9th International Conference on Computer and Knowledge Engineering (ICCKE 2019), October 24-25 2019, Ferdowsi University of Mashhad

Object Tracking Methods:A Review

Zahra Soleimanitaleb Mohammad Ali Keyvanrad Ali Jafari
M.S student Assistant Professor Assistant Professor
Malek-Ashtar University of Malek-Ashtar University of Malek-Ashtar University of
Technology Technology Technology
[email protected] [email protected] [email protected]

Abstract— Object tracking is one of the most important tasks in

computer vision that has many practical applications such as traffic TABLE I
monitoring, robotics, autonomous vehicle tracking, and so on. SOME OF THE OBJECT TRACKING CHALLENGES[12]
Different researches have been done in recent years, but because of
different challenges such as occlusion, illumination variations, fast challenge describe example
motion, etc. researches in this area continues. In this paper, various
methods of tracking objects are examined and a comprehensive the illumination
classification is presented that classified tracking methods into four in the target
Illumination
main categories of feature-based, segmentation-based, estimation- region is
Variation
based, and learning-based methods that each of which has its own significantly
sub-categories. The main focus of this paper is on learning-based changed
methods, which are classified into three categories of generative the background
methods, discriminative methods, and reinforcement learning. One near the target
of the sub-categories of the discriminative model is deep learning. Background
has a similar
Because of high-performance, deep learning has recently been very Clutters
color or texture
much considered. as the target

Keywords—Object tracking, Generative learning, the number of

Discriminative Learning, Reinforcement Learning, Deep Learning pixels inside the
Low
ground-truth
I. INTRODUCTION Resolution
bounding box is
Object tracking is one of the important tasks in computer low
vision that tries to detect and track objects in image sequences. In the ratio of the
object tracking, the target specifies in the first frame and must be bounding boxes
detected and tracked in the next frames of the video. Scale of the first frame
Object tracking has different applications. Object tracking Variation and the current
applicable in areas such as traffic monitoring (e.g. monitoring of frame is out of
traffic flow [1] and detection of traffic accidents [2]), robotics the range
(e.g. ASIMO humanoid robot [3]), autonomous vehicle tracking
(e.g. path-tracking [4]–[6]), medical diagnosis systems (e.g. the target is
tracking of ventricular wall [7] and medical instruments control Occlusion partially or fully
[8]), and activity recognition (e.g. learning activity patterns [9] occluded
and human activity recognition [10]).
There are many challenges in object tracking that have led to
During the
ongoing research in this area. Some of these challenges in the movement, the
OTB1 dataset [11] are presented in TABLE I. Change the
target may be
There are two OTB models, OTB100 and OTB50, where 100 target
rotated,
and 50 represent the number of sequences available. The position
deformed, and so
sequences in these datasets manually tag with 9 attributes, which on.
represents the challenging aspects in visual tracking.
the motion of the
Fast Motion ground truth is
large

