4 tracking objects of deformable shapes

International Journal of of Computer Science Engineering Research and Development (IJCSERD), ISSN
International Journal Computer Science and and Engineering
IJCSERD
2248-9363and Development (IJCSERD), ISSN 2248-9363 1, April-June (2011)
Research (Print), ISSN 2248-9371 (Online), Volume 1, Number
(Print), ISSN 2248-9371 (Online)
Volume 1, Number 1, April-June (2011)
© PRJ PUBLICATION
pp. 28-37 © PRJ Publication
http://www.prjpublication.com/IJCSERD.asp

TRACKING OBJECTS OF DEFORMABLE SHAPES

K. Mahesh
Associate Professor, Department of Computer Science and Engg
Alagappa University, Karaikudi,TN,India.
mahesh.alagappa@gmail.com

Dr.K.Kuppusamy
Associate Professor, Department of Computer Science and Engg
Alagappa University, Karaikudi,TN,India.
kkdisamy@yahoo.com

G. Radha Priyadharsini
Department of Computer Science and Engineering
Alagappa University, Karaikudi,TN, India.
rpgmphil@gmail.com

ABSTRACT

We propose a solution to determine the optimal elastic matching of a deformable template to an
image. The central idea is to cast the optimal matching of each template point to a corresponding
image pixel as a problem of finding a minimum cost cyclic path in the three-dimensional product
space as well as in four-dimensional product space spanned by the template and the input image.
We introduce a cost functional associated with each cycle, which consists of three terms: a data
fidelity term favoring strong intensity gradients, a shape consistency term favoring similarity of
tangent angles of corresponding points, and an elastic penalty for stretching or shrinking. The
functional is normalized with respect to the total length to avoid a bias toward shorter curves.
Optimization is performed by Lawler’s Minimum Ratio Cycle algorithm parallelized on state-of-
the-art graphics cards. The algorithm provides the optimal segmentation and point
correspondence between template and segmented curve in computation times that are essentially
linear in the number of pixels.A new approach to 4-D shape-based segmentation and tracking of
multiple,deformable anatomical structures used in cardiac MR images can be implemented
here.We propose to use an energy-minimizing geometrically deformable template(GDT) which
can deform into similar shapes under the influence of image forces.The degree of deformation of
the template from its equilibrium shape is measured by a penalty function associated with
mapping between the two shapes.By minimizing this term along with the image energy terms of
the classic deformable model,the deformable template is attracted towards objects in the image
whose shape is similar to its equilibrium shape.This allows the simultaneous segmentation of
multiple deformable objects using intra-as well as inter-shape information.Simulated
Annealing(SA),a stochastic relaxation technique is used for segmentation while Iterated
Conditional Modes(ICM),a deterministic relaxation technique is used for tracking.

Keywords: Image segmentation, tracking, elastic shape priors, discrete optimization, dynamic
programming, minimum ratio cycles, Simulated annealing, real-time applications

28

International Journal of Computer Science and Engineering Research and Development (IJCSERD), ISSN
2248-9363 (Print), ISSN 2248-9371 (Online), Volume 1, Number 1, April-June (2011)

1. INTRODUCTION

IMAGE segmentation and the tracking of objects are two of the most prominent topics in
computer vision. Numerous authors have tried to solve these problems based on low level
information such as edges or region statistics [25], [34],[3], [42], [21]. However, their success has
been limited: In real-world images, the low-level information is often corrupted, e.g., by changing
lighting conditions and low contrast between object and background. As an example, consider
Figure. 1, where a car is tracked in rainy weather. To cope with such challenges, researchers have
endeavored to integrate prior knowledge into the respective segmentation processes. In numerous
studies [19], [4], [13], [7], this was shown to significantly improve the resulting segmentations.
However, most of these methods find local minima, and hence, require an initialization in the
vicinity of the solution. The one that does find globally optimal segmentations [13] has a
quadratic memory complexity. It is hence well suited for tracking tasks, but for pixel-accurate
image segmentation, only rather coarse resolutions can be handled. In this work, we present the
first globally optimal shape based segmentation method able to yield pixel-accurate
segmentations in effectively linear time.

The segmentation problem in complex images cannot be addressed adequately without the
anatomical a-priori knowledge which usually aids in making decisions about the image
segmentation. In this case two major sources of a-priori knowledge can be identified:
A-priori information about the mean shape and the variability of anatomical objects.
A-priori information about the mean location, orientation and size of the objects with respect to
each other and their variability.

