In this paper, we describe an adaptive Markov chain Monte Carlobased particle filter that effectively addresses realtime multiface tracking on mobile platforms. Because traditional approaches based on a particle filter require an enormous number of particles, the processing time is high. This is a serious issue, especially on low performance devices such as mobile phones. To resolve this problem, we developed a tracker that includes a more sophisticated likelihood model to reduce the number of particles and maintain the identity of the tracked faces. In our proposed tracker, the number of particles is adjusted during the sampling process using an adaptive sampling scheme. The adaptive sampling scheme is designed based on the average acceptance ratio of sampled particles of each face. Moreover, a likelihood model based on color information is combined with corner features to improve the accuracy of the sample measurement. The proposed tracker applied on various videos confirmed a significant decrease in processing time compared to traditional approaches.
1. INTRODUCTION
The realtime object detection or tracking is one of the fundamental steps for a number of advanced systems in computer vision such as humancomputer interaction, augmented reality and videosurveillance. Moreover, human faces play an important role in human communication. Thus, face detection and tracking has been a research interest of many researchers on computer vision. Over the past years, numerous methods have been proposed on face detection and tracking.
The most basic approach to solve face tracking problem is to employ face detection
[1]
on every frame. However, despite much progress performed in recent years on multiface detection, there are indeed many situations where faces are not detected, which is especially due to the variations of face appearance, lighting conditions or partial or full occlusion of the face. Face detectors are normally applied on simple scenarios, where people predominantly look towards the camera. However, it is the less common head poses that people naturally take. Besides, processing time of face detectors is normally considerable. As a consequence, it reduces the number of frames recorded per second, thus lowering the quality of recorded videos. Therefore, in practice, robust face trackers are combined with face detectors not only to improve the detection results but also to reduce the processing time. Numerous methods for visual tracking of faces have been proposed in the literature. These methods can be classified into two classes, singleface tracking and multiface tracking.
The complexity of singleface video screen is less than that of multiface video screen. Thus, a lot of effective tracking methods have been proposed for singleface tracking. Yui Man Lui et al.
[2]
presented an adaptive framework for condensation algorithms in the context of humanface tracking. He addressed the face tracking problem by making factored sampling more efficient and appearance update more effective. Ruian Liu et al.
[3]
used adaboost for face detection and adaptive mean shift algorithm for face tracking. P. Jimenez et al.
[4]
proposed a method for robust tracking and estimating the face pose of a person using stereo vision. In this method, a face model is automatically initialized and constructed online: a fixed point distribution is superposed over the face when it is frontal to the cameras, and several appropriate points close to those locations are chosen for tracking. Vidit Saxena et al.
[5]
presented a realtime face tracking system using rank deficient face detection. Motion estimation and compensation are then incorporated in the system to ensure robust tracking, to minimize false detections, and for persistent tracking of the desired face. Liang Wang et al.
[6]
combined two sophisticated techniques of motion detection and template matching for detection and tracking of human faces. He used a statistical model of skin color and shape information to detect face in the first frame, and initialize it as an appearancebased intensity template for subsequent tracking. Derek Magee et al.
[7]
presented an efficient and general framework for the incorporation of statistical prior information, based on a wide variety of detectable point features, into level set based object tracking. The level set evolution is based on the interpolation of likelihood gradients using kernels centered at the features. Jun Wang et al.
[8]
proposed an improved camshiftbased particle filter algorithm for face tracking. He presented a novel feature extraction method called the block rotationinvariant uniform local binary pattern, and combine with color features to represent the appearance model of face in tracking tasks.
For multiface tracking, due to its complexity in video screen, particle filter
[9]