1
Object Tracking Benchmark

978-1-7281-5075-8/19/$31.00 ©2019 IEEE

282

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on May 17,2021 at 07:14:30 UTC from IEEE Xplore. Restrictions apply.
 The aspects of tracking have serious challenges and there's a comprehensive classification of object tracking methods is
some expectations from tracking system, so when designing provided and details of each one is given below.
each tracking algorithm, should try to address those expectations
as much as possible. Some of these features are listed below. A. Feature-Based Methods
This method is one of the simple ways of object tracking. To
Robustness: Robustness means that the tracking system can track objects, features, such as color, texture, optical flow, etc.,
track the target even in complicated conditions such as are extracted first. These extracted features must be unique so
background clutters, occlusion and illumination variation. that objects can be easily distinguishable in the feature space.
Adaptability: In addition to the environment changes, the Once the features are extracted, then the next step is to find the
target is to changes, such as the complex and sudden movement most similar object in the next frame, using those features, by
of the target. In order to solve this problem, the tracking system exploiting some similarity criterion.
must be able to detect and track the current apparent One of the problems with these methods is at the extraction
characteristics of the target. step because the unique, exact and reliable features of the object
Real-time processing of information: A system that deals should be extracted so that it can distinguish the target from
with image sequences should have high processing speeds. So other objects. Here are some of the features that are used for
there is a need to implement a high-performance algorithm. object tracking.
1) Color
Color feature can show appearance of the object. This
II. CLASSIFICATION OF THE OBJECT TRACKING METHOD feature can be used in different ways, one of the most common
methods to use this feature is color histogram.
Object tracking methods have different categories, for The color histogram shows the distribution of colors in an
example, Fiaz et al. have a comprehensive study of tracking image, in fact, shows the different colors and the number of
methods that categorize tracking methods into two groups of pixels of each color in the image. The disadvantage of color
methods based on the correlation filter and the noncorrelation histograms is that the representation just depends on color of
filter [13]. Li et al. have reviewed and compared deep learning- object and ignore shape and texture of object , so two different
based tracking methods [14], Verma has reviewed methods of objects may have same histogram.
object detecting and tracking and categorized tracking methods There are some papers that use color histogram for tracking
into five categories of feature-based, segmentation-based, such as[16]–[18].
estimation-based, appearance-based and learning-based
methods [15].
In this paper, the last grouping is used and as shown in Figure 1,

Figure 1: Classification of object tracking methods

283

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on May 17,2021 at 07:14:30 UTC from IEEE Xplore. Restrictions apply.
 2) Texture 2) Joint Based Method
Texture is a repeated pattern of information or arrangement In bottom-up method, foreground segmentation and
of the structure with regular intervals. Texture features are not tracking are two separate tasks; one of the problems with this
obtained directly. These features are generated by image method is that the segmentation error propagates forward,
preprocessing techniques. causing error tracking. To solve this problem, as shown in
The texture feature is an important feature in the image, it Figure 3, the researchers merged the foreground segmentation
can be used along with the color feature to describe the contents and tracking method, witch improved tracking performance.
of an image or a region of the image. Because the color feature There are some papers that use joint based method for
is not sufficient to identify the similar objects and sometimes it tracking such as [32]–[34].
can be seen that different images have the same histogram.
Gabor wavelet [19] is one of the most studied texture
features. The most important property of Gabor filters is their
invariance to illumination, rotation, scale, and translation which
makes it suitable for object tracking. A method to detect the
location and body shapes of moving animals using a Gabor
filter is presented in [20]. The local binary pattern is another Figure 3: Joint based framework [26]
textual descriptor [21]. Zhao et al. have used LBP to describe C. Estimation Based Methods
moving objects and also have used a Kalman filter for target Estimation methods formulate the tracking problem to an
tracking [22]. estimation problem in which an object is represented by a state
vector. The state vector describes the dynamic behavior of a
3) Optical Flow system, such as the position and velocity of an object. The
Optical flow is the apparent motion of brightness patterns general framework for the dynamic mode estimation problem
in the image. apparent motion can be caused by lighting is taken from Bayesian methods [15].
changes without any actual motion. The optical flow algorithm The Bayesian filters allow the target to continuously update
calculates the displacement of brightness patterns from one its position on the coordinate system, based on the latest sensor
frame to another. Algorithms that calculate displacement for all data. This algorithm is recursive and consists of two steps:
image pixels are known as dense optical streaming algorithms, prediction and updating.
while light-flow algorithms estimate displacement tension for a The prediction step estimates the new position of the target
selective number of pixels in an image [23]. in the next step using the state model, while the updating step
There are some papers that use optical flow for tracking uses the current observation to update the target position using
such as[24], [25]. the observation model. The prediction and updating steps are
performed on each frame of the video. Here are some examples
B. Segmentation Based Methods of this method.
Segmenting foreground objects from a video frame is 1) Kalman Filter
fundamental and the most critical step in visual tracking. To use the Kalman filter [35] in object tracking, should
Foreground Segmentation is done to separate foreground design a dynamic model of the target movement. The Kalman
objects from the background scene. Normally, the foreground filter is used to estimate the position of a linear system assumed
objects are the objects that are moving in a scene. To track these that the errors are Gaussian. In many cases, dynamic models
objects, they are to be separated from the background scene are nonlinear, so in this case, the Kalman filter is not used and
[15]. In the following, some of the object tracking methods other suitable algorithms are used. One of these algorithms is
based on the segmentation are examined. the extended Kalman Filter [36]. The framework for using the
1) Bottom-Up Based Method Kalman filter is shown in Figure 4.
In this type of tracking, there must be two separate tasks, There are some papers that use kalman filter for tracking
first the foreground segmentation and then the object tracking. such as[22], [37].
The foreground segmentation uses a low-level segmentation
to extract regions in all frames, and then some features are
extracted from the foreground regions and tracked according to
those features[26], as shown in Figure 2.
There are some papers that use a bottom-up based method
for tracking such as [27]–[31].