Here we address the segmentation and tracking problem in these images using Geometrically
Deformable Templates (GDT).This novel approach differs in the following points from
previously described models: Its deformation is controlled via a penalty function rather than via
its parameterization .This penalty function is associated with a thin-plate spline(TPS)mapping
function which maps the templates in its equilibrium configuration into a deformed configuration.
One can visualize thin plate spline mapping function as an imaginary rectangular grid associated
with the model in equilibrium. Any deformation of the model would also deform the rectangular
grid. The penalty function requires energy for any non-affine deformation of the grid but does not
penalize affine deformations. Moreover , the model can incorporate not only information about
the mean location, orientation and size of the anatomical objects with respect to each other and
their variability. Thus , the model can be used to segment multiple objects simultaneously.

1.1 Related Work
Image segmentation and tracking are closely related problems, yet each with its own history. We,
therefore, review them separately.

1.1.1 Tracking Deformable Objects
The tracking of objects has traditionally been based on feature points [17], [22]. Starting from the
KLT-tracker [42],subsequent feature-based methods appeared in [21], [30].More recently,
methods have become popular that treat the object as an entity [11], [6], [20] rather than an
independent number of parts. Denzler and Niemann [11] consider a set of patches that is linked
by a ray model. Cremers [6] models the temporal evolution of shapes by a dynamical,
autoregressive model in a level set framework. This is extended by Gui et al. [20] to the case of
competing priors. While many of these methods are based on minimizing a suitable energy, none
guarantees to find the global optimum. However, such methods give neither a guarantee to find
good (i.e., low energy) solutions nor a means to verify if a solution is optimal. To determine

29


global optima in the presence of significantly deforming curves has remained an open challenge.
Furthermore, real-world applications typically require fast algorithms that can run in real time.

1.2 Shape-Based Image Segmentation

The task to partition an image into meaningful regions has received considerable attention in the
past. When the limits of low-level methods [25], [34], [3] became apparent, researchers
endeavored to integrate prior knowledge into segmentation processes. The amount of prior
knowledge varies from a part-based (deformable) object structure [15], [14], [35] over a
collection of shapes [19], [4], [29], [43], [9], [37], [7], [8], [28] to a single shape [13].

Figure 1. Tracking a car in bad weather: Despite bad visibility, reflections, and camera shake, the
proposed method allows reliable tracking over a hundreds of frames

Such methods are bound to find local optima of the energy they are optimizing and heavily
depend on the initial contour. In addition, they are based on rather simple shape similarity
measures, which do not attempt to establish correspondences of parts or points.

Recently, Cremers et al. [8] dealt with the first point: Starting from an implicit representation of
shapes and segmentations, they are able to find globally optimal
segmentations while taking into account shape similarity to a number of training shapes.
However, the lack of point correspondences remains. In contrast, a number of discrete approaches
do allow shape priors based on point correspondences while guaranteeing global optimality:
Coughlan et al. [5] are able to match open contours to images, taking into account an elastic
shape similarity measure [31], [32]. Being based on dynamic programming, the method is, in
principle, parallelizable. However, it is limited to open contours, and hence, does not provide a
segmentation. Although the method could be extended to closed contours by performing a
complete search over the start point; in practice, this would be far too time-consuming.

The first globally optimal shape-based segmentation algorithm was proposed by Felzenszwalb
[13]. It is based on dynamic programming in chordal graphs. The algorithm is easy to parallelize
and invariant with respect to translation, rotation, and scale changes. In practice, however, due to
its quadratic memory complexity, pixel accurate segmentations can only be computed on rather
coarse resolution.

1.2 Contribution

In this paper, we present an effectively linear-time algorithm to match contours to images closed
contours reduce the bias toward short curves by reverting to ratio functional and minimum ratio
cycle computation. The proposed method supports different amounts of invariances, including

30


translational and rotational ones. By exploiting its high parallelizability, real-time tracking
becomes feasible.

Optimization of GDT’S
Estimating the maximum a posteriori (MAP) solution directly is usually impossible due to the
size of the configuration space, even for template models with very few vertices. Instead we have
implemented two different optimization techniques for the segmentation and tracking process:
Simulated Annealing(SA) minimization technique is used during the segmentation process and
Iterated Conditional Modes(ICM) as an efficient local minimization technique is used during the
tracking process. Simulated Annealing is a stochastic relaxation technique which generates
randomly new configurations by sampling. Iterated Conditional Mode is a deterministic
relaxation algorithm. It is very well suited for tracking objects if the temporal resolution is high
enough.

