Object Tracking using Correction Filter Method with Adaptive

Feature Selection

Xiang Zhang, Yonggang Lu and Jiani Liu

School of Information Science and Engineering, Lanzhou University, Lanzhou, Gansu, China

Keywords: Correlation Filters, Lab Color Space, Confidence Score, Feature Selection.

Abstract: Correlation filter based tracking algorithms have shown favourable performance in recent years. Nonetheless,

the fixed feature selection and potential model drift limit their effectiveness. In this paper, we propose a novel

adaptive feature selection based tracking method which keeps the strong discriminating ability of the

correlation filter. The proposed method can automatically select either the HOG feature or color feature for

tracking based on the confidence scores of the features in each frame. Firstly, the response map of the color

features and the HOG features are extracted respectively using correlation filter. The Lab color space is used

to extract the color features which separate the luminance from the color. Secondly, the confidence region

and the possible location of the target are estimated using the average peak-to-correlation energy. Thirdly,

three criteria are used to select the proper feature for the current frame to perform tracking adaptively. The

experimental results demonstrate that the proposed tracker performs superiorly comparing with several state-

of-the-art algorithms on the OTB benchmark datasets.

1 INTRODUCTION

Visual object tracking is a very active part of the

domain in computer vision, which has many

applications in areas such as surveillance, automation

and robotics (Wu, Lim and Yang, 2013). It is often

used to track the object in a series of frames when a

target is located in one of them. In many visual

tracking tasks, the location of the target is known only

in the first frame, and the estimation of the possible

locations in other frames need to be made. Many

existing algorithms assume that the target location

changes little over time, and determine the target

within a search window centered at the previous object

location, which can be referred to as a motion model

(Wang, Shi, Yeung and Jia, 2015). However, these

algorithms may not be suitable for handling some

complex scenarios, such as illumination variation,

scale variation, occlusion and scale variations

(Marvasti-Zadeh, Seyed Mojtaba and Cheng, 2019).

Visual object tracking has many algorithms to

learn an appearance model of the target using either

generative or discriminative methods. The generating

model (Kwon and Lee, 2011; Liu, Huang, Yang and

Kulikowski, 2011; Sevilla-Lara and Learned-Miller,

2012), such as sparse coding model (Mei, Ling, Wu

and Blasch, 2013), transforms the problem of object

tracking into the problem of sparse approximation,

such as CT tracker (Zhang, Zhang and Yang, 2012).

When the noise level is high, the tracking process is

prone to the drift of the target frame. Most tracking

algorithms use the discriminative methods, and the

main idea is to train an online updated classifier,

which gives the target position in each frame. In this

kind of method, the correlation filter technology

performs the complex convolution operation

efficiently in the frequency domain, which improves

the timeliness of target tracking, so it has attracted

much research interest in the research field. In recent

years, there are many algorithms showed a fusion of

correlation filtering and deep learning. Depth features

use different depth convolution features of

convolution neural networks, which contain more

advanced semantic information. Based on the

correlation filtering algorithm of depth features, the

target location is determined by calculating the

relevant confidence responses of convolution layer

features of different depths (Li, Ma and Wu, 2019).

Although the depth convolution feature has a strong

ability to identify the target, it is hard to understand

the feature transformation in the “black box”, and it

also introduces a high computational complexity such

that the resulting algorithm cannot achieve the real-

480

Zhang, X., Lu, Y. and Liu, J.

Object Tracking using Correction Filter Method with Adaptive Feature Selection.

DOI: 10.5220/0010196604800487

In Proceedings of the 10th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2021), pages 480-487

ISBN: 978-989-758-486-2

time performance (Wang, Zhou, Tian, Hong and Li,

2018). So, it is still necessary to study the correlation

filter technology without using the depth convolution

features.

The discriminative model mainly distinguishes

the target from the background by training the

classifier. The convolution theorem shows that the

time-consuming convolution operation can be

converted into an efficient element dot product

operation in the Fourier domain. Based on the

observation, the correlation filtering technology is

introduced into object tracking. Bolme et al. (2010)

applied the correlation filtering algorithm to the target

tracking task for the first time, and proposed the

minimum mean square error (MOSSE) tracking

algorithm which can perform tracking with high

speed. However, the tracking accuracy of the

algorithm is not ideal because of the grayscale

characteristics of the single channel. Henriques et al.