[12]
, also known as the sequential Monte Carlo
[13]
, becomes the most popular framework chosen by researchers. The basic concept of the particle filter is to use a set of weighted particles to approximate the true filtering distribution. Particle filters offer a degree of robustness to unpredictable motion and can correctly handle complicated, nonlinear measurement models. When tracking multifaces, simply running one individual particle filter for each face is not a viable option. Particle filter itself can not address the complex interactions between faces and leads to frequent tracker failures. Whenever faces pass close to one another, the face with the best likelihood score typically affects the filters of nearby faces.
This is illustrated in
Fig. 1
[23]
. To address this issue, Zia Khan et al.
[14]
proposed a method by incorporating a Markov random field (MRF) to model interactions and improve tracking when faces interact. Besides, he replaced the traditional importance sampling step in the particle filter, which suffers from exponential complexity in the number of tracked faces, with a Markov chain Monte Carlo (MCMC) sampling step. Recently, many other researchers have improved Khan’s method and proposed their improvement based on MCMCbased particle filter. I. Zuriarrain et al.
[15]
presented a MCMCbased particle filter to track multiple persons dedicated to video surveillance applications. He used saliency map proposal distribution to limit the wellknown burst in terms of particles and MCMC iterations. Stefan Duffner et al.
[16]
presented a multiface tracking algorithm that effectively deals with missing or uncertain detections in a principled way. The tracking is formulated in a multiobject statespace Bayesian filtering framework solved with MCMC. AnhTuyet Vu et al.
[17]
proposed a new multitarget tracking algorithm capable of tracking an unknown number of targets that move close and/or cross each other in a dense environment. Xiuzhuang Zhou et al.
[18]
proposed a samplingbased tracking scheme for the abrupt motion problem in the Bayesian filtering framework. Rather than simply adopting the sequential importance resampling or standard MCMC sampling algorithm, he proposed a more effective dynamic sampling scheme to sample from the filtering distribution by using the stochastic approximation Monte Carlo (SAMC) algorithm and present a sequential SAMC sampling algorithm for the tracking of abrupt motion, which demonstrates superiority in dealing with the localtrap problem with less computational burden. These presented methods run efficiently under its purpose, and the authors claim that they can run in realtime. However, the experiments of these methods were done in videos with low frame rate, (e.g. 1015 fps
[16]
, 25 fps
[18]
), and the testing environment is high performance computer, (e.g. 2.8 GHz
[18]
, 3.16 GHz
[16]
).
Problem of particle filter, (a) two faces tracked using independent particle filters, (b) the face with the best likelihood score typically affects the filter of the nearby face, (c) resulting tracker failure
As, nowadays, smart phones mounting high resolution cameras are widely used in the world, a fast face tracking approach running in a low performance device is highly demanded in human life. Thus, in this paper, we propose a novel method using MCMC and particle filter that effectively deals with realtime multiface tracking on mobile. Based on our observations, there are two factors that affect the processing time of particle filter, the number of sampled particles and the likelihood computation. These recent methods require a huge number of particles, for examples, 500 particles
[16]
and 300 particles
[18]
. To reduce the number of particles, we design an adaptive sampling scheme to track the acceptance ratio of sampled particles of each face. Since, the smaller the number of particles used, the lower the accuracy achieved, we need to develop a more sophisticated likelihood model to accurately measure the sampled particles. However, likelihood computation is one of two factors that affects the processing time of particle filter. Thus, we choose a color based histogram model
[12]
as the main model for likelihood measurement because of its simple computation. Since the raw data taken from mobile cameras is in YUV color space, we keep using this data to reduce the converting time to another color space. A weight mask is applied to each particle to increase the meaning of face center in likelihood computation. Moreover, a fast corner detector will be used to improve the likelihood score of each particle, depending on the proportion of detected corners inside each particle to the total number of detected corners.
The rest of the paper is organized as follows. Section 2 presents our adaptive MCMCbased particle filter (AMCMCPF) method. Implementation details and experimental results are presented in section 3. Conclusions can be found in section 4.
2. PROPOSED METHOD
Our primary goal in multiface tracking is to estimate the posterior distribution
P
(
X
^{'}
_{t}

Z
_{1:t}
) over the state
X
^{'}
_{t}
at the current time step t, given all observations
Z
_{1:t}
= {
Z
_{1}
,...,
Z
_{t}
} up to that time step, according to:
Here
c
is normalization constant, the likelihood
P
(
Z
_{t}

X
^{'}
_{t}
) expresses the
measurement model
or
likelihood model
, the probability we would have observed the measurement
Z
_{t}
given the state
X
^{'}
_{t}
at time
t
, and the
motion model P
(
X
^{'}
_{t}

