28-04-2014, 02:18 PM
Active Contour-Based Visual Tracking by Integrating Colors, Shapes, and Motions
[attachment=62608]
Abstract
In this paper, we present a framework for active
contour-based visual tracking using level sets. The main
components of our framework include contour-based tracking
initialization, color-based contour evolution, adaptive shape-
based contour evolution for non-periodic motions, dynamic
shape-based contour evolution for periodic motions, and the
handling of abrupt motions. For the initialization of contour-
based tracking, we develop an optical flow-based algorithm for
automatically initializing contours at the first frame. For the
color-based contour evolution, Markov random field theory is
used to measure correlations between values of neighboring pixels
for posterior probability estimation. For adaptive shape-based
contour evolution, the global shape information and the local
color information are combined to hierarchically evolve the con-
tour, and a flexible shape updating model is constructed. For the
dynamic shape-based contour evolution, a shape mode transition
matrix is learnt to characterize the temporal correlations of
object shapes. For the handling of abrupt motions, particle swarm
optimization is adopted to capture the global motion which is
applied to the contour in the current frame to produce an initial
contour in the next frame.
INTRODUCTION
VISUAL object tracking is an active research topic in
computer vision. In contrast to general object tracking
which uses predefined coarse shape models, such as rectangles
or ellipses, to represent objects [5], active contour-based
tracking [36] provides more detailed object shape information,
but is, in general, more difficult than general tracking of the
same object in the same real-world situation. This is because
contour tracking aims to recover finer details of the object,
i.e., the boundary of the object, and the determination of
the boundary of the object is susceptible to influences from
the background disturbance. In videos taken by stationary
cameras, object motion regions can often be extracted using
background subtraction, and object contours can be produced
by tracing the edges of the motion regions. But in videos taken
by moving cameras, background subtraction cannot be used to
extract object motion regions, making the contour-based track-
ing more difficult than in videos taken by stationary cameras.
Related Work
There are in general two ways to describe object contours:
explicit representations which are characterized by parame-
terized curves such as snakes [1] and implicit representations,
such as level sets [3], which represent a contour using a signed
distance map. The level set representation is more popular than
the explicit representation because it has a stable numerical
solution and it is capable of handling topological changes.
Active contour evolution methods are classified into three
categories: edge-based, region-based, and shape prior-based.
1) Edge-Based Methods: Edge-based methods mainly con-
sider the local information around contours, such as the grey
level gradient. Kass et al. [1] propose the snake model which
is the best known edge-based active contour method. Caselles
et al. [12] propose a geodesic model which reflects more
intrinsic geometric image measures than the snake, using the
prior knowledge that the larger the gradient at a pixel, the
higher the probability that the pixel belongs to an object’s
edge. Paragios and Deriche [13] improve the geodesic model
in [12] using level sets to describe contours and using a
gradient descent algorithm to optimize contours.
Construction of Shape Subspace
The contour shape subspace is constructed in the following
way which is similar to that of [18]. From a training sequence,
we manually obtain a series of training shape samples of
the object which is to be tracked in the test sequence. All
the signed distance maps of each sample are aligned using the
shape registration. The level set embedding function values
in each distance map are flattened into a column vector [19].
The mean vector μ is obtained by taking the mean of the
column vectors for all the samples. A matrix X whose columns
are obtained by subtracting μ from each sample column
vector is constructed. Using the singular value decomposition
(SVD) for X, the diagonal matrix k with the first k largest
singular values and the corresponding singular column vector
matrix Uk are obtained. The shape model is represented by
{μ, Uk , k }.
CONCLUSION
In this paper, we have presented an effective framework for
tracking object contours. We have the following conclusions:
1) Our color-based contour evolution algorithm which applies
the MRF theory to model the correlations between pixel values
for posterior probability estimation is more robust to back-
ground disturbance than the region-based method which does
not consider correlations between the values of neighboring
pixels for posterior probability estimation. 2) Our adaptive
shape-based contour evolution algorithm, which efficiently
fuses the global shape information and the local color infor-
mation and uses a flexible shape model updating algorithm, is
robust to partial occlusions, weak contrast at the boundaries,
and motion blurring, etc. 3) Our dynamical shape prior model
effectively characterizes the temporal correlations between
contour shapes in periodic motions, and thus it obtains more
accurate contours than the existing autoregressive model.
4) Our PSO-based algorithm can deal effectively with contour
tracking for videos with abrupt motions, and it outperforms
the particle filter-based algorithm.