2 MATCHING’S AS CYCLES IN A PRODUCT SPACE

We are given a prior contour S :S11 -> IR2 (where S1 is the unit circle) with a uniform
parameterization. The task is to match this contour to a given image I : -> IR, where ,<- IR2 is
the (typically rectangular) image domain. The placed contour C:S1-> should be similar to the
input contour and fulfill some data-driven criteria. In this work, we want it to be located at image
edges. Figure 2 gives an illustration of our approach: When a contour and an image are input, the
algorithm locates a deformed version of the contour in the image and computes an alignment to
the prior contour.

Figure 2 Starting from (a) a prior contour and (b) an input image, the proposed method
simultaneously locates (c) the (possibly deformed) contour in the image and computes (d) a
correspondence function between the two curves

31


Figure 3 Cyclic paths in a 3D graph (no edges are shown): For any point on the prior contour,
there are K copies of the image in the graph. Any assignment of pixels in the image to
corresponding points on the template contour corresponds to a cyclic path in this graph.
correspondence. This allows to use correspondence-based shape similarity measures, which were
shown to be important for reproducing human notions of shape similarity [18], [26].

While this was known for open curves [5], the computationally much more challenging case of
closed curves has so far not been solved. The product space arises by combining the functions C
and m into a single function Ґ: S1 -> ×S1 which is called a cycle. The space in which these
cycles live is visualized in Figure. 3. It has the form of a torus and arises by placing a copy of the
image for each point on the (onedimensional) prior contour. When splitting a closed contour at
some point, it can be viewed as an open one. The space would then be a solid block. When
additionally imposing that start and end points are identical, the respective end faces of the block
have to meet and the torus is formed.A curve (with winding number 1) in this space now allows
to read off the desired information: The curve C is obtained by projecting _ to the first two
dimensions. The correspondences of the points on C can be read off in the third dimension.

3 ASSIGNING A COST TO EACH CYCLE

We now present an exemplary energy functional for matching shapes to images. The presented
method applies to a much larger class of functionals. For example, in [40], we used a more
sophisticated data term based on patch comparison. Before we state the cost function, we briefly
discuss how curves are represented.
3.1 Representing Curves

There are infinitely many ways to parameterize a specific curve. Naturally, an optimization
problem should not depend on the chosen parameterization.

Figure 4 The three ingredients of the proposed method: (a) An edge detector function assigning
low values to high image gradients.(b) Computation of tangent angles of the contours C and S
(shown for C, the tangent is drawn in black). (c) Computation of length distortion.

32


The functional we consider in this paper is indeed invariant with respect to reparameterizations.
For most of this section, we will therefore not assume any specific parameterization of the
contour C to be optimized. Yet, in a few places, it will be convenient to have a uniform
parameterization, i.e., with constant derivative ║Cs(s)║=║C║ kCk everywhere. In the given
setting, the correspondence function m is dependent on the contour C: m(s) will always denote
the correspondence of the point C(s). Hence, if the parameterization of C is changed, the function
m changes as well.In subsequent sections, we prefer the combined function Ґ: S1 -> ×S1. Since
the objective function is not invariant against reparameterizations of Ґ (data term and shape
measure are not coupled), we state it in terms of C and m.

We allow self-intersecting contours C since we have no means to exclude them. We found them
to arise only seldom as long as the desired object is truly contained in the image.

4 DISCRETIZING COST AND PRODUCT SPACE

To optimize over the cycles , , both the cost function and the product space are
discretized. This section deals with the discretization, the optimization algorithm is detailed in the
next one. The key idea is to represent C as a polygonal curve with (an a priori unknown number
of) vertices on the pixel grid. In addition, the correspondence m is assumed to be linear along
each polygonal line segment. It is therefore uniquely defined by assigning point correspondences
to the two end points of such a segment. Specifically, we consider line segments connecting
neighboring pixels on the pixel grid,where we choose an 8-neighborhood. A cycle can now be
composed out of a finite set of basic parts 4.1 Discretizing Prior Contour and
Correspondence In addition to the cycle , the prior contour S is also discretized. we represent it
in the same form as the contour C, i.e., as an ordered set of points on a suitable pixel
grid, where—for ease of notation— is represented twice. To get a dense representation
of the contour, we require that si be among the eight closest neighbors of The discrete
correspondence function assigns each image pixel on C one of these prior points. To
ensure a monotone matching, we enforce that the start pixel of a segment is assigned a shape
point with index lower than or equal to that of the endpoint. Closure of the matching is obtained
by the fact that .The length distortion penalizer gives two hard constraints, which limit
the minimal and maximal distortion ratio. In the discrete setting, these are realized in terms of the
indices of the two shape points assigned to a line segment. The upper limit corresponds to an
index difference of at most K. Ensuring the lower limit is more intricate since here several line
segments may correspond to the same part of S. We therefore allow the two indices
to be equal.