Figure 4: Kalman filter framework [38]

Figure 2: Bottom-Up based framework [26]

284

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on May 17,2021 at 07:14:30 UTC from IEEE Xplore. Restrictions apply.
 2) Particle Filter affected it and increased its accuracy is object tracking. Due to
Most tracking issues are non-linear. Therefore, particle filter this increasing accuracy, we will focus on deep learning
has been considered for solving such problems. The particle methods which is explained below. In this paper, deep learning
filter is a recursive Monte Carlo statistical calculation method methods are classified into feature extraction based and end-to-
that is often used for non-gaussian noise measurement models. end methods.
The main idea of the particle filter shows the distribution of a
set of particles. Each particle has a probability weight, which i. FEATURE EXTRACTION BASED METHODS
represents the probability of sampling that particle from the Wang et al. [48] show that feature extraction is also a very
probability density function. One of the problems with this important issue in the design of a robust tracker. From the
method is that the particles that have more probability to be research experience of classical tracking algorithms, it can be
selected several times, and resampling is used to solve this concluded that any development in the methods of the feature
problem. The framework for using particle filters is shown in extracting or machine learning techniques may lead to the
Figure 5. development of tracking. Therefore, since deep learning
Particle filters are also used in [39]–[41]. techniques have shown great abilities in feature extraction and
object classification, it can be concluded that the use of deep
learning can improve tracking performance. Due to the success
of the deep features in image classification, some of the deep-
network tracking methods are used to extract features that are
known as a feature extraction network.
These methods separate the detection and tracking sections.
Detection uses deep learning methods that can extract deep
features, but for tracking, these methods can be classified into
two categories of tracking with classical methods and tracking
with Deep methods.
Figure 5: Particle filter framework [42] Yang et al have used Faster RCNN, KCF2 and Kalman filter
trackers [49]. Chahyati et al. have used Faster RCNN and the
D. Learning-Based Methods Siamese network [50]. Agarwal and Suryavanshi have used
In learning based methods, the features and appearance of Faster RCNN and GOTURN [51]. Ghaemmaghami has used
different targets and their prediction are learned in next frames, SSD, YOLO, and LSTM, and an example of the architecture of
and then in the test time, based on what they learned, they can this method is shown in Figure 6.
detect the object and track it in next frames. Learning based
methods are often divided into three types of generative,
discriminative and reinforcements learning.
1) Discriminative Methods
Discriminative trackers usually consider tracking as a
classification problem that discriminates the target from the
background. Discriminative learning is divided into two
Figure 6: Feature extraction based tracking example [52]
categories contains Shallow learning and Deep learning.
a) Shallow Learning ii. END-TO-END METHODS
Object tracking can be considered as a decision-making End-to-end methods train a network to conduct both feature
process where classification is trained to learn the extraction and candidate evaluation. In this paper, end-to-end
discrimination between the target and the background. After the methods are classified into three categories of Siamese trackers,
training, at the test time, it decides that the object is target or no. patch learning and graph-based trackers.
Features are extracted from different objects and the various
methods such as support vector machines [43]–[46] can be ™ Siamese tracker
used for classification.
Siamese networks have two inputs and produce one output,
b) Deep Learning it captures two inputs and measures the similarity between the
Shallow learning with fewer layers predicts the model, but two images to determine whether or not the same objects exist
deep learning has too many layers. Another difference is that in the two images.
shallow learning requires important and discriminatory features These types of networks are capable of learning similarities
extracted by professionals, but deep learning itself extracts and common features. There are some papers that use the
these important features. Deep learning [47] has made siamese tracker method for tracking such as [53]–[56]. An
impressive developments in various areas, including computer example of tracker architecture is shown in Figure 7.
vision, in recent years. One of the areas where deep learning has

