Fast Semi-automatic Target Initialization based on Visual Saliency for

Airborne Thermal Imagery

C¸a˘glar Aytekin

, Emre Tunalı

and Sinan

Middle East Technical University, Department of Electrical and Electronics Engineering, Ankara, Turkey

Image Processing Department, ASELSAN Inc. Microelectronics, Guidance and Electro-Optics Division, Ankara, Turkey

Keywords:

Real-time Target Initialization, Saliency, Distinctive Feature Selection, Long-term Tracking, Center-surround

Histogram Difference, Error Compensation.

Abstract:

In this study, a semi-automatic target initialization algorithm is introduced based on a recently proposed visual

saliency approach. First, a center-surround difference based initial window selection is utilized around the

input point coordinate provided by the user, in order to select the window which is most likely to contain

the actual target and background satisfying piecewise connectivity. Then, a recently proposed visual saliency

algorithm is exploited in order to detect the bounding box encapsulating the most salient part of the object.

The experiments support that the saliency based tracking window initialization is capable of handling marking

errors, i.e. erroneous user inputs, and boosts the performance of several tracking algorithms in terms of the

number of frames in which successful tracking is achieved, when compared with several ﬁxed window size

initializations.

1 INTRODUCTION

Target tracking is a classical problem and has many

important applications such as surveillance, activity

or behavior detection. Hence, a diverse set of track-

ing algorithms are proposed in the literature. Majority

of tracking algorithms assume predetermined target

location and size for initialization of tracking (Shi

and Tomasi, 1994; Sand and Teller, 2006; Ramanan

et al., 2007; Dowson and Bowden, 2005; Kwon and

Lee, 2010; Bibby and Reid, 2008; Grabner et al.,

2010; Collins et al., 2005a; Avidan, 2007; Grabner

and Bischof, 2006; Babenko et al., 2009; Grabner

et al., 2008; Stalder et al., 2009). In many appli-

cations, target size and location are required as in-

put from human-users. Therefore, target initialization

can drastically change the performance of the tracker

since this initial window determines for the tracker

what to track, i.e. the features (Shi and Tomasi,

1994; Sand and Teller, 2006), appearance (Ramanan

et al., 2007; Dowson and Bowden, 2005; Kwon and

Lee, 2010), contours (Bibby and Reid, 2008). Hence,

any insigniﬁcant or false information, i.e. parts of ob-

jects similar to common background or patches from

background, may result in a mislearning of target ap-

pearance. Some tracking algorithms (Grabner et al.,

2010; Collins et al., 2005a; Avidan, 2007; Grabner

and Bischof, 2006; Babenko et al., 2009; Grabner

et al., 2008; Stalder et al., 2009) try to deal with

this problem inherently: They classify the foreground

and background of the selected window by deﬁning

the regions close to selected window as foreground

samples, and the ones distant from the selected win-

dow as background priors. Still, this attempt does not

completely compensate the false initialization, specif-

ically in scenarios with high clutter or in crowded

scenes. Hence, false target initialization is still a prob-

lem. Indeed, in most of the real-time applications,

erroneous input is usually provided by the user due

to obligation to mark the target instantly. This er-

roneous input usually results in track losses prema-

turely. Therefore, if long-term tracking performance

desired to be achieved this erroneous input should be

compensated. Moreover, even in the case that user

provides a perfect bounding box or the center of the

target to be tracked, depending on the appearance of

the target; this initialization may not always be pre-

ferred. For example, in Fig. 1(a) an object with a

similar appearance with the background is illustrated.

An initialization like that of in Fig. 1(b) may result in

redundant features or deceptive appearance depend-

ing on the type of tracker, which may not provide

long-term tracking. Therefore, we propose that tar-

get should be selected as the most salient part of an

490

Aytekin Ç., Tunalı E. and Öz S..

Fast Semi-automatic Target Initialization based on Visual Saliency for Airborne Thermal Imagery.