(2012, 2015) added kernel function to the framework

of correlation filtering algorithm, replaced the

grayscale feature of single channel with multi-

channel HOG feature, and proposed kernel

correlation filter (KCF) tracking algorithm. Through

the cyclic matrix, the sampling problem in the

training stage of the correlation filter template is

equivalent to the cyclic shift operation of the Eigen

matrix to complete the intensive sampling of the

training samples. The cyclic sampling effectively

increases the number of samples, and the robustness

of the tracker is further improved. As for the scale

adaptation, Danelljan et al. (2014) put forward the

DSST method on the basis of MOSEE, which uses

two single filters to estimate the position and scale.

While the single feature or multi-feature fusion is

usually used to feature extraction in the most trackers,

Danelljan et al. (2014a, 2014b) used characteristic

Color Names to extend CSK algorithm, which have

obtained good results in the tracking of color video

sequence. In the aspect of multi-feature fusion, HOG

features and Color features are superimposed directly

in a work (Li and Zhu, 2014). The Staple algorithm

(Bertinetto, Valmadre and Golodetz, 2016) fuses

HOG features and color histogram features

(Possegger, Mauthner and Bischof, 2015) with

weighting proportions of 0.7, which effectively

promotes the robustness of tracking, but lacks

pertinence to specific video scenes.

In the process of tracking, if the discrimination

power of one feature is very different from another

feature, using the two features together may result in

low tracking accuracy. To deal with this problem, a

correlation filter tracker based on adaptive feature

selection using the confidence score of response map

is proposed. In this paper, we calculate the confidence

score of the response map using correlation filter for

each feature (HOG feature and color feature in Lab

color space), which can be used to select the features

adaptively. For example, when the object is deformed,

the HOG feature will be affected a lot. The color

feature that is less affected than the HOG features will

be selected for tracking. In the other case when the

color feature of the object is disturbed a lot, the HOG

feature will be selected for tracking. The difference

between this method and other correlation filtering

methods is that the more suitable features are selected

according to the actual situation of each frame, which

increases the flexibility in the feature selection in the

complex scene, and can describe the target more

accurately. The algorithm is tested on the evaluation

criteria of OTB benchmark dataset and 100 video

sequences, and its performance is compared with

many mainstream algorithms.

The architecture of this paper is arranged as

follows: Some related works are briefly reviewed in

Section 2. Then, we introduce the proposed method

based on the confidence score in HOG feature and

color feature in Lab color space in Section 3. The

experiment results are shown in Section 4. Finally,

conclusions are given in Section 5.

2 RELATED WORK

2.1 Correlation Filters

Correlation filtering was first used in signal

processing to describe the correlation between two

signals. Initially, correlation filtering was applied to

visual object tracking in grayscale images by Bolme

et al. (2010). Later, the extension to multiple feature

channels and HOG feature achieved the state of the

art performance (Henriques, Caseiro, Martins, 2012).

In 2014, DSST (Danelljan, Häger and Khan, 2014), a

multi-scale template for Discriminative Scale-Space

Tracking on the basis of MOSSE is used to deal with

the scale change, with two filters to track the position

change and the scale change respectively. The

position filter is used to determine the new target

position, and the scale filter is used for scale

estimation. One deficiency of Correlation Filters is

that they are constrained to learn from all circular

shifts. Several recent works have sought to resolve

this issue, and one of work, called Spatially

Regularized Formulation (SRDCF) (Danelljan,

Hager and Shahbaz, 2015), has demonstrated

excellent tracking results. However, this

improvement is achieved at the cost of real-time

Object Tracking using Correction Filter Method with Adaptive Feature Selection

481

operation. The Staple algorithm (Bertinetto,

Valmadre and Golodetz, 2016) propose a simple

combination of template and histogram scores that are

learnt independently to preserve real-time operation.

The resulting tracker outperforms significantly more

complex state-of-the-arts trackers in several

benchmarks, but there is still the problem of

sacrificing some of the characteristic advantages.