X
^{'}
_{t}
_{1}
) predicts the state
X
^{'}
_{t}
given the previous state
X
^{'}
_{t}
_{1}
. These models are described more detail in the following sections.
 2.1 State Space
A state
X
^{'}
_{t}
of faces contains all the information to identify faces, including the position, scale and eccentricity (i.e. the ratio between height and width) of the face bounding box. In addition to this necessary information, due to interactions and occlusions in multiface tracking, faces may appear or disappear under observation. Thus the number and identity of the faces need to be estimated. To model this, a new variable, namely the set of identifiers
k_{t}
of faces currently in view
[19]
, is introduced. Suppose that
M
is the maximum number of faces visible at a current time step, we can define a state
X
^{'}
_{t}
as:
where
X
_{t}
= {
X
_{i,t}
}
_{i=1..M}
and
k
_{t}
= {
k
_{i,t}
}
_{i=1..M}
. Each
X
_{i,t}
contains the position, scale and eccentricity of face
i
at time
t
, and each
k
_{i,t}
denotes the status of face
i
at time
t
(Eq. 3).
 2.2Motion Model
The motion model describes the relationship between the current state
X
^{'}
_{t}
and the previous state
X
^{'}
_{t}
_{1}
. In overall motion model is defined as:
Here
π
_{0}
(
X
_{t}

k
_{t}
) is the interaction prior which prevents faces to become too close to each other.
π
_{0}
is defined as:
where the set Ω ={(
i
,
j
)(
k
_{i,t}
=1)∧(
k
_{j,t}
=1)∧(
i
≠
i
)} consists all possible pairs of faces that are visible, and the
ϕ
function describes the interaction between two visible faces. In our framework,
π
_{0}
is estimated as:
where
is the penalty function describing the proportion of the intersection area to the average area of two bounding boxes
B
_{i}
and
B
_{j}
defined by
X
_{i,t}
and
X
_{j,t}
, respectively.
λ_{g}
is a constant factor that controls the strength of the interaction prior. The motion of each face is describe more precisely as:
To describe the motion
P
(
X_{i,t}

X
_{i,t1}
) of each visible face, the first order autoregressive model is used to update the position parameter. Since the scale and eccentricity parameters are more stable than the position parameter, these parameters are only updated when a face detector tries to reinitialize the state space after a number of frames.
 2.3 Likelihood Model
 2.3.1 ColorBased Likelihood Model
As mentioned in section 1, the processing time of particle filter depends on the likelihood computation. Therefore, to strike a balance between robustness and computational complexity, we choose a simple but effective likelihood model based on color information for multiface tracking.
Assuming that the face observations
Z
_{i,t}
are conditionally independent given the state
X
_{i,t}
, we can define the likelihood model as the product of likelihoods of the visible faces:
To compute the likelihood of each face
P
(
Z
_{i,t}

