Sie sind auf Seite 1von 30

Journal of Mathematical Imaging and Vision 17: 207236, 2002

c 2002 Kluwer Academic Publishers. Manufactured in The Netherlands.


A Group-Theoretic Construction with Spatiotemporal Wavelets
for the Analysis of Rotational Motion
JEAN-PIERRE LEDUC
Department of Mathematics, University of Maryland, 1301 Mathematics Bldg, College Park,
MD 20742-4015, USA
jleduc@math.umd.edu
Abstract. This paper presents a group-theoretic approach for the analysis of rotational motion in image sequences.
This method relies on Lie algebras, Lie groups and Lie group representations to provide not only the continuous
wavelets but also the related tools of harmonic analysis. This approach can be referred to research works presented
in J.S. Byrnes et al. (Wavelets and their Applications, Kluwer Academic Publishers, 1994) who strongly inuenced
this topic. For the purpose of modeling motion transformations, this paper introduces the concepts of Lie algebras
and Lie groups as the actual mathematical foundations of all the observable kinematics embedded in spatio-temporal
signals. Rotational motion analysis focuses on the estimation of angular velocity and angular accelerations embedded
in image sequences. Rotational motion is usually carried on a trajectory, the complete problem at hand consists in
estimating not only the angular velocity and its temporal derivatives but also the position, the translational velocity
and its temporal derivatives along the carrier trajectory. The paper starts with the usual afne and Galilei groups
and proceeds by successive extensions and sections to the rotational group. The theory of group representations is
central to provide families of continuous wavelets, special functions, PDEs, ODEs and integral transforms as new
mathematical tools of motion analysis in image sequence to perform optimal and selective detection, estimations,
tracking, and reconstructions. This paper denes rotational wavelets and proposes a structured approach to perform
estimation and tracking in image sequences which ts to Kalman lters. Simulations on real digital image sequences
are also presented with tracking and estimation.
Keywords: Lie group theory, motion analysis, tracking, continuous wavelet transforms, image sequence analysis
1. Introduction
This paper addresses new applications of Lie group
representation theory for digital image processing.
More precisely, the purpose of this work is the anal-
ysis of rotational motion in digital image sequences.
To achieve this goal, a theoretical framework is built
on the following structural steps: construct the ap-
propriate Lie groups and Lie algebras that model the
physics and the kinematics [1, 3] of rotational mo-
tion, construct the related group representations in
the function spaces of signals, dene new families of
spatio-temporal continuous wavelets, develop the re-
lated harmonic analysis and design the related track-
ing algorithms. In this context, continuous wavelets
are dened from square-integrable unitary irreducible
representations of Lie groups. Following this vein, this
paper derives rotational wavelets froma Lie group con-
struction that starts with the multidimensional afne
group and the Galilei group. The multidimensional
afne group is a homogeneous extension of the uni-
dimensional afne group. The Galilei group is the
group of classical mechanics which deals with veloc-
ity transformation. The construction proceeds to the
denition of rotational wavelets as families of con-
tinuous wavelets indexed by the following kinemat-
ical parameters: scale, spatial and temporal position,
translational velocity, angular position and angular ve-
locity. In this construction, any family of continuous
wavelets stems from the action of a Lie group G onto
208 Leduc
operators in the functional space of the signals. This
action is dened in terms of a homomorphism i.e. a
one-to-one mapping from the group elements g G
to linear and invertible operators T
g
. These operators
are built as Unitary Irreducible Representations (UIRs)
of a Lie group fulll square integrability under ad-
ditional conditions. Square-integrable representations
dene continuous wavelet transforms. These continu-
ous wavelets are endowed with more properties than
those rooted in an usual cross-correlation function or
a matched lter [31]. The continuous wavelet trans-
form is an isometry dened on the Hilbert space of
observation H
o
= L
2
(R
n
. d x) whose range is a closed
subspace of the Hilbert space of the state parameters
H
s
= L
2
(G. dg). The set of kinematical parameters g
denes a Lie group G and characterizes the state of the
moving system. dg is an invariant Haar measure on G
to be dened. The continuous wavelet transformis also
optimal to many regards; it performs Minimum-Mean-
Square-Error (MMSE) motion estimation [31], edge
detection, image noise rejection, motion jitter smooth-
ing, frame decompositions, interpolations, predictions,
and motion-selective reconstructions [2, 26, 30].
The denition of rotational motion [3] depends upon
the position of the axis around which the motion takes
place. If the axis is the center of inertia of the object, the
motion refers to a spin. If the object revolves around
an external axis, the motion refers to an orbit. Fur-
thermore, the rotations are expressed through unitary
matrices of transformation namely
R(
i

i
) =

cos(
i

i
) sin(
i

i
)
sin(
i

i
) cos(
i

i
)

