Journal of Signal and Information Processing, 2011, 2, 105-111
doi:10.4236/jsip.2011.22014 Published Online May 2011 (http://www.SciRP.org/journal/jsip)
Copyright © 2011 SciRes. JSIP
105
Tracking of Non-Rigid Object in Complex Wavelet
Domain
Om Prakash1,2, Ashish Khare1
1Department of Electronics & Communication, University of Allahabad, Allahabad, India; 2Centre of Computer Education, Univer-
sity of Allahabad, Allahabad, India.
Email: au.omprakash@gmail.com, ashishkhare@hotmail.com
Received December 13th, 2010; revised March 18th, 2011; accepted April 25th, 2011.
ABSTRACT
In this paper we have proposed an object tracking method using Dual Tree Complex Wavelet Transform (DTCxWT).
The proposed method is capable of tracking the moving object in video sequences. The object is assumed to be deform-
able under limit i.e. it may change its shape from one frame to another. The basic idea in the proposed method is to
decompose the image into two components: a two dimensional motion and a two dimensional shape change. The motion
component is factored out while the shape is explicitly represented by storing a sequence of two dimensional models.
Each model corresponds to each image frame. The proposed method performs well when the change in the shape in the
consecutive frames is small however the 2-D motion in consecutive frames may be large. The proposed algorithm is
capable of handling the partial as well as full occlusion of the object.
Keywords: Object Tracking, Dual Tree Complex Wavelet Transform, Model Based Tracking, Biorthogonal Filters
1. Introduction
In computer vision [1] applications, one of the challeng-
ing problems is the tracking of objects in video, in clut-
tered environment [2,3]. Object tracking and its applica-
tions are found in many diverse areas including medical
imaging [4,5], recogn ition and interpretation o f obj ects in
a scene [6], video surveillan ce, target detection and iden-
tification [4] etc.
Various spatial and frequency domain techniques are
used for these applications, but recent trend is to use
wavelet transforms for such applications [7]. However,
the form of wavelet used in these techniques has some
major drawbacks [8]. To overcome those drawbacks
some new form of wavelet transforms like complex
wavelet transform [9-14] has been used, which give bet-
ter results. The major limitations of Discrete Wavelet
Transform (DWT) are its shift sensitivity, poor direc-
tionality and ab sence of phase information. These limita-
tions are removed or at least reduced by using complex
wavelet Transforms. Dual tree complex wavelet trans-
form (DTCxWT) is one of the commonly used complex
wavelet transforms. Although Dual Tree Complex
Wavelet Transform (DTCxWT) suffers from the high
computational cost due to the redundancy of coefficients
but is free from shift variance [15] and directional selec-
tivity [11] problems which are useful in segmentation
and tracking of objects in different scenes. Many re-
searchers have used DTCxWT for tracking of rigid objects.
Although in some cases Daubechies Complex Wavelet
Transform has also been used [13]. Perfect reconstruc-
tion is one of the desirable properties for construction of
filters. Kingsbury [16,17] has introduced a complex
wavelet transform called DTCxWT which gives Perfect
Reconstruction and other properties like shift-invariance
and directional selectivity.
Rest of the paper is organized as follows: Section 2 of
this paper gives overview of complex wavelets and con-
struction of dual tree complex wavelet coefficients, Sec-
tion 3 describes the proposed method and its implemen-
tation while the Section 4 explains the results which are
tested and reported for different videos. At last, the Sec-
tion 5 of this paper describes the conclusions and future
scope.
2. Complex Wavelet Transform
Although different researchers have shown the magical
results after using real valued wav elet transforms but real
valued wavelet transforms suffer from the serious disad-
vantages like shift sensitivity and poor directionality.
One of the solutions to overcome these shortcomings is
Tracking of Non-Rigid Object in Complex Wavelet Domain
106
the use of Complex Wavelet Transform.
Kingsbury introduced a more computationally efficient
approach to shift invariance [15], the Dual-Tree Complex
Wavelet Transform (DTCxWT). Furthermore the DTCxWT
gives better directional selectivity when filtering multi-
dimensional signals. In summary, it has the following
properties:
Approximate shift invariance;
Good directional selectivity in 2-dimensions (2-D)
with Gabor-like filters (also true for higher dimen-
sionality, m-D);
Perfect reconstruction (PR) using short linear-
phase filters;
Limited redundancy, independent of the number of
scales, 2:1 for 1-D (2m:1 for m-D);
Efficient order-N computation – only twice the
simple DWT for 1-D (2m times for m-D).
Kingsbury [11] observed that the approximate shift
invariance with a real DWT can also be achieved by
doubling the sampling rate at each level of the tree. For
this to work, the samples must be evenly spaced. One
way to double all the sampling rates in a conventional
wavelet tree, such as Tree a of Figure 1, is to eliminate
the down-sampling by 2 after the level 1 filters, H0a and
H1a. This is equivalent to having two parallel fully- de-
cimated trees, a and b in Figure 1, provided that the de-
lays of filters H0b and H1b are one sample offset from the
delays of H0a and H1a, which ensures that the level 1
down samplers in tree b pick the opposite samples to
those in tree a. Then it is found that to get uniform inter-
vals between samples from the two trees below level 1,
the filters in one tree must provide delays that are half a
sample different (at each filter’s inpu t rate) from those in
the opposite tree. For linear phase filters, this requires
odd-length filters in one tree and even-length filters in the
other. Greater symmetry between the two trees occurs if
each tree uses odd and even filters alternately from level
to level, it has also been shown that the positions of the
wavelet basis functions when the filters are arranged to
be odd and even as in Figure 1.
2-band reconstruction block
Figure 1. Dual tree of the real filters for the complex wavelet transform (DTCxWT), giving real and imaginary parts of the
omplex coefficients . c
Copyright © 2011 SciRes. JSIP
Tracking of Non-Rigid Object in Complex Wavelet Domain107
To invert the DTCxWT, each tree in Figure 1 is in-
verted separately using biorthogonal filters G, designed
for perfect reconstruction with the corresponding analy-
sis filters H in the 2-band reconstruction block, shown in
Figure 1. Finally the two tree outputs are averaged in
order to obtain an approximately shift invariant system.
This system is a wavelet frame with r edu nd an cy two ; and
if the filters are designed such that the analysis and re-
construction filters have very similar frequency re-
sponses, then it is an almost tight frame, which means
that energy is approximately preserved when signals are
transformed into the DTCxWT domain. The basis func-
tions were obtained by injecting unit pulses separately
into the inverse DTCxWT at each scale in turn. The real
and imaginary parts were obtained by injecting the unit
pulses into trees a and b in turn.
3. The Proposed Object Tracking Approach
In the proposed method an object represented as a se-
quence of binary images, are corresponding to the each
frame of the input image sequence. Each frame of the
model corresponds to a set of pixels in a given region of
the corresponding image frame. A model Mt is assumed
to consist of m points (i.e. in the binary representation of
Mt there are m non-zero pixels). i.e. the image feature at
any instant of time consists of the model at that time.
These sets of pixels represent the shape of the object at
particular frame.
The main ideas of our proposed approach are as be-
low:
1) The 3D moving object is divided into two parts: a
2D shape change and a 2D motion. The change in the
shape should be small while there is no restriction on the
motion.
2) Change in 2D shape between successive image
frames is captured using 2D geometric models.
3) Perform 2D model matching using minimum
Hausdorff distance [18] which can be defined, for any
two point sets P and Q, as