X
_{i,t}
), we observe its color information. Because of its advantage with respect to human perception, HSV color space is the most popular color model
[12]
,
[16]
. However, the raw data taken from mobile cameras is in YUV color space, and in our experiments, the processing time to convert an image, 640x480 pixels, from YUV color space to HSV color space is approximately 40ms (see section 3.1 for detail environment setup). Furthermore, as similar to HSV color space, YUV color space also takes human perception into account. Thus, in our likelihood model, YUV color space is chosen to represent the color information of faces.
Fig. 2
shows a sample image in both RGB and YUV color spaces. In YUV color space, Y stands for the brightness component, and U and V are the chrominance components. Hence, we obtain color information with
N
=
N_{u}
×
N_{v}
bins using only the U and V channels.
N_{u}
and
N_{v}
are the number of bins of U and V channels, respectively.
A sample image in different color spaces, (a) RGB color space, (b) YUV color space
Given a state
X
_{i,t}
of face
i
at time
t
, the candidate region in which color information will be gathered is named as
R
_{i,t}
. Within this region a kernel density estimate
q
_{i,t}
of color distribution at time
t
is given by:
where
Here
k
is a normalization constant ensuring
w
is a weighting constant,
δ
is Kronecker delta function,
p
is a pixel located inside the region
R
_{i,t}
, and
is the bin index associated with the color at pixel location
p
. The most basic weighting function is
w
≡1, which means the kernel density is equal to standard color histogram. However, the meaning of the color pixels near region center is normally higher than the meaning of the color pixels far from region center.
For example, when the face involves a large head tilt as shown in
Fig. 3
a, its bounding box includes a lot of background pixels, which are meaningless. If we weigh the background pixels equally to face’s pixels, the likelihood measurement will be less accurate, and the bounding box will never fit the tracked face. There are several weighting functions to address this problem, such as Gaussian
[20]
and radius distance
[21]
. Based on these weighting functions, we adopt our weighting function by dividing the bounding box into 4 by 4 blocks (
Fig. 3
b) and weigh each block with a power of two.
A sample image with and without grid box, (a) with bounding box, (b) with grid box
The weight of each block is shown in
Table 1
. Weight values are assigned to be power of two for more efficient computation.
Weight values of each block
Weight values of each block
At time
t
, the observation
Z
_{i,t}
=
q
_{i,t}
is compared to the reference color distribution
to define the observation likelihood for a tracked face. An approximation of observation likelihood is given as:
where
D
denotes the Euclidean distance between two color distributions and
λ_{D}
is a constant factor that controls the strength of the observation likelihood. The reference color distribution is gathered at the initial time
t_{0}
and updated every time step. Let
denotes the color distribution of the mean state of tracked face
i
at time
t
 1. The reference color distribution of face
i
at time
t
is defined as:
where ε is the update factor that controls how fast the reference color distribution is updated.
 2.3.2 Corner Features
Likelihood model based on color information is fast. However, color information is not stable for occlusion. Thus, in this paper, we combine color information with corner information to increase the accuracy of likelihood measurement. In our observation, there are a lot of corner points concentrated on human faces.
For example, a face with detected corner points is shown in
Fig. 4
. Thus, a good particle contains a lot of corner points inside. To represent this characteristic, we update the likelihood of a particle with the ratio of the number of detected corners outside particle to the number of detected corners. For a better efficiency, we detect only the corner points in a limited region. Suppose that
B
_{i,t}
_{−1}
is the tracked bounding box of face
i
at time
t
 1. At time
t
, we detected only corner points for face
i
in the bounding box
B
_{i,t}
_{−1}
± Δ, where Δ is a small constant value describing the expansion and movement of
B
_{i,t}
_{−1}
. If
C
_{t}
is the number of detected corners, and
C
_{o}
is the number of detected corners outside particle, the observation likelihood in Eq. 12 can be reestimated as:
Detected corner points in a face
For fast corner detection, we employ the FAST (Features from Accelerated Segment Test) corner detector
[22]
.
 2.4 Tracking Model
At each time step, the tracking model proceeds in two main stages: estimating the states of the tracked faces and identifying the visibility status of these faces.
 2.4.1 NCNCbased Particle Filter
To estimate the states of the tracked faces, we use a MCMC sampling scheme, which allows efficient sampling in high dimensional state space of interacting faces
[14]
. Suppose that at time
t
– 1, the state of the tracked faces is represented by a set of samples
In which,
N
is the total number of particles and
N_{b}
is the number of “burnin” particles. Hence, the detailed steps of the MCMC sampling scheme are proposed as follows:
1) Initialize the MCMC sampler at time
t
with the sample
X_{t}
^{'(0)}
obtained by randomly selecting a particle from the set
and sampling state of every visible faces
i
in
X_{t}
^{'(0)}
using the motion model
P
(
X_{i,t}

X
_{i,t1}
)
2) Sample iteratively
N
particles from the posterior distribution (Eq. 1) using the MetropolisHasting (MH) algorithm. Discard the first
N_{b}
samples to account for sampler burnin. The detailed steps of MH are described as follows.