However, for any shape point si, there may be at most K parts, where both endpoints of
correspond to si.In practice, this is realized by modifying the correspondence function m: In the
discrete setting, m maps to pairs (i,k) where i gives the shape point and k < K gives the number of
image pixels already corresponding to si. If m maps to the same i at the beginning and end of the
contour segment, then the index k must be one higher for the end node. This is formalized in the
following section.

33


Figure 5 Segmentation with a single template: Despite significant deformation, scale change, and
translation, the initial template curve (red) is accurately matched to each template.

5 SHAPE - BASED IMAGE SEGMENTATION

We treat images with significant distortion. As a consequence, we allow K=5 image pixels to be
matched to a single shape point and set a low length distortion weight with λ=0.1. The tangent
angles are given more weight with ν=0.5—this term really drives the process.

5.1 Translation-Invariant Matching

In Figure 5, the contour of a rabbit (viewed from the side) is matched to images from two
different sequences. In the first sequence, the rabbit is shown from different viewpoints but at the
same scale. Despite low contrast between object and background, the algorithm relocates the
object reliably.

5.2 On the Effect of Length Normalization

We introduced the length normalization to reduce the bias toward shorter curves. This effect is
demonstrated in Figure. 6: The figure shows the global optima for the ratio functional and for the
numerator integral alone. The latter corresponds to the geodesic energy we proposed in [41]:

Figure 6 The length normalization removes the bias toward shorter curves.

34


and is minimized globally by a combination of branch and bound and the shortest path
algorithms. Clearly, the ratio functional produces longer curves. We observe this whenever there
is low contrast in some regions along the desired curve.

5.3 Including Rotational Invariance

Aside from translational invariance, sometimes one also wants rotational invariance. The
proposed framework can be easily extended to include this: One simply samples the rotation
angle in sufficiently dense intervals. The prior contour is rotated by the specified amount and the
obtained contour is matched to the image.

6 SHAPE-BASED TRACKING

In this section, we present the problem of tracking deformable objects (or contours). In the first
frame, the contour S is given. Then subsequently, we map the contour determined for the
previous frame to the current frame. This performs better than keeping a fixed template since
large-scale deformations are decomposed into a sequence of smaller ones.

7 CONCLUSION

In this paper, we introduced a polynomial-time solution for matching a given contour to an image
despite translation, rotation, scale change, and deformation. The central idea is to cast the
assignment of an image pixel to each template point as a problem of finding optimal ratio cycles
in a 3D graph that represents the product space of image and template. The energy that is
optimized globally consists of an edge-based data term and a shape similarity measure favoring
similarity of local edge angles and minimal distortion (stretching/shrinking) of the template
curve.

We propose to use an energy-minimizing geometrically deformable template(GDT) which can
deform into similar shapes under the influence of image forces. This allows the simultaneous
segmentation of multiple deformable objects using intra-as well as inter-shape information.
Simulated Annealing(SA).

REFERENCES

[1] R.E. Bellman, “On a Routing Problem,” Quarterly Applied Math., vol. 16, pp.
87-90, 1958.
[2] A. Blake and M. Isard, Active Contours. Springer Verlag, 1998.
[3] V. Caselles, R. Kimmel, G. Sapiro, and C. Sbert, “Minimal Surfaces Based Object
Segmentation,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 19, no. 4, pp. 394-
398, Apr. 1997.
[4] T. Cootes and C.J. Taylor, “Active Shape Model Search Using Local Grey-Level Models: A
Quantitative Evaluation,” Proc. British Machine Vision Conf., pp. 639-648, 1993.
[5] J. Coughlan, A. Yuille, C. English, and D. Snow, “Efficient Deformable Template Detection
and Localization without User Initialization,” Computer Vision and Image Understanding, vol.
78, no. 3, pp. 303-319, 2000.
[6] D. Cremers, “Dynamical Statistical Shape Priors for Level Set Based Tracking,” IEEE Trans.
Pattern Analysis and Machine Intelligence, vol. 28, no. 8, pp. 1262-1273, Aug. 2006.