DOI: 10.5220/0004668904900497

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 490-497

ISBN: 978-989-758-009-3

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

(a) (b) (c)

Figure 1: (a) A thermal image containing a car as a target,

(b) The actual bounding box including the whole target is

indicated as red, (c) The bounding box detected by the pro-

posed target initialization algorithm.

object, most distinctive segment from background, as

in Fig. 1(c), in order to achieve long-term tracking

performance.

In many industrial applications the initialization

of target needs to be done in real time and most of

the time the only user interaction is just a marking

operation on the object. Using only marking input,

many systems initialize target, being target center at

the marked point with predeﬁned discrete bounding

box sizes. In this study, a real-time target initializa-

tion framework is proposed which takes a single (x, y)

image coordinate from the user and returns the most

salient region bounding box in the neighborhood. It

is also shown that this initialization results in bet-

ter discriminative targets with respect to background

and provides a long-term tracking even in trackers

that claim to deal with discrimination problem during

tracking.

Automatic target initialization methods exist in

the literature, with several limitations. One main ap-

proach is motion detection (Veeraraghavan et al.,

2006) which cannot deal with stationary targets. A

more systematic approach is used in (Toyama and

Wu, 2000), however it requires a target model. There

are other existing methods (Mahadevan and Vascon-

celos, 2011; Yilmaz et al., 2003) that can handle pre-

viously mentioned issues. These algorithms process

exhaustive search over the entire image to extract tar-

gets which takes a lot of time, thereby making them

inappropriate for real-time applications. Hence, we

do not prefer exploiting these works and we follow a

semi-automatic approach instead, since we are con-

centrated on real-time applications. Furthermore, we

avoided using object window detector as (Alexeet al.,

2010) or segmentation algorithms with user interac-

tion (Rother et al., 2004), since we have obligation

to select the initial bounding box in real-time and we

only retain a coordinate input.

The rest of the paper is organized as follows: The

proposed target initialization method is explained in

Section 2, the conducted experiments are analyzed in

Section 3, ﬁnally the study is concluded in Section 4

where discussions were made.

2 PROPOSED METHOD

The target initialization method proposed in this study

consists of three main steps: First, an initial window

selection method based on center-surround histogram

difference is processed, then saliency map in this win-

dow is calculated by the method (Wei et al., 2012)

and the saliency map is thresholded. Finally, the

connected component having maximum total saliency

with minimum distance to the center is selected as ini-

tial target location. These steps are extensively ana-

lyzed in this chapter.

Figure 2: Three main steps of the proposed initial target

window selection algorithm.

2.1 Initial Window Selection

The motivation of selecting an initial window is to in-

clude the foreground and to provide a feasible back-

ground for the saliency evaluation method (Wei

et al., 2012) which assumes that most of the im-

age boundaries belong to the background and the

background patches are piecewise connected. This

makes the saliency detection sensitive to the image

boundaries; hence a proper window selection is re-

quired. The main approach to an initial window se-

lection is based on the well-known center-surround

histogram difference (CSD). First, we calculate the

center-surround histogram distances in windows of

multiple sizes around the pixel marked by the user.

In order to ensure the piecewise connectivity assump-

tion in the saliency detection, we choose the window

which gives the ﬁrst local maximum of histogram dis-

tance HD vector which is deﬁned as follows:

HD(i) = K (B

+ F

), (1)

where K is a distance measure between B

and F

which are the foreground and background histograms

of window w

and i = (1, 2, ., N).

Local maxima other than the ﬁrst one can have a

larger CSD. Actually, these maxima appear in win-

dows with layered background patches, where the

background region may correspond to only some

patches of real background, whereas the foreground

window may cover the actual foreground together

with another layer of background (See Fig. 3). How-

ever, we wish to obtain an initial window where back-

ground patches are piecewise connected, which may

not be case in some local maxima other than the ﬁrst

FastSemi-automaticTargetInitializationbasedonVisualSaliencyforAirborneThermalImagery

491

(a)

(b)

Figure 3: (a) Three window sizes and blue and red win-

dows indicating the foreground and background region bor-

ders respectively, (b) Histogram distance of foreground and

background regions obtained at each window number.