a) Sample a new particleX''tfrom the proposal distribution

b) Compute the acceptance ratio
c) If
a
≥1 then accept the particle, set
X_{t}
^{'(r+1)}
=
X^{''}_{t}
. Otherwise, add a copy of the current particle to the new sample set with probability
a
.
The particle set
at time t represents an estimation of the posterior of the tracked faces.
 2.4.2 Adaptive Sampling Scheme
The number of particles
N
is a factor that affects the processing time of particle filter. Thus, instead of using a fixed number of particles, we try to adjust it during sampling process.
Fig. 5
shows the adaptive sampling scheme of our tracker. In this scheme, we need to define three factors:
P_{min}
,
P_{max}
and
P_{step}
.
P_{min}
is the minimum number of particles that the sampler has to generate. In contrast with
P_{min}
,
P_{max}
is the maximum number of particles that the sampler can generate. Thus, we have
P_{min}
≤
N
≤
P_{max}
.
P_{step}
is the number of particles sampled more if the average acceptance ratio
γ
of sampled particles is less than threshold
λ_{a}
.
Adaptive Sampling Scheme
 2.4. 3 Visibility Status Identification
To identify the visibility status of tracked faces, we assume that if a face is still tracked correctly, the observation likelihood should be high and the variances in x and y direction of the bounding box should be low. Let
y_{i,t}
be the observation likelihood of the mean state of face
i
at time
t
and
v_{i,t}
be the maximum of the variances in x and y direction of the bounding box of the mean state of face
i
at time
t
. The average likelihood of face
i
over all time steps is computed as:
The average variance of bounding box of face
i
over all time steps is computed as:
With the computation in Eq. 17 and Eq. 18, we can skip the likelihood and variance of invisible faces. A tracked face is marked as invisible if:
In here,
λ_{y}
and
λ_{v}
are the constant factors that control the confident of the tracker. A face is invisible more than a number of frames will be removed from the tracker. In contrast, an invisible face at time
t
– 1 will be marked as a visible face at time t if:
3. EXPERIMENTS AND RESULTS
 3.1 Experiment Setup
To test the empirical performance of our proposed tracker, we construct a video database using a Samsung Galaxy S2 mobile phone. Our video database contains 90 videos recorded in various scenarios, including single face, multifaces, different illumination (i.e. bright, dark, normal) and different movement (i.e. head movement, camera movement). Each video is recorded with a resolution of 640x480 pixels, and frame rate of 30 fps. Our tracker is implemented in Android NDK environment and runs in Samsung Galaxy S2 with a dualcore 1.2 GHz processor.
Below are the specific implementation choices for constant factors presented in previous section.

o We use a uniform density centered on the previous pose.

o The range of the uniform distribution is (32, 32).

o The constant factor for the interaction prior:λg= 4.

o The numbers of bins of channel U and V:Nu=Nv= 16.

o The constant factor for likelihood computation:λD= 16.

o The update factor:ε= 0.5

o The constant value for the expansion and movement of bounding box: Δ = 8

o We discard 25% of the samples to let the sampler burn in, regardless of the total number of samples.

