

ever, the present system clearly demonstrates how ob-

ject and distractor models allow to recognize difﬁcult

Figure 13: An instant of face tracking.

tracking situations and to respond to them appropri-

ately. This brings us one step closer to building track-

ing systems with human level performance.

ACKNOWLEDGEMENTS

This work was supported by the BMBF Project

“Bernstein Fokus: Neurotechnologie Frankfurt, FKZ

01GQ0840”. We thank S

ebastien Forestier for the

participation and help with the experiments.

REFERENCES

Chandrashekhariah, P., Spina, G., and Triesch, J. (2013).

Let it learn: A curious vision system for autonomous

object learning. In VISAPP.

Erdem, C., Sankur, B., and Tekalp, A. (2004). Performance

measures for video object segmentation and tracking.

Image Processing, IEEE Transactions on, 13(7):937–

951.

Falotico, E. and Laschi, C. (2009). Predictive tracking

across occlusions in the icub robot. In Humanoid

Robots, 2009, pages 486–491.

Ginhoux, R. and Gutmann, S. (2001). Model-based object

tracking using stereo vision. In ICRA, Volume: 2, cole

Nationale Suprieure de Physique de.

Kalal, Z., Mikolajczyk, K., and Matas, J. (2012). Tracking-

learning-detection. Pattern Analysis and Machine In-

telligence, IEEE Transactions on, 34(7):1409–1422.

Nelson, R. C. and Green, I. A. (2002). Tracking objects

using recognition. In In International Conference on

Pattern Recogntion, pages 1025–1030. Prentice Hall.

Pattacini, U. (2010). Modular Cartesian Controllers for

Humanoid Robots: Design and Implementation on the

iCub. Ph.D. dissertation, RBCS, IIT, Genova.

Rosten, E. and Drummond, T. (2006). Machine learning for

high-speed corner detection. In ECCV, pages 430–

443.

Ta, D., Chen, W., Gelfand, N., and Pulli, K. (2009). Surf-

trac: Efﬁcient tracking and continuous object recogni-

tion using local feature descriptors. In CVPR09.

Triesch, J., Ballard, D. H., and Jacobs, R. A. (2002). Fast

temporal dynamics of visual cue integration. Percep-

tion, 31(4):421–434.

Wiskott, L., Fellous, J., Kr

uger, N., and v.d. Malsburg,

C. (1997). Face recognition by elastic bunch graph

matching. IEEE Trans. Pattern Anal. Mach. Intell.

Yin, F., Makris, D., and Velastin, S. A. (2007). Perfor-

mance Evaluation of Object Tracking Algorithms. In

PETS2007.

HideandSeek-AnActiveBinocularObjectTrackingSystem

591