2
Kernelized Correlation Filter

285

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on May 17,2021 at 07:14:30 UTC from IEEE Xplore. Restrictions apply.
 The main idea of the correlation filter is to estimate an
optimal image filter so that this filter produces optimal output
on the input image.
In the first frame, the target is identified by a bounding box
and the correlation filter is trained on it, then at each time step,
the patch is cropping to its predicted position for tracking. Then,
as shown in Figure 10, different features can be extracted from
the input data. A cosine window is usually used to smooth
boundry effects.
Figure 7: Siamese tracker example [53] Then, the correlation between the current input and the
frequency filter learned in the frequency domain is made based
on the Convolution theorem. After the correlation, a spatial
™ Patch learning tracker confidence map is obtained by Inverse FFT (IFFT), The
In patch learning method positive and negative samples are maximum value of that can be predicted as the new position of
extracted, then train the model on these samples. The model is target. Then, appearance at the newly estimated position is
tested on the number of samples, and the maximum response extracted for training and updating the correlation filter with a
indicates the target position. There are some papers that use output.
patch learning method for tracking such as [57]–[59]. An
example of this method shown in Figure 8. There are some papers that use the correlation filter method
for tracking such as [64]–[66].

Figure 8: Learning patch tracker example [60]

Figure 10: Correlation filter tracker framework [67]

™ Graph-based tracker
In computer vision, graph theory has been successfully 3) Reinforcement Learning
applied to solve many tasks. Graphs offer a rich and compact In a reinforcement learning[68] problem, we encounter a
representation of the spatial relations between objects in an agent that interacts with the environment through trial and error
image and they can be used to model many types of relations and learns to select the optimal action to achieve the goal.
and processes. There are some papers that use a reinforcement learning
There are some papers that use graph-based methods for method for tracking such as [69]–[71]. An example of this
tracking such as [61]–[63]. An example of this method shown method shown in Figure 11.
in Figure 9.

Figure 9: Graph-based tracker example[61]

Figure 11: Reinforcement learning tracker example [72]

2) Generative Methods
These methods focus on searching in areas that are more
similar to object. Correlation Filter based methods are one of
the examples of these trackers.