The tracking-by-detection paradigm is used in the

Staple algorithm to calculate the response graph

matrix in the correlation filters as:









,, t-1

F,;

wh wh t

tFTxp





(1)

where x

represents the t-th frame image, w and h

represent the position in the frame, T represents an

image transformation function, p represents a

rectangular window which gives a target location in

, F represents a candidate target, and θ is the model

parameter. Let S

represents the set of candidate

targets of the t-th frame, the score F(w, h) of all

candidate target F ∊ S

is the response map matrix of

the frame. The parameter θ is calculated by:









t-1 1

=argmin {L ; + R }









(2)

where the L(θ; x

t-1

) represents the loss function,



tii

XXxp



represents the sequence of historical

frame tracking results, R(θ) represents a regular term

to prevent overfitting and the λ is regular term

parameter. The loss function is a linear combination

of all sample losses, as shown in (3) and (4):

  

,,,

tiii

Lx xp













(3)

 



,, , , ;

x p d p argmax f T x q









(4)

where d(p, q) defines the cost of choosing rectangle q

when the correct rectangle is p. In our paper, we use

the h and β represent the specific model parameters of

HOG feature and color feature respectively which is

solved by two independent ridge regression problems

such as (5) and (6):



hargmin ;

h temp1 t temp1

hX h















(5)



arg min ;

temp2 t temp















(6)

2.2 Color Features in Lab Color Space

At present, the input and output of most image

capture devices are based on RGB color space, the

traditional method is to covert the color images into

RGB color space and extract the color features. But

the three channels in RGB color space contain

luminance information, and there is a strong

correlation among them. Therefore, it may not be

possible to obtain the desired effect by using these

directly. Different from the most common RGB color

space, Lab color space does not rely on light or

pigments. It is a color space determined by the

International Lighting Committee (CIE), which can

theoretically describe all colors in nature. In the 3-

channel component of Lab color space, L represents

luminance whose range is [0, 100] and the luminance

increases with the numerical value. The ranges of a

and b channel are both [-128, 127]. In our

experiments, the color features in the Lab color space

is used for the object tracking.

3 THE PROPOSED APPROACH

In this section, we first present the problem

formulation of the adaptive feature selection method

based on the confidence score. Then, we design an

algorithm that builds up a bridge between the known

correlation filter and our problem formulation.

How to judge whether the tracking result is

accurate is a very important problem, because this

determines the update strategy of the model. Many

algorithms, such as KCF, DSST, SRDCF, Staple, do

not judge the reliability of tracking results, and the

results of each frame are updated either immediately

or every N frames. This is unreliable, especially when

the target is occluded, or the tracking has not followed

well, and then update the model, which will only

make the tracker more and more unable to recognize

the target, which is the problem of model drift. In the

correlation filter tracking, ideally, the response map

of each frame is a single peak, and the corresponding

position of the maximum value of the response map

is the tracking result position of the current frame.

The sharper the single spike in the response map

means that the tracking result is more credible;

conversely, the flatter the single spike means the less

credible the tracking result is. However, because the

actual scene is disturbed by complex factors, the

actual response map may be multi-peak, and the

correct target position can be at the highest peak of

the response map. It can also be at the secondary peak

or other peak.

In order to evaluate the effectiveness accurately,

the average peak-to-correlation energy and the

maximum response score F

max

of the response map

ICPRAM 2021 - 10th International Conference on Pattern Recognition Applications and Methods

482

are used together. The F

max

is defined as:



,t-1

=,;

max w h t

FmaxFTxp





(7)

Generally, the value of the highest point F

max

response map represents the result of center position,

but, if only this condition is used for feature selection,

the phenomenon of model drift will occur when there

are multiple peaks in the response map. The F

max

does

not reflect the degree of oscillation of the response

map, a new criterion called average peak-to-

correlation energy (APCE) is used, which is:

max min

,min

(( ))

APCE

mean F F







(8)

where F

max

, F

min

, F

w,h

represents respectively the

response at the highest, lowest and the w-th row and

h-th column scores of F(w, h). This criterion can

reflect the degree of oscillation of the response map.

When the APCE suddenly changes, the target is

occluded, or the target is lost in this feature which is

unreliable in the frame.

When we begin a new tracking using the

discriminative methods, the goal is to learn a suitable

classifier which can extract the target from the

background in real time. In the first frame, the

location of the target is known. It can directly get the

HOG feature and color feature using the correlation

filter. In this paper, we use the confidence score to

evaluate the reliability of each feature and select the

suitable feature adaptively during the target tracking.