Figure 4: Foreground (blue),deadzone (red) and back-

ground borders for several window selections.

one. Therefore, we select the window corresponding

to ﬁrst local maximum of HD(i).

During the selection of initial window, various ob-

ject shapes are desired to be handled also. However,

since the foreground and background regions of each

window are selected as square boxes, objects with

shapes deviating from a square can be problematic.

For example, if foreground window is selected as en-

capsulating the whole target in Fig. 4, this square win-

dow also includes regions from backgroundwhich de-

crease CSD. To compensate this effect we introduce a

dead-zone region when calculating CSD. Empirically

we choose the foreground and background to cover

the 25 and 20 percent of the selected window area re-

spectively and the remaining area to be a dead-zone.

This selection is realized to help dealing with targets

with large deviation from square shape (Fig. 4).

Another issue is the calculation of histogram dis-

tances of foreground and background. For this pur-

pose, quadratic-chi histogram distance (Pele and

Werman, 2010) is utilized, since it suits to the prob-

lem in two aspects. First, it is capable of handling

quantization effects that occurs when close intensity

values are mapped into different histogram bins by

using similarity matrix which is taking care of cross-

bin relationship. Second, it suppresses the distances

resulting from bins with high values. Formulation of

quadratic-chi histogram distance is given in Eqn. 2

as:

(P, Q) = (2)

∑

− Q

)

(

∑

+ Q

)

− Q

)

(

∑

+ Q

)

where P and Q represent N dimensional nonnega-

tive bounded foreground and background histograms,

”i,j” are histogram bins, A is the nonnegative sym-

metric bin similarity matrix which is NxN and m is

the normalization factor retaining distance due to high

bin values.

2.2 Saliency Map Calculation

The saliency map of the window, selected by the al-

gorithm above, is extracted by a recently proposed

fast saliency extraction method (Wei et al., 2012) in

which the saliency problem is tackled from different

perspective by focusing on background more than the

object. Although there are various saliency detection

algorithms (Hou and Zhang, 2007; Achanta et al.,

2009; Goferman et al., 2010; Cheng et al., 2011), the

main motivation of using this method is its capability

of extracting a saliency map within few milliseconds;

however, it has two basic assumptions that should be

guaranteed, namely boundary and connectivity. The

boundary assumption is reﬂection of a basic tendency

that photographer/cameraman do not crop salient ob-

jects among the frame. Therefore, the image bound-

ary is usually background. The connectivity assump-

tion comes from the fact that background regions are

generally tend to be large and homogenous, i.e. sky,

grass. In other words, most image patches can be eas-

ily connected to each other piecewisely. Satisfying

these two conditions, the salient regions are assumed

to be the patches that are extracted by downscaling

or by any super pixel extraction algorithm with high

geodesic distance from the boundaries of the image

that is assumed to correspond to piecewise-connected

background regions. The geodesic saliency of a patch

p is the accumulated edge weights along the shortest

path from p to virtual background node b in an undi-

rected weighted graph p ∈ {υ, ε},

S(p) = min

=p, p

,..., p

n−1

∑

i=1

weight(p

, p

i+1

), (3)

s.t. (p

, p

i+1

) ∈ ε,

For this purpose a shortest path algorithm is ex-

ploited (Toivanen, 1996) in order to calculate the

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

492

shortest distance to the image boundaries from each

patch. The higher this value is the more salient the

patch. Furthermore, since patches close to the center

of the image requires a longer path in order to reach

the background, accumulation of weights tend to be

larger in the center patches. Therefore, this method

also favors the center image regions as more salient

which is reasonable since salient regions tend to oc-

cur around the center of image.