286

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on May 17,2021 at 07:14:30 UTC from IEEE Xplore. Restrictions apply.
 III. CONCLUSION [13] M. Fiaz, A. Mahmood, and S. K. Jung, “Tracking Noisy Targets: A
Review of Recent Object Tracking Approaches,”
arXiv:1802.03098 [cs], Feb. 2018.
In this paper, a comprehensive classification of object [14] P. Li, D. Wang, L. Wang, and H. Lu, “Deep visual tracking: Review
tracking algorithms is presented. In this category, tracking and experimental comparison,” Pattern Recognition, vol. 76, pp.
algorithms are divided into feature-based, segmentation-based, 323–338, 2018.
[15] R. Verma, “A Review of Object Detection and Tracking Methods,”
estimation-based and learning-based categories. This paper
International Journal of Advance Engineering and Research
focuses on learning-based approaches. Learning-based tracking Development, vol. 4, pp. 569–578, Oct. 2017.
algorithms, and especially deep learning, have recently received [16] X. Li and N. Zheng, “Adaptive target color model updating for
much attention. Deep learning is a new and exciting field in visual tracking using particle filter,” in 2004 IEEE International
various fields, especially computer vision, and in many fields it Conference on Systems, Man and Cybernetics (IEEE Cat. No.
is very accurate and has made a lot of progress. 04CH37583), 2004, vol. 4, pp. 3105–3109.
[17] M. Fotouhi, A. R. Gholami, and S. Kasaei, “Particle filter-based
object tracking using adaptive histogram,” in 2011 7th Iranian
Deep learning networks have many layers and extract Conference on Machine Vision and Image Processing, 2011, pp. 1–
different features of the object in each layer so it can have good 5.
accuracy in object tracking as well because it has many layers [18] A. Zhao, “Robust histogram-based object tracking in image
and model complexity is slower than shallow networks. sequences,” in 9th Biennial Conference of the Australian Pattern
However, it can’t be said that deep learning works best in all Recognition Society on Digital Image Computing Techniques and
cases and always should be used, but by knowing the Applications (DICTA 2007), 2007, pp. 45–52.
[19] D. Barina, “Gabor Wavelets in Image Processing,”
advantages and disadvantages of all methods one can find out arXiv:1602.03308 [cs], Feb. 2016.
which method can work best in problem solving. [20] D. A. Wagenaar and W. B. Kristan, “Automated video analysis of
animal movements using Gabor orientation filters,”
IV. REFERENCES Neuroinformatics, vol. 8, no. 1, pp. 33–42, 2010.
[21] T. Ojala, M. Pietikainen, and D. Harwood, “Performance
evaluation of texture measures with classification based on
[1] B. Tian, Q. Yao, Y. Gu, K. Wang, and Y. Li, “Video processing Kullback discrimination of distributions,” in Proceedings of 12th
techniques for traffic flow monitoring: A survey,” in 2011 14th International Conference on Pattern Recognition, 1994, vol. 1, pp.
International IEEE Conference on Intelligent Transportation 582–585.
Systems (ITSC), 2011, pp. 1103–1108. [22] Z. Zhao, S. Yu, X. Wu, C. Wang, and Y. Xu, “A multi-target
[2] J.-C. Tai, S.-T. Tseng, C.-P. Lin, and K.-T. Song, “Real-time image tracking algorithm using texture for real-time surveillance,” in
tracking for automatic traffic monitoring and enforcement 2008 IEEE International Conference on Robotics and Biomimetics,
applications,” Image and Vision Computing, vol. 22, no. 6, pp. 485– 2009, pp. 2150–2155.
501, 2004. [23] J. L. Barron, D. J. Fleet, S. S. Beauchemin, and T. A. Burkitt,