Assume there are t frames X = [X

, X

, …, X

Give the location of the target in X

, we learn a

discriminative correlation filters in HOG feature and

get the score F

HOG

at each pixel in the response map

with (1). Then the color of RGB image is converted

to the Lab color space. We also learn a correlation

filters in color feature to computed the score F

color

with (1). Because the location of the target in the first

frame is known, the (9) is used to get a standard

confidence score M

HOG

and M

color

in the first frame to

measure the reliability of the following frame. The

threshold is set to evaluate the reliability of the next

frame, which is:

30%*

HOG HOG

color color

threshold M





(9)

If the score of the next frame is greater than the

threshold, the frame is considered to be a valid frame

and added to the benchmark queue, otherwise the

frame is considered invalid, which will cause the

model drift and is discarded. The benchmark queue

stores the scores of three valid frames and takes the

average score as the benchmark:

base

PCE





(10)

In this paper, we use three indicators to select the

optimal features adaptively. First, the score of the

response map APEC represents the confidence degree

of the current frame. We put the two kinds of scores

under the same standard and compare them. Because

the M

HOG

and M

color

is the standard score, we set a ratio

to transform the APCE

color

to APCE

HOG

, which is:

color HOG color HOG

color

APCE M M APCE

ratio







(11)

If the ratio>0, the color feature is better than the

HOG feature. Second, a rate of changing the score is

set to measure the confidence degree, which is:

base

PCE S

change_rate





(12)

When the rate of the HOG feature is greater than

the color feature, we choose the stable feature of color

feature in this case. Otherwise, the other one will be

selected. Finally, we calculate the position offset of

the maximum value of the response map F

max

in the

current frame from the previous frame, which is:

 

tt-1 1

x-x



  



(13)

Because a normal tracking changes smoothly

between neighboring frames, if there is a sudden

change in the offset of the maximum value, the model

drift is considered to have occurred. We choose the

smaller distance of each feature by comparing it

separately. A confidence score is designed to measure

the three criteria to select the suitable feature, which is:

  

color

=_ _ ,

color HOG

HOG

C Score Sign Sign Sign

ratio

change rate change rate























 



，



(14)

where the Sign(x) is the symbolic function. In the

definition of the C_Score, if the score is positive, the

Object Tracking using Correction Filter Method with Adaptive Feature Selection

483

color feature is better than the HOG feature in at least

two criteria, we select the color feature in the current

frame for tracking and model update. Otherwise, the

HOG feature is selected. An overview of the proposed

method is summarized in Algorithm 1.

Algorithm 1: The proposed tracking algorithm.

Input: Frames {It}, initial target location p

Output: Target locations of each frame {p

, t ≠1}

1: REPEAT

2: Crop an image region from the las

location p

t−1

and extract its response map in

HOG feature.

3: Convert the RGB image to the Lab colo

space and extract the response map.

4: Calculate the F

max

and the APCE fro

the response map via (7) and (8).

5: Calculate three adaptively selected

conditions via (11), (12) and (13).

6: IF the C_Score via (14) is positive

THEN

7: Select the response map of colo

feature to update the model

8: ELSE

9: Select the response map of HOG

feature to update the model

10: END IF

11: UNTIL end of video sequence.

4 EXPERIMENTAL RESULTS

AND ANALYSIS

In this section, the proposed method is verified on two

benchmark datasets, OTB13 (Wu, Lim and Yang,

2013) and OTB15 (Wu, Lim and Yang, 2015), and

compared with several recent algorithms such as

Staple, LMCF (Wang, Liu and Huang, 2017), SRDCF,

DSST (Danelljan, Hager and Khan, 2017) and KCF.

It is implemented in MATLAB with an Intel Core i7

3.60GHz CPU and 8GB of RAM.

We follow the evaluation standard provided by the

benchmark OTB15 which includes 100 video

sequences with various targets and backgrounds. In

OTB15, four indices are used to evaluate all the

compared algorithms with one-pass evaluation (OPE)

such as bounding box overlap, center location error,

distance precision and overlap precision. We evaluate

the trackers according to the result with an error

threshold of 20 pixels for the precision plots. For the

success plots, the trackers are evaluated by the AUC

scores.

Figure 1 shows the performance of our method

with the other correlation filters on OTB-15. The

proposed performs significantly better than the other

methods. In the precision plot, our tracker performs 6%

better than the Staple algorithm. The tracker also

shows 7% better than the Staple in the success plot.