 

,max,,,
H
PQhPQh QP
where,

,maxmin
qQ
pP
hPQp q
,
||.|| denotes Euclidean distance. This Hausdorff distance
simply measures the proximity of points in the two sets
which represents the difference between fixed point sets
whereas we require the difference between shapes of
point sets. Therefore, for some transformation group G,
minimum differ ence between two shapes w.r.t. th e group
action is given by
 

,min ,
GgG
DPQ HgPQ
In other words, the distance between two shapes is
minimum difference between them under all possible
transformations of one shape w.r.t. the other. DG(P,Q)
satisfies metric properties (identity, symmetry and trian-
gle inequality)
4) If DG(P,Q) is zero then two shapes are same other-
wise change in shape is measured.
5) Huttenlocher, et al. [18] has defined the partial dis-
tance or Rank order, which is used in tracking the 2-D
shape change, as
,
,min
th
KpPqQ
hPQKpq

This minimum directed partial Hausdorff distance [18]
is used to find the current location of the object. This
tracks the 2-D motion of the object. Also, the distance
between image and transformed model is used to select
those set of pixels which are the part of the next model,
gives us change in the 2-D shape of the object.
3.1. Algorithm for Tracking of Non-Rigid
Objects
1) Initialize the model Mt, which is assumed to consist of
m points (i.e. in the binary representation of Mt there are
m non-zero pixels), move it to the next time frame It + 1.
2) Find the new model Mt + 1, from Mt and It + 1.
3) Locate the object in the new image frame by com-
puting d as