For the adaptive sampling scheme

o The average acceptance ratio threshold:λa= 0.25

o The minimum number of samples:Pmin= 64

o The maximum number of samples:Pmax=128

o The number of samples are increased each iteration:Pstep= 16

For visibility status identification

o The confident factor:λy= 0.25 andλv= 4

o A face is invisible more than 16 frames will be removed from the tracker.
Notice that these constant factors are assigned to be power of two for faster computation.
In our tracking progress, the face detector et al.
[1]
is used to initialize in the first frame, and reinitialize after every 64 frames. The reinitialization is necessary for a stable tracker. However, in some frames, the face detector fails to detect the faces. In that case, we will keep tracking undetected faces without reinitialization.
 3.2 Performance Measures
To measure performance of our proposed algorithm, we use Precision (P), Recall (R) and Fmeasure (
F_{β}
), which are defined as following equations.
Here,
n
is the number of annotated faces in a frame,
B_{g,i}
is the ground truth rectangle of face
i
and
B_{t,i}
is the rectangle output of face
i
from face detection or tracking. We use
β
^{2}
= 0.3 to weigh recall more than precision.
 3.3 Results
We compared our proposed algorithm to the following algorithms: [MCMCPF] A stateofart multitracking method based on MCMC and particle filter
[14]
. [PP1] Our proposed algorithm without applying grid box in likelihood computation. [PP2] Our proposed algorithm without using corner information. [PP3] Our proposed algorithm with a fixed 64 number of particles. [PP4] Our proposed algorithm with a fixed 128 number of particles. [PP5] Our proposed algorithm but using HSV color space instead of YUV color space.
Table 2
shows features of the above methods. And the accuracies of tracking algorithms are shown in
Table 3
.
Face tracking algorithms with different features
Face tracking algorithms with different features
Accuracies of tracking algorithms
Accuracies of tracking algorithms
The processing time of tracking algorithms is shown in
Table 4
.
Processing time of tracking algorithms
Processing time of tracking algorithms
The accuracy of MCMCPF is worse than that of other algorithms. This means that our extended likelihood model with corner information is more effective than the simple colorbased likelihood model. The accuracy of PP1 is worse than that of our proposed algorithm. This proves the effectiveness of grid box in likelihood computation. The accuracy of PP2 is far lower than that of our proposed algorithms. This means that the corner information is one of the most important feature in our proposed algorithms.
The accuracy of our proposed algorithm is more competent than that of PP3, but not as good as that of PP4. However, regarding processing time, our algorithm has proved its advantage. Because PP4 uses a large number of particles, its processing time is almost twice the processing time of our proposed algorithm or the processing time of PP3. Therefore, to strike a balance between tracking accuracy and processing time, our proposed algorithm is the most suitable for the realtime processing in low performance devices, such as smart phones. The accuracy of PP5 shows that the HSV color space has some advantage in comparing with YUV color space. Because of converting time between two color spaces, the processing time PP5 is much more than that of our proposed algorithm. Thus, HSV color space is not suitable for our proposed algorithm.
Fig. 6
shows the results of our face tracker in several different scenarios.
Results of the proposed face tracker with several different scenarios, 1st row: free moving style, 2nd row: rotation, 3rd row: scaling, 4th row: multiface scaling, 5th row: multiface rotation
4. CONCLUSIONS
We have presented an adaptive MCMCbased particle filter framework for robust realtime multiface tracking in various scenarios. In the proposed tracking algorithm, we have introduced an adaptive sampling scheme that concurrently reduces the number of particles and processing time. Furthermore, we have extended the likelihood model based on color information by combining with corner information. The extended likelihood model can effectively deal with occlusions and increase the tracking accuracy. Extensive experimentation has indicated that our method outspeeds other alternatives and can run in realtime in low performance devices, such as mobile phones. Our further study will concentrate on improving the tracking accuracy while sustaining the processing time.
Acknowledgements
“This work was supported by SAMSUNG ELECTRONICS CO., LTD.” And "This research was supported by the MSIP(Ministry of Science, ICT and Future Planning), Korea, under the ITRC(Information Technology Research Center) support program (NIPA2014H0301141014) supervised by the NIPA(National IT Industry Promotion Agency)."
BIO
In Seop Na