[3] Y. Sakagami, R. Watanabe, C. Aoyama, S. Matsunaga, N. Higaki, “Performance of optical flow techniques,” in Proceedings 1992
and K. Fujimura, “The intelligent ASIMO: System overview and IEEE Computer Society Conference on Computer Vision and
integration,” in IEEE/RSJ international conference on intelligent Pattern Recognition, 1992, pp. 236–242.
robots and systems, 2002, vol. 3, pp. 2478–2483. [24] J. Hariyono, V.-D. Hoang, and K.-H. Jo, “Moving object
[4] M. Brown, J. Funke, S. Erlien, and J. C. Gerdes, “Safe driving localization using optical flow for pedestrian detection from a
envelopes for path tracking in autonomous vehicles,” Control moving vehicle,” The Scientific World Journal, vol. 2014, 2014.
Engineering Practice, vol. 61, pp. 307–316, 2017. [25] D.-S. Kim and J. Kwon, “Moving object detection on a vehicle
[5] M. Menze and A. Geiger, “Object scene flow for autonomous mounted back-up camera,” Sensors, vol. 16, no. 1, p. 23, 2016.
vehicles,” in Proceedings of the IEEE Conference on Computer [26] R. Yao, G. Lin, S. Xia, J. Zhao, and Y. Zhou, “Video Object
Vision and Pattern Recognition, 2015, pp. 3061–3070. Segmentation and Tracking: A Survey,” arXiv preprint
[6] V. A. Laurense, J. Y. Goh, and J. C. Gerdes, “Path-tracking for arXiv:1904.09172, 2019.
autonomous vehicles at the limit of friction,” in 2017 American [27] J. Son, I. Jung, K. Park, and B. Han, “Tracking-by-segmentation
Control Conference (ACC), 2017, pp. 5586–5591. with online gradient boosting decision tree,” in Proceedings of the
[7] A. C. F. Colchester and D. J. Hawkes, Eds., Information Processing IEEE International Conference on Computer Vision, 2015, pp.
in Medical Imaging: 12th International Conference, IPMI ’91, 3056–3064.
Wye, UK, July 7-12, 1991. Proceedings. Berlin Heidelberg: [28] A. Adam, E. Rivlin, and I. Shimshoni, “Robust fragments-based
Springer-Verlag, 1991. tracking using the integral histogram,” in 2006 IEEE Computer
[8] S. Walker et al., “Systems and methods for localizing, tracking Society Conference on Computer Vision and Pattern Recognition
and/or controlling medical instruments,” Apr-2017. (CVPR’06), 2006, vol. 1, pp. 798–805.
[9] C. Stauffer and W. E. L. Grimson, “Learning patterns of activity [29] T. Zhou and D. Tao, “Shifted subspaces tracking on sparse outlier
using real-time tracking,” IEEE Transactions on Pattern Analysis for motion segmentation,” in Twenty-Third International Joint
and Machine Intelligence, vol. 22, no. 8, pp. 747–757, Aug. 2000. Conference on Artificial Intelligence, 2013.
[10] R. Bodor, B. Jackson, and N. Papanikolopoulos, “Vision-based [30] Z. Cai, L. Wen, Z. Lei, N. Vasconcelos, and S. Z. Li, “Robust
human tracking and activity recognition,” in Proc. of the 11th deformable and occluded object tracking with dynamic graph,”
Mediterranean Conf. on Control and Automation, 2003, vol. 1. IEEE Transactions on Image Processing, vol. 23, no. 12, pp. 5497–
[11] Y. Wu, J. Lim, and M.-H. Yang, “Online object tracking: A 5509, 2014.
benchmark,” in Proceedings of the IEEE conference on computer [31] C. Gu and M.-C. Lee, “Semiautomatic segmentation and tracking
vision and pattern recognition, 2013, pp. 2411–2418. of semantic video objects,” IEEE Transactions on Circuits and
[12] “Visual Tracker Benchmark.” [Online]. Available: Systems for Video Technology, vol. 8, no. 5, pp. 572–584, 1998.
http://cvlab.hanyang.ac.kr/tracker_benchmark/datasets.html. [32] C. Bibby and I. Reid, “Robust real-time visual tracking using pixel-
[Accessed: 11-Jun-2019]. wise posteriors,” in European Conference on Computer Vision,
2008, pp. 831–844.