2.3 Binarization of Saliency Map

Since we look for a fast target initialization, we wish

to keep the computational cost at minimum even in

the binarization step. Hence, a fast binarization ap-

proach is proposed here exploiting the local max-

ima of saliency map. The threshold is selected by

a weighted average of local maxima of the saliency

map (Eqn. 4). In this sense, for fast binarization

process, fast local maxima detection is required. In

order to achieve fast local maxima detection, a fast

local maxima detection algorithm (Pham, 2010) is

used. After detection of local maxima, we form a

vector LocalMax

sorted

by sorting the local maxima

in descending order, and the normalized laplacian of

this vector is used as weights for local maxima. This

is meaningful since the local maxima with higher

laplacian represent a distinctive fall within local max-

ima. We shall favor these values when calculating

the threshold level since distinctive falls are indicators

of split between regions with higher saliency with re-

spect to their surroundings. Hence, a threshold would

be suitable for binarization around the most distinc-

tive fall; greatest weight is given to that local maxi-

mum in the weighted average of local maxima.

Thr = LocalMax

sorted

.∇

norm

(LocalMax

sorted

), (4)

where

∇

norm

( f) =

∇

( f) − min



∇

( f)



∑

∇

( f)|

− min(∇

( f))

, (5)

In order to sort the local maxima in a fast man-

ner, we generate a binary tree with heap property in

the phase of local maxima selection. Then, sorting is

accomplished in classical sense by selecting the ﬁrst

element,highest, and then re-ordering the heap at each

turn until all local maxima are sorted.

After thresholding the saliency map, the con-

nected component maximizing the regularization en-

ergy given by Eqn. 6, i.e. the most salient region with

minimum distance to the center, is selected as the tar-

get.

argmax

sqrt



− x

)

+ (y

− y

)



, (6)

where C

is the vectorized form obtained by raster

scanning the 2D label matrix with values 1 and 0

as foreground and background respectively, S is the

saliency map vectorized similarly and (x

, y

), (x

, y

)

are the centers of each connected component and the

inital window respectively.

Based on the explainations above, the entire initial

target window selection algorithm is summarized in

Algorithm 1.

Algorithm 1: Semi-Supervised Target Initialization.

Input: (x, y) coordinates from user indicating the target

location roughly

Initial window selection: Given input (x, y), select a

square window which is most likely to include actual target

and the background satisfying piecewise connectivity from

a set of windows W ∈ {w

, , w

};

Initialize Dist = {};

for W ∈ {w

, , w

} do

Obtain histogram of square foreground window

: 25% of area(W), centered at (x, y)

Obtain histogram of background window

: 20% of area(W), from the boundaries of W

Calculate quadratic-chi histogram distance

HD(W) = QC

, B

)

Dist = Dist ∪ HD(w);

end for

return window size that corresponds to 1st local

maximum of Dist as initial window size

Saliency Map Calculation: Given the initial window size

return saliency map using Eqn. 3

Binarization of Saliency Map: Given the saliency map

Find local maxima and sort in descending order to

obtain LocalMax

sorted

Calculate normalized laplacian of LocalMax

sorted

using Eqn. 5

Obtain the threshold Thr using Eqn. 4

Select the connected component satisfying Eqn. 6

Output: Target bounding box enclosing the selected

connecting component.

3 EXPERIMENTS

The proposed method was tested for two aspects with

two different procedures. For both stages, in ini-

tial window selection phase, nine window sizes from

20x20 to 100x100 with regular grid were used and his-

togram bin number was selected as 25. In the calcu-

lation of histogram difference, the parameters m and

bin similarity matrix A were selected empirically as m

FastSemi-automaticTargetInitializationbasedonVisualSaliencyforAirborneThermalImagery

493

Figure 5: Erroneous marking input (red dot) and the output

target window (blue bounding box).

is to be 0.5 and A to be in 5 bin neighborhood which

cover one ﬁfth of total number of bins. It should be

noted that increase in the number of bins of histogram

results in better resolution; therefore, better discrim-

inability between foreground and background. How-

ever, it also yields increase in computational com-

plexity which reduces time efﬁciency. Considering

the compromise between time efﬁciency and resolu-

tion, histogram bin number was decided to be 25. The

same consideration is also valid for the saliency map

calculation in which 2x2 patches were used for mean

intensity computation. Thus, these parameters may

be modiﬁed according to system in use.