1
1
,
min ,
min min
tt
Ktt
gG
th
gGpM qI
dhgMI
K
gp q
 
computed value of d identifies the transformation
g
G
of Mt which minimizes the rank order i.e. it
identifies the best “position”, g*, of Mt in the image
frame It + 1.
3.2. Finding the Model’s New Location
1) Identify all possible locations of the model M
t in the
next time frame It + 1.
2) Compute the set of transformations of Mt, say X
such that the partial Hausdorff distance is not larger than
some value τ i.e.
1
:,
Kt t
XxhMxI

where
is the Minkowski sum notation. The X can be
computed as
Compute distance transform Dt + 1 of It + 1 which is
an array specifying each location in the image the
distance to the nearest non-zero pixel of It + 1.
Compute directed Hausdorff distance, hK(Mt
x,
It + 1) for a given translation x of Mt.
This distance is the Kth largest of the m valued Dt + 1
Copyright © 2011 SciRes. JSIP
Tracking of Non-Rigid Object in Complex Wavelet Domain
108
Frame No. Spatial domain Real Wavelet Domain Complex Wavelet Domain
Frame-1
Frame-5
Frame-10
Frame-15
Frame-20
Frame-25
Frame-30
Frame-35
Frame-40
Figure 2. Tracking for duck video.
Copyright © 2011 SciRes. JSIP
Tracking of Non-Rigid Object in Complex Wavelet Domain109
Frame No. Spatial domain Real Wavelet Domain Complex Wavelet Domain
Frame-1
Frame-5
Frame-10
Frame-15
Frame-20
Frame-25
Frame-30
Frame-35
Frame-40
Figure 3. Tracking for stuart video.
Copyright © 2011 SciRes. JSIP
Tracking of Non-Rigid Object in Complex Wavelet Domain
Copyright © 2011 SciRes. JSIP
110
specified by the non-zero points of Mt + x. that is for a
given translation x, for each non-zero pixel p of Mt find
the location p + x of Dt + 1. The Kth largest of these m
values found so for is the partial directed Hausdorff dis-
tance hK(Mt + x, It + 1).
3.3. Updating the Model
The new model, Mt + 1 constructed from finding the
translation x of Mt with respect to the It + 1 by selectively
choosing the non zero pixels of It + 1 which falls within a
distance of
of non zero pixels of Mtx. This can be
obtained by dialating Mt by a disc radius
, shifting this
by x, and then computing the logical AND of It + 1 with
the dialated and translated model.
4. Experimental Results
The initial model, corresponding to the first frame of the
image sequence is taken. The user needs to take a rectan-
gle in the first frame containing this initial model. Then
image is processed to select the subset of the edge pixels
in this rectangle. This is done by assuming that the cam-
era will not move in first two consecutive frames and
using this fact to filter the first image frame based on the
second image frame. Those edge pixels in the user se-
lected rectangle that moved between the first and the
second frame are used as initial model. Thus the only
input from the user is the rectangle in the first frame of
the image sequence that contains the moving object.
The proposed method of tracking of non-rigid object is
applied on various videos and the results are compared
by computing the dual tree complex wavelet coefficient
to that by real and spatial domain coefficients. It has been
observed that:
In spatial domain, tracking is more accurate but
high execution time
In real wavelet domain, tracking is not as good as
in the spatial domain but the execution time is ap-
proximately 4 times less
Tracking in Complex wavelet domain is not only
fast as compared to real wavelet domain but also it
is comparable to spatial domain.
Also, we observed that in the spatial domain and real
wavelet domain the tracking work well up to the 15
frames and thereafter tracker is not able to track the ob-
ject accurately while as in complex wavelet domain it is
more accurate and efficient. Here the results are only
mentioned up to the 40 frames. The comparative results
for two representative videos are presented: in spatial
domain, real wavelet domain and dual tree complex
wavelet domain in Figure 2 and Figure 3.
5. Conclusions and Future Scope
The experiments are performed on various moving vid-
eos and non-rigid objects are tracked in spatial domain,
real wavelet domain and in complex wavelet domain.
Comparing the tracking results in three domains, we
found that the tracking in spatial domain is very accurate
because the operations are performed pixel-wise basis
but the computational time in spatial domain is high. It
has been observed that in real wavelet domain, the com-
putational complexity is low (approximately 4 times less)