287

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on May 17,2021 at 07:14:30 UTC from IEEE Xplore. Restrictions apply.
[33] A. Milan, L. Leal-Taixé, K. Schindler, and I. Reid, “Joint tracking [53] D. Held, S. Thrun, and S. Savarese, “Learning to track at 100 fps
and segmentation of multiple targets,” in Proceedings of the IEEE with deep regression networks,” in European Conference on
Conference on Computer Vision and Pattern Recognition, 2015, Computer Vision, 2016, pp. 749–765.
pp. 5397–5406. [54] R. Tao, E. Gavves, and A. W. Smeulders, “Siamese instance search
[34] F. Schubert, D. Casaburo, D. Dickmanns, and V. Belagiannis, for tracking,” in Proceedings of the IEEE conference on computer
“Revisiting robust visual tracking using pixel-wise posteriors,” in vision and pattern recognition, 2016, pp. 1420–1429.
International Conference on Computer Vision Systems, 2015, pp. [55] X. Wang, C. Li, B. Luo, and J. Tang, “Sint++: robust visual
275–288. tracking via adversarial positive instance generation,” in
[35] N. Najafzadeh, M. Fotouhi, and S. Kasaei, “Object tracking using Proceedings of the IEEE Conference on Computer Vision and
Kalman filter with adaptive sampled histogram,” in 2015 23rd Pattern Recognition, 2018, pp. 4864–4873.
Iranian Conference on Electrical Engineering, 2015, pp. 781–786. [56] K. Chen and W. Tao, “Once for all: a two-flow convolutional
[36] “A Constrained Extended Kalman Filter for Target Tracking - neural network for visual tracking,” IEEE Transactions on Circuits
Google Search.” [Online]. Available: and Systems for Video Technology, vol. 28, no. 12, pp. 3377–3386,
https://www.google.com/search?q=A+Constrained+Extended+Ka 2017.
lman+Filter+for+Target+Tracking&oq=A+Constrained+Extended [57] H. Nam and B. Han, “Learning multi-domain convolutional neural
+Kalman+Filter+for+Target+Tracking&aqs=chrome..69i57.380j0 networks for visual tracking,” in Proceedings of the IEEE
j7&sourceid=chrome&ie=UTF-8. [Accessed: 13-Jun-2019]. Conference on Computer Vision and Pattern Recognition, 2016,
[37] P. R. Gunjal, B. R. Gunjal, H. A. Shinde, S. M. Vanam, and S. S. pp. 4293–4302.
Aher, “Moving Object Tracking Using Kalman Filter,” in 2018 [58] I. Jung, J. Son, M. Baek, and B. Han, “Real-time mdnet,” in
International Conference On Advances in Communication and Proceedings of the European Conference on Computer Vision
Computing Technology (ICACCT), 2018, pp. 544–547. (ECCV), 2018, pp. 83–98.
[38] “Object Tracking: Kalman Filter with Ease.” [Online]. Available: [59] B. Han, J. Sim, and H. Adam, “Branchout: Regularization for
/Articles/865935/Object-Tracking-Kalman-Filter-with-Ease. online ensemble tracking with convolutional neural networks,” in
[Accessed: 14-Jun-2019]. Proceedings of the IEEE Conference on Computer Vision and
[39] L. Mihaylova, P. Brasnett, N. Canagarajah, and D. Bull, “Object Pattern Recognition, 2017, pp. 3356–3365.
tracking by particle filtering techniques in video sequences,” [60] “Learning Multi-Domain Convolutional Neural Networks for
Advances and challenges in multisensor data and information Visual Tracking arXiv : [cs.CV] v1 2015, v Hyeonseob Nam,
processing, vol. 8, pp. 260–268, 2007. Bohyung Han Dept. - ppt video online download.” [Online].
[40] P. Pérez, C. Hue, J. Vermaak, and M. Gangnet, “Color-based Available: https://slideplayer.com/slide/11571058/. [Accessed: 19-
probabilistic tracking,” in European Conference on Computer Jun-2019].
Vision, 2002, pp. 661–675. [61] H. Nam, M. Baek, and B. Han, “Modeling and propagating cnns in
[41] Y. Li, H. Ai, T. Yamashita, S. Lao, and M. Kawade, “Tracking in a tree structure for visual tracking,” arXiv preprint
low frame rate video: A cascade particle filter with discriminative arXiv:1608.07242, 2016.