At the ﬁrst experiment, proposed target initializa-

tion method and its error compensation capability, i.e.

robustness to marking errors, were tested. Ground

truth for performance evaluation was generated by 10

different users. 5 speciﬁc images

illustrated in Ap-

pendix were given to all users. They were kindly

requested to generate bounding box for anything in

the image that appears to them as target. After the

bounding box generation, we selected targets among

the images generated by at least 6 different users as

targets and the ground truth of target bounding boxes

were decided to be the average of all bounding boxes

generated by users. Input marks were given as the

summation of the center coordinates of ground truth

bounding box and uniformly distributed pseudo ran-

dom integers, representing marking errors, for each

direction as illustrated in Fig 5.

The noise amount was decided to be 5 pixels at

maximum for the targets that are smaller than 20 pix-

els, however for the targets which are greater than 20

pixels noise amount becomes one fourth of the target

size. For each selected target, the target initialization

algorithm was run for 10 times with erroneous input

and numbers of successful initializations were noted.

The success measure for initialization is the ratio of

intersection and union of the ground truth and the tar-

get bounding box (overlap measure) and successful

initialization assumed to be achieved when the mea-

sure is higher than 0.5 as it is used in many different

To obtain detailed results and the input images, please

contact with the authors.

Table 1: Performance of proposed method for target initial-

ization and marking error compansation.

Image& # of Succesful Target Max # of Fully Matched

Target No Initilization (for 10 trial) Target Window

1 1 10 10

2 10 10

3 10 10

4 9 9

5 10 10

6 10 10

7 0 X

8 9 2

9 10 10

10 10 10

11 6 5

12 9 2

13 0 X

2 1 10 10

2 9 9

3 10 10

4 9 5

5 7 6

6 0 X

7 10 10

8 10 10

3 1 0 X

2 0 X

3 9 7

4 10 10

5 0 X

4 1 10 10

2 10 8

3 10 10

4 10 10

5 10 10

6 10 10

7 10 10

8 10 10

9 9 3

10 10 10

11 0 X

12 10 10

13 10 10

14 0 X

5 1 10 8

2 7 7

3 10 6

4 5 5

5 0 X

6 10 7

7 3 3

8 10 6

9 10 6

Initialization Performance: 75.7

Robustness: 87.3

applications. Since, the purpose is to obtain exactly

same target bounding box for erroneous user input,

we counted the maximum number of target bounding

boxes which are exactly matched among the target

bounding boxes that are considered as successfully

initialized and their ratio is used for robustness mea-

sure. It should be noted that in most of cases (≈ 87%)

the input error was compensated by the proposed al-

gorithm and robust target initialization(exactly same

target bounding box) was achieved. Results of the

ﬁrst part are given in Table 1.

At the second stage of experiments we tested the

effect of our target initialization algorithm on tracking

by evaluating the tracking performance of three track-

ing algorithms (Grabner and Bischof, 2006; Grab-

ner et al., 2008; Stalder et al., 2009) in terms of the

number of frames that the tracker successfully tracks

against the condition that the distance from the center

of the object to be tracked to the target window center

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

494

does not change more than 5 pixels. The marking in-

put was given with a random error similar to previous

stage but in this case initialization and tracking was

executed only once. We compared the performance

of the proposed algorithm with that of ﬁxed window

size initializations, which is the case for many real-

time applications. 16x16, 32x32 and 64x64 window

sizes were used whose centers marked by the user. It

should be mentioned that initializing the target 5 pix-

els wider than extracted window by the proposed al-

gorithm at each direction helps the trackers to learn

the model of foreground better, hence we chose the

initial windows wider. The experiments were con-

ducted over various datasets from the work of Collins

et al. (Collins et al., 2005b) with several scenar-

ios which includes airborne videos. Before compar-

ing the tracking performances, we ﬁrst analyzed the

location of the initialization window determined by

the proposed algorithm when initial marks provided

by the user were erroneous, i.e. far from the cen-

ter of the target. In Fig. 6 the marks and the pro-

posed target windows are illustrated. As it is real-

ized, the algorithm is quite tolerant to erroneous input

and achieves high performance of selecting the salient

part of the target. On the other hand, when the track-

ing window was selected with ﬁxed size target initial-

ization, the initial window tended to include various

parts of background image, decreasing the tracking

performance. The tracking performance evaluations

in terms of unsuccessful frame numbers are given in

Table 2. In this sense, 0 means target is successfully

tracked throughout the scenario.