but results of tracking are very poor. Tracking in com-
plex wavelet domain is fast as well as it produce good
results as compared to real wavelet domain. Here in
complex wavelet domain, both real as well as imaginary
components are taken which provide us better segmenta-
tion and tracking as compared to that of real wavelets.
The proposed tracking algorithm can be extended for
tracking of multiple objects in video sequences.
6. Acknowledgements
The authors are thankful to the Department of Science &
Technology, New Delhi and the University Grants
Commission (UGC), New Delhi, India for providing re-
search grant vide its grant nos. SF/FTP/ETA-023/2009
and 36-246/2008( SR) for research project.
REFERENCES
[1] A. K. Jain, “Fundamentals of Digital Image Processing,”
Prentice Hall of India Pvt. Ltd., New Delhi, 2001.
[2] S. Nigam and A. Khare, “Curvelet Transform Based Ob-
ject Tracking,” Proceedings of IEEE International Con-
ference on Computer and Communication Technologies,
Allahabad, 17-19 September 2010, pp. 230-235.
[3] M. Khare, T. Patnaik and A. Khare, “Dual Tree Complex
Wavelet Transform Based Video Object Tracking,” Com-
munications in Computer and Information Science, Vol.
101, No. 2, 2010, pp. 281-286.
[4] H. Goszczynska, “A Method for Densitometric Analysis
of Moving Object Tracking in Medical Images,” Machine
Graphics & Vision International Journal, Vol. 17, No. 1,
2008.
[5] Z. M. Budimlija, M. Leckpammer, D. Popiolek, F. Fogt,
M. Forenderick and R.Bieber, “Forensic Applications of
Capture Laser Microdissections: Use in DNA-Based Pa-
rentage Testing and Plateform Validation,” Croatian
Medical Journal, Vol. 46, No. 4, 2005, pp. 549-555.
[6] R. Anderson, N. Kingsbury and J. Fauqueur, “Coarse
Level Object Recognition Using Interlevel Products of
Complex Wavelets,” Proceedings of IEEE Conference on
Image Processing, Genoa, September 2005.
[7] N. G. Kingsbury and J. F. A. Magarey, “Wavelet Trans-
forms in Image Processing,” Proceedings of 1st European
Conference on Signal Analysis and Prediction, Prague,
24-27 June 1997, pp. 23-34.
[8] I. W. Selesnick, R. G. Baraniuk and N. G. Kingsbury,
“The Dual-Tree Complex Wavelet Transform,” IEEE
Tracking of Non-Rigid Object in Complex Wavelet Domain111
Signal Processing Magazine, Vol. 22, No. 6, 2005, pp.
123-151. doi:10.1109/MSP.2005.1550194
[9] F. C. A. Fernandes, R. L. C. Spaendonck and C. S. Burrus,
“A New Framework for Complex Wavelet Transform,”
IEEE Transactions on Signal Processing, Vol. 51, No. 7,
2003, pp. 1825-1837. doi:10.1109/TSP.2003.812841
[10] A. A. Bharath and J. Ng, “A Steerable Complex Wavelet
Construction and Its Applications to Image Denoising,”
IEEE Transactions on Image Processing, Vol. 14, No. 7,
2005, pp. 948-959. doi:10.1109/TIP.2005.849295
[11] N. G. Kingsbury, “Shift Invariant Properties of the
Dual-Tree Complex Wavelet Transform,” Proceedings of
IEEE Conference on Acoustics, Speech and Signal Proc-
essing, Phoenix, Vol. 3, 16-19 March 1999.
[12] N. Kingsbury, “Rotation-Invariant Local Feature Match-
ing with Complex Wavelets,” Proceedings of European
Conference on Signal Processing, Florence, 4-8 Septem-
ber 2006.
[13] A. Khare and U. S. Tiwary, “Daubechies Complex Wave-
let Transform Based Moving Object Tracking,” IEEE
Symposium on Computational Intelligence in Image and
Signal Processing, Honolulu, 1-5 April 2007, pp. 36-40.
[14] N. G. Kingsbury, “Image Processing with Complex
Wavelets,” Philosophical Transactions of Royal Society
London A, Special Issue for the Discussion Meeting on
“Wavelets: The Key to Intermittent Information?” Vol.
357, No. 1760, 1999, pp. 2543-2560.
[15] N. G. Kingsbury, “The Dual-Tree Complex Wavelet
Transform: A New Technique for Shift Invariance and
Directional Filters,” Proceedings of 8th IEEE DSP
Workshop, Bryce Canyon, August 1998.
[16] N. G. Kingsbury, “Complex Wavelets for Shift Invariant
Analysis and Filtering of Signals,” Applied and Computa-
tional Harmonic Analysis, Vol. 10, No. 3, 2001, pp.
234-253. doi:10.1006/acha.2000.0343
[17] N. G. Kingsbury, “The Dual-Tree Complex Wavelet
Transform: A New Efficient Tool for Image Restoration
and Enhancement,” The 9th European Signal Processing
Conference, Rhodes, September 1998.
[18] D. P. Huttenloche r, J. J. Noh and W. J. Rucklidge, “Track-
ing Non-Rigid Objects in Complex Scenes,” Proceedings
of 4th International Conference on Computer Vision,
Berlin, 11-14 May 1993, pp. 93-101.
Copyright © 2011 SciRes. JSIP