observers of different life spans,” IEEE Transactions on Pattern [62] T. Wang and H. Ling, “Gracker: A graph-based planar object
Analysis and Machine Intelligence, vol. 30, no. 10, pp. 1728–1740, tracker,” IEEE transactions on pattern analysis and machine
2008. intelligence, vol. 40, no. 6, pp. 1494–1501, 2017.
[42] “Object Tracking: Particle Filter with Ease.” [Online]. Available: [63] C. Gomila and F. Meyer, “Graph-based object tracking,” in
/Articles/865934/Object-Tracking-Particle-Filter-with-Ease. Proceedings 2003 International Conference on Image Processing
[Accessed: 14-Jun-2019]. (Cat. No.03CH37429), Barcelona, Spain, 2003, vol. 3, pp. II-41–4.
[43] S. Avidan, “Support vector tracking,” IEEE transactions on pattern [64] M. Danelljan, A. Robinson, F. S. Khan, and M. Felsberg, “Beyond
analysis and machine intelligence, vol. 26, no. 8, pp. 1064–1072, correlation filters: Learning continuous convolution operators for
2004. visual tracking,” in European Conference on Computer Vision,
[44] M. Tian, W. Zhang, and F. Liu, “On-line ensemble SVM for robust 2016, pp. 472–488.
object tracking,” in Asian conference on computer vision, 2007, pp. [65] M. Danelljan, G. Hager, F. Shahbaz Khan, and M. Felsberg,
355–364. “Learning spatially regularized correlation filters for visual
[45] T. Malisiewicz, A. Gupta, and A. A. Efros, “Ensemble of exemplar- tracking,” in Proceedings of the IEEE International Conference on
SVMs for object detection and beyond.,” in Iccv, 2011, vol. 1, p. 6. Computer Vision, 2015, pp. 4310–4318.
[46] S. Zhang, X. Yu, Y. Sui, S. Zhao, and L. Zhang, “Object tracking [66] M. Danelljan, G. Bhat, F. S. Khan, and M. Felsberg, “ECO:
with multi-view support vector machines,” IEEE Transactions on Efficient Convolution Operators for Tracking.,” in CVPR, 2017,
Multimedia, vol. 17, no. 3, pp. 265–278, 2015. vol. 1, p. 3.
[47] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. [67] Z. Chen, Z. Hong, and D. Tao, “An Experimental Survey on
521, no. 7553, pp. 436–444, May 2015. Correlation Filter-based Tracking,” Sep. 2015.
[48] N. Wang, J. Shi, D.-Y. Yeung, and J. Jia, “Understanding and [68] M. E. Harmon and S. S. Harmon, “Reinforcement Learning: A
diagnosing visual tracking systems,” in Proceedings of the IEEE Tutorial.,” WRIGHT LAB WRIGHT-PATTERSON AFB OH,
International Conference on Computer Vision, 2015, pp. 3101– 1997.
3109. [69] J. Choi, J. Kwon, and K. M. Lee, “Visual tracking by reinforced
[49] F. Bu, Y. Cai, and Y. Yang, “Multiple Object Tracking Based on decision making,” arXiv preprint arXiv:1702.06291, 2017.
Faster-RCNN Detector and KCF Tracker.” [70] W. Luo, P. Sun, F. Zhong, W. Liu, T. Zhang, and Y. Wang, “End-
[50] D. Chahyati, M. I. Fanany, and A. M. Arymurthy, “Tracking People to-end active object tracking via reinforcement learning,” arXiv
by Detection Using CNN Features,” Procedia Computer Science, preprint arXiv:1705.10561, 2017.
vol. 124, pp. 167–172, 2017. [71] S. Yun, J. Choi, Y. Yoo, K. Yun, and J. Young Choi, “Action-
[51] A. Agarwal and S. Suryavanshi, “Real-Time* Multiple Object decision networks for visual tracking with deep reinforcement
Tracking (MOT) for Autonomous Navigation,” Technical report, learning,” in Proceedings of the IEEE conference on computer
2017. vision and pattern recognition, 2017, pp. 2711–2720.
[52] M. P. Ghaemmaghami, “Tracking of Humans in Video Stream [72] D. Zhang, H. Maei, X. Wang, and Y.-F. Wang, “Deep
Using LSTM Recurrent Neural Net- work,” p. 50. reinforcement learning for visual object tracking in videos,” arXiv
preprint arXiv:1701.08936, 2017.

288

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on May 17,2021 at 07:14:30 UTC from IEEE Xplore. Restrictions apply.