For a more specific analysis, the performance of our

tracker approach can be affected by several

challenges as shown in Figure 2. It shows the

performance of tracking method for various

challenging attributes provided in the benchmark

OTB-15 such as Illumination Variation (IV), Scale

Variation (SV), Occlusion (OCC), Deformation

(DEF), Motion Blur (MB), Fast Motion (FM), In-

Plane-Rotation (IPR), Out-of-Plane-Rotation (OPR),

Out-of-View (OV), Background Clutters (BC), Low

Resolution (LR). Our method is effective in BC, DEF,

IV, OPR and OCC compared to the existing

approaches and more robust than the compared

trackers for deformable object. Because the color

feature

of our algorithm is in Lab space, it can better

Figure 1: The precision plots(left) and success plots(right) of OPE on OTB-15. The numbers in the legend shows the precision

scores and AUC scores for each tracker.

ICPRAM 2021 - 10th International Conference on Pattern Recognition Applications and Methods

484

Figure 2: The success plots for 11 challenging attributes including background clutter, illumination variation, occlusion,

deformation, out-of-plane rotation, out-of-view, scale variation, in-plane rotation, motion blur, fast motion, low resolution.

The proposed tracker performs best or second best in almost all the attributes.

recognize the change of the different frame and

enhance the tracking effect than other correlation

filtering algorithms in RGB color space in IV. In

addition, if the color feature does not change much,

the HOG feature can enhance the tracking effect in

the DEF. It can also be seen from Figure 2 that our

tracker achieves better performance than the other

trackers when the object suffers from fast motion and

background clutters.

Figure 3 shows the results of different tracking

algorithms. It can be seen that the proposed approach

can handle different situations well. There are four

types of video tracking results shown in Figure 3. The

first row is the skiing sequence which represents a

deformation scenario. In the video sequence the

athlete is in a state of high-speed and deformation.

The color feature of the player changes obviously in

each frame and the HOG feature is not useful. The

proposed method automatically selects the color

features

for tracking. Compared with other trackers,

Object Tracking using Correction Filter Method with Adaptive Feature Selection

485

Figure 3: Several representative frames of different tracking algorithms. Our tracker exhibits robustness in challenging

scenarios like deformation (row 1), fast motion (row 2), background clutter (row 3), occlusion (row 4). These sequences come

from the OTB15 benchmark (Skiing, Deer, Soccer and Basketball). The red rectangle indicates the bounding box obtained

by the proposed tracker.

it does not fuse the color feature with the HOG feature

and improves the tracking effect. The second row is

the deer sequence which represents the fast motion

scenario. In the video sequence a deer runs in a high

speed. The color feature is also automatically selected

by the proposed method. The third row shows the

results of the soccer sequence which represents a

background clutters scenario. The target is the player

dressed in red in the video sequence. From 70th frame

to 113th frame, the background color also becomes

red, so the APEC score of the color feature is lower

than that of the HOG feature. The proposed tracker

only uses the HOG feature for tracking and produces

more accurate tracking results, while the other

trackers use both the color feature and HOG feature

and the tracking performance is deteriorated,

resulting in the loss of tracking targets. The fourth

row is the basketball sequence which represents an

occlusion scenario. The target is partially or fully

occluded in the video sequence. When the player is

occluded, the color feature is affected and the HOG

feature is automatically selected by the propose

method to deal with this situation.

These results demonstrate that the proposed

tracker is able to handle various categories of objects

by selecting the feature adaptively.

5 CONCLUSIONS

In this paper, to improve the precision and success in

target tracking, a method is proposed to select a

feature adaptively based on a confidence score. The

confidence score can be used to select the most

suitable feature for tracking at each frame adaptively.

The adaptive feature selection at the frame level is

shown to be effective for improving the robustness in

tracking. One disadvantage of the method is the

introduction of the additional computation of the

confidence score, although the real time tracking can

still be performed. We will continue to work on

improving the speed of the proposed method in the

future work.

ICPRAM 2021 - 10th International Conference on Pattern Recognition Applications and Methods

486

ACKNOWLEDGMENT

This work is supported by the National Key R&D

Program of China (Grants No. 2017YFE0111900,

2018YFB1003205).

REFERENCES

Wu, Y. , Lim, J. , & Yang, M. H. . (2013). Online Object

Tracking: A Benchmark. Computer Vision & Pattern

Recognition. IEEE.

Wang, N. , Shi, J. , Yeung, D. Y. , & Jia, J. . (2015).

Understanding and diagnosing visual tracking systems.

Marvasti-Zadeh, Seyed Mojtaba & Cheng, Li & Ghanei-

Yakhdan, Hossein & Kasaei, Shohreh. (2019). Deep

Learning for Visual Tracking: A Comprehensive Survey.

Sevilla-Lara, L. , & Learned-Miller, E. . (2012).