35


[7] D. Cremers, S.J. Osher, and S. Soatto, “Kernel Density Estimation and Intrinsic Alignment
for Shape Priors in Level Set Segmentation,” Int’l J. Computer Vision, vol. 69, no. 3, pp. 335-
351, Sept. 2006.
[8] D. Cremers, F.R. Schmidt, and F. Barthel, “Shape Priors in Variational Image Segmentation:
Convexity, Lipschitz Continuity and Globally Optimal Solutions,” Proc. IEEE Int’l Conf.
Computer Vision and Pattern Recognition, June 2008.
[9] D. Cremers, F. Tischha¨user, J. Weickert, and C. Schno¨ rr, “Diffusion Snakes: Introducing
Statistical Shape Knowledge into the Mumford-Shah Functional,” Int’l J. Computer Vision, vol.
50, no. 3, pp. 295-313, 2002.
[10] F. Dellaert and C. Thorpe, “Robust Car Tracking Using Kalman Filtering and Bayesian
Templates,” Proc. Conf. Intelligent Transportation Systems, 1997.
[11] J. Denzler and H. Niemann, “Active Rays: Polar-Transformed Active Contours for Real-
Time Contour Tracking,” Real-Time Imaging, vol. 5, pp. 203-213, 1999.
[12] A. Doucet, N. de Freitas, and N. Gordon, Sequential Monte Carlo Methods in Practice
(Statistics for Engineering and Information Science). Springer Verlag, 2001.
[13] P.F. Felzenszwalb, “Representation and Detection of Deformable Shapes,” IEEE Trans.
Pattern Analysis and Machine Intelligence, vol. 27, no. 2, pp. 208-220, Feb. 2005.
[14] P.F. Felzenszwalb and D. Huttenlocher, “Pictorial Structures for Object Recognition,” Int’l J.
Computer Vision, vol. 61, no. 1, pp. 55-79, 2005.
[15] M.A. Fischler and R.A. Eschlager, “The Representation and Matching of Pictorial
Structures,” IEEE Trans. Computers, vol. 22, no. 1, pp. 67-92, Jan. 1973.
[16] L.R. Ford, “Network Flow Theory,” Paper P-923, The Rand Corporation, 1956.
[17] W. Fo¨rstner and E. Gu¨ lch, “A Fast Operator for Detection and Precise Localization of
Distinct Points, Corners and Circular Features,” Proc. Intercommission Conf. Fast Processing of
Photogrammetric Data, pp. 281-305, 1987.
[18] Y. Gdalyahu and D. Weinshall, “Flexible Syntactic Matching of Curves and Its Application
to Automatic Hierarchical Classification of Silhouettes,” IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 21, no. 12, pp. 1312-1328, Dec. 1999.
[19] U. Grenander, Y. Chow, and D.M. Keenan, Hands: A Pattern Theoretic Study of Biological
Shapes. Springer Verlag, 1991.
[20] L. Gui, J. Thiran, and N. Paragios, “Joint Object Segmentation and Behaviour Classification
in Image Sequences,” Proc. IEEE Int’l Conf. Computer Vision and Pattern Recognition, 2007.
[21] G.D. Hager and P.N. Belhumeur, “Real-Time Tracking of Image Regions with Changes in
Geometry and Illumination,” Proc. IEEE Int’l Conf. Computer Vision and Pattern Recognition,
pp. 403-410, 1996.
[22] C. Harris and M. Stephens, “A Combined Corner and Edge Detector,” Proc. Fourth Alvey
Vision Conf., pp. 147-151, 1988.
[23] A. Jalba, M. Wilkinson, and J. Roerdink, “CPM: A Deformable Model for Shape Recovery
and Segmentation Based on Charged Particles,” IEEE Trans. Pattern Analysis and Machine
Intelligence, vol. 26, no. 10, pp. 1320-1335, Oct. 2004.
[24] I.H. Jermyn and H. Ishikawa, “Globally Optimal Regions and Boundaries as Minimum
Ratio Weight Cycles,” IEEE Trans.Pattern Analysis and Machine Intelligence, vol. 23, no. 10, pp.
1075-1088, Oct. 2001.
[25] M. Kass, A. Witkin, and D. Terzopoulos, “Snakes: Active Contour Models,” Int’l J.
Computer Vision, vol. 1, no. 4, pp. 321-331, 1988.
[26] L.J. Latecki and R. Laka¨mper, “Shape Similarity Measure Based on Correspondence of
Visual Parts,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 22, no. 10, pp. 1185-
1190, Oct. 2000.
[27] E.L. Lawler, “Optimal Cycles in Doubly Weighted Linear Graphs,” Proc. Int’l Symp.
Theory of Graphs, pp. 209-213, 1966.