We observed that the proposed initialization

achieved high mean performance over the scenarios

when compared with that of each ﬁxed size initial-

izations, and it gave comparable results with each

single window initialization almost in each scenario.

The performance boost achieved with the proposed

algorithm results from handling a subset of scenarios

very well such as crowded scenes, where target is sur-

rounded with many objects; or a small part of back-

ground around the neighborhood of the target includ-

ing very strong, discriminative features. In the former

case, a erroneous initialization may result in tracking

”jumps” where the tracker starts to track the other tar-

get, whereas in the latter case, the tracker may learn

the discriminative features as the ones that belong to

background and not the actual target. Both cases oc-

cur if the window is selected large or localized at an

erroneous position from wrong marking. On the other

hand, a small window initialization may prevent the

tracker to learn the features discriminative enough to

track the target, resulting in the loss of track. The

proposed initialization is able to compensate for such

Figure 6: Some examples of target initialization windows:

(black)64x64 window, (green) 32x32 window, (red) 16x16

window, (blue) window extracted by the proposed method

and the red asterisk is the click input provided by the user.

effects and achieves good performance. Furthermore,

in cases of occlusion and illumination changes, we

observed that the tracker is more likely to redetect the

target and continues tracking. Although redetecting

target is a trait of the tracking algorithm itself, a more

discriminative initial target window selection is ob-

served to help boosting the performance in these sce-

narios.

The algorithm is applicable to real-time and a

C++ implementation of the proposed target initial-

ization algorithm takes about 12 ms on average in

TMS320C6713 Floating-Point Digital Signal Proces-

sor @270 MHz which is enough to achieve target ini-

tialization in a frame for 60 fps systems.

4 CONCLUSIONS

In this work, we have shown that target initializa-

tion can dramatically change the performance of the

tracker, since the initial window determines for the

tracker what to track. In order to achieve a better

tracking performance; we proposed a fast, saliency

based algorithm for target initialization. Performance

boost of tracker is mainly based on two key fea-

tures of target initialization algorithm: It is capable

of compensating erroneous user input; also select-

ing the most distinctive, salient part of object as tar-

get, so better discrimination is achieved between the

target and background. Experimental results show

that tracking performance is boosted in scenarios, in

which the tracking is initialized by the proposed algo-

rithm. Very low computational cost and requirement

of only a point coordinate as input in the neighbor-

hood of the target make this approach preferable in

real time tracking applications.

FastSemi-automaticTargetInitializationbasedonVisualSaliencyforAirborneThermalImagery

495

Table 2: Performance comparison of proposed method and ﬁxed sized windows for three different trackers.

BooostingTracker SemiBooostingTracker BeyondSemiBooostingTracker

Scenario (Grabner and Bischof, 2006) (Grabner et al., 2008) (Stalder et al., 2009)

Initial

Sequence Frame Frames Proposed 16x16 32x32 64x64 Proposed 16x16 32x32 64x64 Proposed 16x16 32x32 64x64

1.pktest01 0 450 315 310 307 210 274 314 299 191 157 178 275 241

2.pktest01 1110 350 76 316 0 155 14 74 63 194 80 160 78 344

3.pktest02 0 470 431 434 433 438 213 437 376 197 202 206 231 193

4.pktest02 770 450 0 0 0 388 360 372 275 373 0 192 102 33

5.pktest02 1185 330 32 279 96 103 5 8 52 262 84 148 273 116

6.egtest03 0 300 0 300 267 16 237 300 300 240 29 300 39 72

7.pktest03 290 230 191 190 185 188 196 192 103 184 139 197 139 199

8.egtest01 0 150 0 150 150 0 0 150 150 0 59 150 150 40

9.egtest03 0 150 0 11 137 77 52 57 141 71 47 150 74 38

10.pktest03 0 415 86 86 80 122 83 88 118 210 94 410 88 186

Average number of frames

with track loss 116.11 221.11 175.00 175.00 150.11 211.56 195.44 190.22 88.56 186.78 151.22 141.78