[attachment=62608]
Abstract
In this paper, we present a framework for active
contour-based visual tracking using level sets. The main
components of our framework include contour-based tracking
initialization, color-based contour evolution, adaptive shape-
based contour evolution for non-periodic motions, dynamic
shape-based contour evolution for periodic motions, and the
handling of abrupt motions. For the initialization of contour-
based tracking, we develop an optical flow-based algorithm for
automatically initializing contours at the first frame. For the
color-based contour evolution, Markov random field theory is
used to measure correlations between values of neighboring pixels
for posterior probability estimation. For adaptive shape-based
contour evolution, the global shape information and the local
color information are combined to hierarchically evolve the con-
tour, and a flexible shape updating model is constructed. For the
dynamic shape-based contour evolution, a shape mode transition
matrix is learnt to characterize the temporal correlations of
object shapes. For the handling of abrupt motions, particle swarm
optimization is adopted to capture the global motion which is
applied to the contour in the current frame to produce an initial
contour in the next frame.
INTRODUCTION
VISUAL object tracking is an active research topic in
computer vision. In contrast to general object tracking
which uses predefined coarse shape models, such as rectangles
or ellipses, to represent objects [5], active contour-based
tracking [36] provides more detailed object shape information,
but is, in general, more difficult than general tracking of the
same object in the same real-world situation. This is because
contour tracking aims to recover finer details of the object,
i.e., the boundary of the object, and the determination of
the boundary of the object is susceptible to influences from
the background disturbance. In videos taken by stationary
cameras, object motion regions can often be extracted using
background subtraction, and object contours can be produced
by tracing the edges of the motion regions. But in videos taken
by moving cameras, background subtraction cannot be used to
extract object motion regions, making the contour-based track-
ing more difficult than in videos taken by stationary cameras.
Related Work
There are in general two ways to describe object contours:
explicit representations which are characterized by parame-
terized curves such as snakes [1] and implicit representations,
such as level sets [3], which represent a contour using a signed
distance map. The level set representation is more popular than
the explicit representation because it has a stable numerical
solution and it is capable of handling topological changes.
Active contour evolution methods are classified into three
categories: edge-based, region-based, and shape prior-based.
1) Edge-Based Methods: Edge-based methods mainly con-
sider the local information around contours, such as the grey
level gradient. Kass et al. [1] propose the snake model which
is the best known edge-based active contour method. Caselles
et al. [12] propose a geodesic model which reflects more
intrinsic geometric image measures than the snake, using the
prior knowledge that the larger the gradient at a pixel, the
higher the probability that the pixel belongs to an object’s
edge. Paragios and Deriche [13] improve the geodesic model
in [12] using level sets to describe contours and using a
gradient descent algorithm to optimize contours.
Construction of Shape Subspace
The contour shape subspace is constructed in the following
way which is similar to that of [18]. From a training sequence,
we manually obtain a series of training shape samples of
the object which is to be tracked in the test sequence. All
the signed distance maps of each sample are aligned using the
shape registration. The level set embedding function values
in each distance map are flattened into a column vector [19].
The mean vector μ is obtained by taking the mean of the
column vectors for all the samples. A matrix X whose columns
are obtained by subtracting μ from each sample column
vector is constructed. Using the singular value decomposition
(SVD) for X, the diagonal matrix k with the first k largest
singular values and the corresponding singular column vector
matrix Uk are obtained. The shape model is represented by
{μ, Uk , k }.
CONCLUSION
In this paper, we have presented an effective framework for
tracking object contours. We have the following conclusions:
1) Our color-based contour evolution algorithm which applies
the MRF theory to model the correlations between pixel values
for posterior probability estimation is more robust to back-
ground disturbance than the region-based method which does
not consider correlations between the values of neighboring
pixels for posterior probability estimation. 2) Our adaptive
shape-based contour evolution algorithm, which efficiently
fuses the global shape information and the local color infor-
mation and uses a flexible shape model updating algorithm, is
robust to partial occlusions, weak contrast at the boundaries,
and motion blurring, etc. 3) Our dynamical shape prior model
effectively characterizes the temporal correlations between
contour shapes in periodic motions, and thus it obtains more
accurate contours than the existing autoregressive model.
4) Our PSO-based algorithm can deal effectively with contour
tracking for videos with abrupt motions, and it outperforms
the particle filter-based algorithm.