36


[28] V. Lempitsky, A. Blake, and C. Rother, “Image Segmentation by Branch-and-Mincut,” Proc.
European Conf. Computer Vision, Oct. 2008.

[29] M. Leventon, W. Grimson, and O. Faugeras, “Statistical Shape Influence in Geodesic Active
Contours,” Proc. IEEE Int’l Conf. Computer Vision and Pattern Recognition, vol. 1, pp. 316-323,
2000.
[30] D. Lowe, “Object Recognition from Local Scale-Invariant Features,” Proc. IEEE Int’l Conf.
Computer Vision, Sept. 1999.
[31] M. Maes, “Polygonal Shape Recognition Using String-Matching Techniques,” Pattern
Recognition, vol. 24, no. 5, pp. 433-440, 1991.
[32] R. McConnell, R. Kwok, J.C. Curlander, W. Kober, and S.S. Pang, Correlation and
Dynamic Time Warping: Two Methods for Tracking Ice Floes in SAR Images,” IEEE Trans.
Geosciences and Remote Sensing, vol. 29, no. 11, pp. 1004-1012, Nov. 1991.
[33] E.F. Moore, “The Shortest Path through a Maze,” Proc. Int’l Symp. Theory of Switching, pp.
285-292, 1959.
[34] D. Mumford and J. Shah, “Optimal Approximations by Piecewise Smooth Functions and
Associated Variational Problems,” Comm. Pure and Applied Math., vol. 42, pp. 577-685, 1989.
[35] D. Ramanan and C. Sminchisescu, “Training Deformable Models for Localization,” Proc.
IEEE Int’l Conf. Computer Vision and Pattern Recognition, pp. 206-213, June 2006.
[36] M. Rousson and D. Cremers, “Efficient Kernel Density Estimation of Shape and Intensity
Priors for Level Set Segmentation,” Proc. Int’l Conf. Medical Image Computing and Computer
Assisted Intervention, vol. 1, pp. 757-764, 2005.
[37] M. Rousson and N. Paragios, “Shape Priors for Level Set Representations,” Proc. European
Conf. Computer Vision, pp. 78- 92, 2002.
[38] T. Schoenemann and D. Cremers, “Globally Optimal Image Segmentation with an Elastic
Shape Prior,” Proc. IEEE Int’l Conf. Computer Vision, Oct. 2007.
[39] T. Schoenemann and D. Cremers, “Globally Optimal Shape-Based Tracking in Real-Time,”
Proc. IEEE Int’l Conf. Computer Vision and Pattern Recognition, June 2008.
[40] T. Schoenemann and D. Cremers, “Matching Non-Rigidly Deformable Shapes across
Images: A Globally Optimal Solution,” Proc. IEEE Int’l Conf. Computer Vision and Pattern
Recognition, June 2008.
[41] T. Schoenemann, F.R. Schmidt, and D. Cremers, “Image Segmentation with Elastic Shape
Priors via Global Geodesics in Product Spaces,” Proc. British Machine Vision Conf., Sept. 2008.
[42] J. Shi and C. Tomasi, “Good Features to Track,” Proc. IEEE Int’l Conf. Computer Vision
and Pattern Recognition, June 1994.
[43] A. Tsai, A. Yezzi, W. Wells, C. Tempany, D. Tucker, A. Fan, E. Grimson, and A. Willsky,
“Model-Based Curve Evolution Technique for Image Segmentation,” Proc. IEEE Int’l Conf.
Computer Vision and Pattern Recognition, pp. 463-468, 2001.
[44] X. Xie and M. Mirmehdi, “MAC: Magnetostatic Active Contour Model,” IEEE Trans.
Pattern Analysis and Machine Intelligence, vol. 30, no. 4, pp. 632-646, Apr. 2008.
[45] C. Xu and J. Prince, “Generalized Gradient Vector Flow External Forces for Active
Contours,” Signal Processing, vol. 71, no. 2, pp. 131-139, 1998.

37

4 tracking objects of deformable shapes

More Related Content

What's hot (15)

Similar to 4 tracking objects of deformable shapes (20)

More from prjpublications (20)

4 tracking objects of deformable shapes