Although a dead zone was introduced in order to

deal with elongated objects, the proposed window se-

lection method may not be effective for all elongated

objects. Speciﬁcally, when signiﬁcant amount of the

object pixels ﬂood into the background zone together

with background pixels in foreground zone affect tar-

get initialization adversely and may yield erroneous

target initialization, which may be the case for elon-

gated objects. Even though the experiments are exe-

cuted in thermal data sets in which target objects has

smooth transition due to heat diffusion equation, the

suggested solution may be well generalized to other

imaging devices.

REFERENCES

Achanta, R., Hemami, S., Estrada, F., and Susstrunk, S.

(2009). Frequency-tuned salient region detection.

In Computer Vision and Pattern Recognition, 2009.

CVPR 2009. IEEE Conference on, pages 1597–1604.

Alexe, B., Deselaers, T., and Ferrari, V. (2010). What is an

object? In Computer Vision and Pattern Recognition

(CVPR), 2010 IEEE Conference on, pages 73–80.

Avidan, S. (2007). Ensemble tracking. Pattern Analy-

sis and Machine Intelligence, IEEE Transactions on,

29(2):261–271.

Babenko, B., Yang, M.-H., and Belongie, S. (2009). Vi-

sual tracking with online multiple instance learning.

In Computer Vision and Pattern Recognition, 2009.

CVPR 2009. IEEE Conference on, pages 983–990.

Bibby, C. and Reid, I. (2008). Robust real-time visual track-

ing using pixel-wise posteriors. In Proceedings of

the 10th European Conference on Computer Vision:

Part II, ECCV ’08, pages 831–844, Berlin, Heidel-

berg. Springer-Verlag.

Cheng, M., Zhang, G., Mitra, N. J., Huang, X., and Hu,

S. (2011). Global contrast based salient region detec-

tion. In Proceedings of the 2011 IEEE Conference on

Computer Vision and Pattern Recognition, CVPR ’11,

pages 409–416, Washington, DC, USA. IEEE Com-

puter Society.

Collins, R., Liu, Y., and Leordeanu, M. (2005a). On-

line selection of discriminative tracking features. Pat-

tern Analysis and Machine Intelligence, IEEE Trans-

actions on, 27(10):1631–1643.

Collins, R., Zhou, X., and Teh, S. K. (2005b). An open

source tracking testbed and evaluation web site. In

IEEE International Workshop on Performance Evalu-

ation of Tracking and Surveillance (PETS 2005), Jan-

uary 2005.

Dowson, N. D. H. and Bowden, R. (2005). Simultaneous

modeling and tracking (smat) of feature sets. In Com-

puter Vision and Pattern Recognition, 2005. CVPR

2005. IEEE Computer Society Conference on, vol-

ume 2, pages 99–105.

Goferman, S., Zelnik-Manor, L., and Tal, A. (2010).

Context-aware saliency detection. In Computer Vision

and Pattern Recognition (CVPR), 2010 IEEE Confer-

ence on, pages 2376–2383.

Grabner, H. and Bischof, H. (2006). On-line boosting

and vision. In Computer Vision and Pattern Recog-

nition, 2006 IEEE Computer Society Conference on,

volume 1, pages 260–267.

Grabner, H., Leistner, C., and Bischof, H. (2008). Semi-

supervised on-line boosting for robust tracking. In

Proceedings of the 10th European Conference on

Computer Vision: Part I, ECCV ’08, pages 234–247,

Berlin, Heidelberg. Springer-Verlag.

Grabner, H., Matas, J., Van Gool, L., and Cattin, P. (2010).