Distribution fields for tracking. Computer Vision &

Pattern Recognition. IEEE.

Liu, B. , Huang, J. , Yang, L. , & Kulikowski, C. A. . (2011).

Robust tracking using local sparse appearance model

and K-selection. IEEE Conference on Computer Vision

& Pattern Recognition. IEEE.

Kwon, J. , & Lee, K. M. . (2011). Tracking by Sampling

Trackers. International Conference on Computer Vision.

IEEE.

Mei, X. , Ling, H. , Wu, Y. , & Blasch, E. P. . (2013).

Efficient minimum error bounded particle resampling

l1 tracker with occlusion detection. IEEE Transactions

on Image Processing, 22(7), 2661-2675.

Zhang, K. , Zhang, L. , & Yang, M. H. . (2012). Real-Time

Compressive Tracking. European Conference on

Computer Vision. Springer, Berlin, Heidelberg.

Li, X. , Ma, C. , Wu, B. , He, Z. , & Yang, M. H. . (2019).

Target-Aware Deep Tracking. 2019 IEEE/CVF

Conference on Computer Vision and Pattern

Recognition (CVPR). IEEE.

Wang, N. , Zhou, W. , Tian, Q. , Hong, R. , & Li, H. . (2018).

Multi-Cue Correlation Filters for Robust Visual

Tracking. 2018 IEEE/CVF Conference on Computer

Vision and Pattern Recognition (CVPR). IEEE.

Bolme, D. S. , Beveridge, J. R. , Draper, B. A. , & Lui, Y.

M. . (2010). Visual object tracking using adaptive

correlation filters. The Twenty-Third IEEE Conference

on Computer Vision and Pattern Recognition, CVPR

2010, San Francisco, CA, USA, 13-18 June 2010. IEEE.

Henriques, J. F. , Caseiro, R. , Martins, P. , & Batista, J. .

(2012). Exploiting the circulant structure of tracking-

by-detection with kernels.

Henriques, J. F. , Caseiro, R. , Martins, P. , & Batista, J. .

(2015). High-speed tracking with kernelized correlation

filters. IEEE Transactions on Pattern Analysis &

Machine Intelligence, 37(3), 583-596.

Danelljan, M. , Gustav Häger, Khan, F. S. , & Felsberg, M. .

(2014). Accurate Scale Estimation for Robust Visual

Tracking. British Machine Vision Conference.

Danelljan, M. , Khan, F. S. , Felsberg, M. , & Weijer, J. V.

D. . (2014). Adaptive Color Attributes for Real-Time

Visual Tracking. IEEE Conference on Computer Vision

& Pattern Recognition. IEEE.

Cai, Z. , Wen, L. , Lei, Z. , Vasconcelos, N. , & Li, S. Z. .

(2014). Robust deformable and occluded object

tracking with dynamic graph. IEEE Transactions on

Image Processing, 23(12).

Li, Y. , & Zhu, J. . (2014). A scale adaptive kernel

correlation filter tracker with feature integration.

Bertinetto, L. , Valmadre, J. , Golodetz, S. , Miksik, O. , &

Torr, P. H. S. . (2016). Staple: Complementary Learners

for Real-Time Tracking. Computer Vision & Pattern

Recognition. IEEE.

Possegger, H. , Mauthner, T. , & Bischof, H. . (2015). In

defense of color-based model-free tracking. Computer

Vision & Pattern Recognition. IEEE.

Danelljan, M. , Hger, G. , Khan, F. S. , & Felsberg, M. .

(2016). Learning spatially regularized correlation filters

for visual tracking.

Wang, M. , Liu, Y. , & Huang, Z. . (2017). Large margin

object tracking with circulant feature maps.

Wu, Y. , Lim, J. , & Yang, M. H. . (2015). Object tracking

benchmark. IEEE Transactions on Pattern Analysis &

Machine Intelligence, 37(9), 1834-1848.

Babenko, B. , Yang, M. H. , & Belongie, S. . (2011). Robust

object tracking with online multiple instance learning.

IEEE Transactions on Pattern Analysis and Machine

Intelligence, 33(8), 1619-1632.

Danelljan, M. , Hager, G. , Khan, F. S. , & Felsberg, M. .

(2016). Discriminative scale space tracking. IEEE

Transactions on Pattern Analysis and Machine

Intelligence, PP(99), 1-1.

Object Tracking using Correction Filter Method with Adaptive Feature Selection

487