He received his B.S., M.S. and Ph.D. degree in Computer Science from Chonnam National University, Korea in 1997, 1999 and 2008, respectively. Since 2012, he has been a research professor in Department of Computer Science, Chonnam National University, Korea. His research interests are image processing, pattern recognition, character recognition and digital library.
Ha Le
He received the B.S in Computer Science from Hanoi University of Science and Technology, Vietnam in 2010. And he recieved the M.E in the Department of Computer Science, Chonnam National University, Korea. In 2013. His main research interests include pattern recognition, image processing, text recognition, object segmentation and object tracking.
Soo Hyung Kim
He received his B.S. degree in Computer Engineering from Seoul National University in 1986, and his M.S. and Ph.D degrees in Computer Science from Korea Advanced Institute of Science and Technology in 1988 and 1993, respectively. From 1990 to 1996, he was a senior member of research staff in Multimedia Research Center of Samsung Electronics Co., Korea. Since 1997, he has been a professor in the Department of Computer Science, Chonnam National University, Korea. His research interests are pattern recognition, document image processing, medical image processing, and ubiquitous computing.
Viola P.
,
Jones M.
2001
“Rapid object detection using a boosted cascade of simple features,” vol. 1
In Proc. of CVPR
511 
518
Man Lui Yui
,
Ross Beveridge J.
,
Darrell Whitley L.
2010
“Adaptive appearance model and condensation algorithm for robust face tracking,”
Trans. Sys. Man Cyber
Part A 40
(3)
437 
448
DOI : 10.1109/TSMCA.2010.2041655
Liu Ruian
,
Zhang Mimi
,
Ma Shengtao
2010
“Design of Face Detection and Tracking System,”
Image and Signal Processing (CISP)
4
1840 
1844
Jime´nez P.
,
Nuevo J.
,
Bergasa L. M.
,
SoteloFace M. A.
2009
“Tracking and pose estimation with automatic threedimensional model construction,”
Computer Vision, IET
3
93 
102
DOI : 10.1049/ietcvi.2008.0057
Saxena Vidit
,
Grover Sarthak
,
Joshi Sachin
2008
“A Real Time Face Tracking System using Rank Deficient Face Detection and Motion Estimation,”
Cybernetic Intelligent Systems, CIS
1 
6
Wang Liang
,
Tan Tieniu
,
Hu Weiming
“Face Tracking Using MotionGuided Dynamic Template Matching,”
Asian Conference on Computer Vision ACCV
Magee Derek
,
Leibe Bastian
2003
“Online Face Tracking Using a Feature Driven Levelset,”
British Machine Vision Conference BMVC'03
Wang Jun
,
Peng Jinye
,
Feng Xiaoyi
,
Li Linqing
,
Li Danjiao
“An improved camshiftbased particle filter algorithm for face tracking,” Yanning Zhang, ZhiHua Zhou, Changshui Zhang, and Ying Li (Eds.)
SpringerVerlag
Berlin, Heidelberg
In Proceedings of the Second Sinoforeigninterchange conference on Intelligent Science and Intelligent Data Engineering (IScIDE'11)
278 
285
Gordon N.
,
Salmond D.
,
Smith A.
1993
“Novel approach to nonlinear/nonGaussian Bayesian state estimation,”
IEE Procedings F.
140
(2)
107 
113
Carpenter J.
,
Clifford P.
,
Fernhead P.
1997
An improved particle filter for nonlinear problems
Department of Statistics, University of Oxford, Tech. Rep.
Arulampalam M. S.
,
Maskell S.
,
Gordon N.
,
Clapp T.
2002
“A tutorial on particle filters for online nonlinear/nonGaussian Bayesian tracking,”
IEEE Trans. Signal Process
50
(2)
174 
188
DOI : 10.1109/78.978374
Perez P.
,
Hue Carine
,
Vermaak Jaco
,
Gangnet Michel
“ColorBased Probabilistic Tracking,”
SpringerVerlag
London, UK, UK
In Proceedings of the 7th European Conference on Computer VisionPart I ECCV '02
661 
675
Doucet A.
,
de Freitas N.
,
Gordon N.
2001
Sequential Monte Carlo Methods in Practice
SpringerVerlag
New York
Khan Z.
,
Balch T.
,
Dellaert F.
2005
“An MCMCbased particle filter for tracking multiple interacting targets,”
IEEE Trans. on PAMI
27
(11)
1805 
1918
DOI : 10.1109/TPAMI.2005.223
Zuriarrain I.
,
Lerasle F.
,
Arana N.
,
Devy M.
2008
“An MCMCbased particle filter for multiple person tracking,”
Pattern Recognition. ICPR
1 
4
Duffner S.
,
Odobez J.
2011
“Exploiting longterm observations for track creation and deletion in online multiface tracking,”
Automatic Face & Gesture Recognition and Workshops FG
525 
530
Vu AnhTuyet
,
Vo BaNgu
,
Evans R.
2011
“Particle Markov Chain Monte Carlo for Bayesian Multitarget Tracking,”
Information Fusion FUSION
1 
8
Zhou Xiuzhuang
,
Lu Yao
,
Lu Jiwen
,
Zhou Jie
2012
“Abrupt Motion Tracking Via Intensively Adaptive MarkovChain Monte Carlo Sampling,”
Trans. Img. Proc.
21
(2)
789 
801
DOI : 10.1109/TIP.2011.2168414
Isard M.
,
MacCormick J.
2001
“BraMBLe: A Bayesian multipleblob tracker,”
Intl. Conf. on Computer Vision ICCV
34 
41
Chen H. T.
,
Liu T. L.
2001
“Trustregion methods for realtime tracking,”
In Proc. Int. Conf. Computer Vision
717 
722
Comaniciu D.
,
Ramesh V.
,
Meer P.
2000
“Realtime tracking of nonrigid objects using mean shift,”
In Proc. Conf. Comp. Vision Pattern Rec.
142 
149
Rosten Edward
,
Drummond Tom
“Machine learning for highspeed corner detection,” Aleš Leonardis, Horst Bischof, and Axel Pinz (Eds.), vol. Part I
SpringerVerlag
Berlin, Heidelberg
In Proceedings of the 9th European conference on Computer Vision  Volume Part I (ECCV'06)
430 
443