Tracking the invisible: Learning where the object

might be. In Computer Vision and Pattern Recogni-

tion (CVPR), 2010 IEEE Conference on, pages 1285–

1292.

Hou, X. and Zhang, L. (2007). Saliency detection: A spec-

tral residual approach. In Computer Vision and Pat-

tern Recognition, 2007. CVPR ’07. IEEE Conference

on, pages 1–8.

Kwon, J. and Lee, K. M. (2010). Visual tracking decompo-

sition. In CVPR, pages 1269–1276.

Mahadevan, V. and Vasconcelos, N. (2011). Automatic ini-

tialization and tracking using attentional mechanisms.

In Computer Vision and Pattern Recognition Work-

shops (CVPRW), 2011 IEEE Computer Society Con-

ference on, pages 15–20.

Pele, O. and Werman, M. (2010). The quadratic-chi

histogram distance family. In Proceedings of the

11th European conference on Computer vision: Part

II, ECCV’10, pages 749–762, Berlin, Heidelberg.

Springer-Verlag.

Pham, T. Q. (2010). Non-maximum suppression using

fewer than two comparisons per pixel. In Blanc-Talon,

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

496

J., Bone, D., Philips, W., Popescu, D., and Scheun-

ders, P., editors, Advanced Concepts for Intelligent Vi-

sion Systems, volume 6474 of Lecture Notes in Com-

puter Science, pages 438–451. Springer Berlin Hei-

delberg.

Ramanan, D., Forsyth, D., and Zisserman, A. (2007).

Tracking people by learning their appearance. Pat-

tern Analysis and Machine Intelligence, IEEE Trans-

actions on, 29(1):65–81.

Rother, C., Kolmogorov, V., and Blake, A. (2004). ”grab-

cut”: interactive foreground extraction using iterated

graph cuts. In ACM SIGGRAPH 2004 Papers, SIG-

GRAPH ’04, pages 309–314, New York, NY, USA.

ACM.

Sand, P. and Teller, S. (2006). Particle video: Long-

range motion estimation using point trajectories. In

Computer Vision and Pattern Recognition, 2006 IEEE

Computer Society Conference on, volume 2, pages

2195–2202.

Shi, J. and Tomasi, C. (1994). Good features to track. In

Computer Vision and Pattern Recognition, 1994. Pro-

ceedings CVPR ’94., 1994 IEEE Computer Society

Conference on, pages 593–600.

Stalder, S., Grabner, H., and Van Gool, L. (2009). Beyond

semi-supervised tracking: Tracking should be as sim-

ple as detection, but not simpler than recognition. In

Computer Vision Workshops (ICCV Workshops), 2009

IEEE 12th International Conference on, pages 1409–

1416.

Toivanen, P. J. (1996). New geodesic distance trans-

forms for gray-scale images. Pattern Recogn. Lett.,

17(5):437–450.

Toyama, K. and Wu, Y. (2000). Bootstrap initialization of

nonparametric texture models for tracking. In Pro-

ceedings of the 6th European Conference on Com-

puter Vision-Part II, ECCV ’00, pages 119–133, Lon-

don, UK, UK. Springer-Verlag.

Veeraraghavan, H., Schrater, P., and Papanikolopoulos, N.

(2006). Robust target detection and tracking through

integration of motion, color, and geometry. Computer

Vision and Image Understanding, 103(2):121–138.

Wei, Y., Wen, F., Zhu, W., and Sun, J. (2012). Geodesic

saliency using background priors. In Proceedings of

the 12th European conference on Computer Vision -

Volume Part III, ECCV’12, pages 29–2, Berlin, Hei-

delberg. Springer-Verlag.

Yilmaz, A., Shaﬁque, K., and Shah, M. (2003). Target

tracking in airborne forward looking infrared imagery.

Image and Vision Computing, 21(7):623 – 635.

APPENDIX

(a)

(b)

(c)

(d)

(e)

Figure 7: Thermal image dataset.

FastSemi-automaticTargetInitializationbasedonVisualSaliencyforAirborneThermalImagery

497