with x
2
1
x
2
2
-invariance for circular rotations and
R(
i

i
) =

cosh(
i

i
) sinh(
i

i
)
sinh(
i

i
) cosh(
i

i
)

with x
2
1
x
2
2
-invariance for hyperbolic rotations. This
paper will focus on both rotational aspects and illus-
trate the rst case within image sequences. Rotational
motion composes with translational motion. This pa-
per shows how estimation and tracking of translational
and rotational parameters can be dissociated fromeach
other.
The approach developed in this paper is based on
Lie group representations for signal processing as orig-
inally developed by Antoine et al. in [2] where the
prime focus was mathematical physics and the wavelets
on the Galilei, Poincar e and de Sitter groups and by
Caelli et al. in [11] where the interest was human vi-
sion modeled by the Lorentz group. Segman and Zeevi
in [38, 41, 42] stepped on Caelli approach and recog-
nized the relevance of the group-theoretic methods for
computer vision. Duval-Destin and Murenzi [12] have
also been focusing on perceptive aspects of the hu-
man visual system to design the rst spatio-temporal
continuous wavelets parametrized with speed. Let us
also remark that the concept of rotating wavelets has
been mentioned in a different context by O.A. Barut
[4]. Baruts wavelet functions are solutions of the Dirac
and wave equations. Let us also mention the early work
of Fleet and Jepson [13] who developed without any
group-theoretic support the rst orientation and veloc-
ity selective lters. This paper proceeds fromAntoines
original point of view to develop new extensions of
the Galilei group [33] based on observable kinemat-
ics in the functional spaces of the signals. The author
prefers to start from the Lie algebra. Indeed, Lie al-
gebras provide an easy and general way to dene the
kinematic and geometry. The construction proceeds to
the calculation of the Lie group and its representations
[28, 29, 31]. At that stage, all the concepts of harmonic
analysis [27] may be revisited and related to the kine-
matics at hand. As a major property of this approach,
all the relevant concepts can be analytically computed
and numerically estimated in the signal. The concepts
of interest in harmonic analysis are as follows: the anal-
ysis of uncertainty relations, ODEs, PDEs, special
functions and integral and wavelet transforms, exis-
tence of frames and orthogonal bases. Moreover, con-
tinuous wavelet transform on motion transformations
ts perfectly to important theory like Kalman ltering
[30, 31] and irregular sampling [27]. Moreover, con-
tinuous wavelet transformprovides a Wold decomposi-
tion of the signal considered as a wide-sense stationary
stochastic processes [15], maximumlikelihood estima-
tions, and links to stochastic calculus and Ito integrals
[27]. From this point of view, the group-theoretic ap-
proach presented in this paper differs fundamentally
from other techniques [32] that have been proposed so
far in the literature [13] such as those based on opti-
cal ow, pel-recursive, block matching and Bayesian
models and addresses the actual physics acting upon
the signal. Continuous wavelet transforms also provide
motion estimations that are robust not only against im-
age noise and blur but also against motion jitter [31].
Moreover, as a result of their spatio-temporal lter-
ing and interpolation properties, the wavelet technique
A Group-Theoretic Construction with Spatiotemporal Wavelets 209
can resolve temporary occlusions problem supply by
interpolation the objects at the occluded positions.
This paper presents a tracking algorithm based on a
Kalman lter which relies on the exponentiation of the
Lie algebra to generate the state equation of the sys-
tem. The Galilean and rotational wavelets perform se-
lective motion estimation and selective reconstruction
by means of a gradient-based technique which provides
an optimal search of parameters. The wavelet transform
eventually contributes as a motion-selective lter in the
observation equation.
The structure of the paper is as follows. Lie groups
and algebras for rotational motion are developed in a
constructive way along with the related wavelets, and
admissibility conditions. This constructive approach
proceeds by successive extensions and sections on al-
gebras and groups. This technique provides an use-
ful way to create many other variants of interest for
applications (for instance, deformations instead of ro-
tations, accelerations instead of velocity). The UIRs
are constructed using the Mackeys theory of induced
representations and imprimitivity [34] on semi-direct
products [5, 14], and the Kirillovs method of orbits. It
is shown how the characters of these new groups gen-
erate special functions similar to Bessel functions [8].
The special functions dene the kernel of integral trans-
forms similar to Hankel transforms and the sampling
of rotational motion which is related to the symmetries
of the system. The connection with the harmonic oscil-
lator in the group SU(2) and SL(2. C) is investigated
to relate this approach to some physical concepts. An
alternative Lie algebraic approach is nally proposed
for motion tracking and estimation in order to build
the motion trajectories. Simulations on real image se-
quences illustrate how to build the rotational wavelets
and how to perform the parameter estimation. To help
the interested reader, four appendixes sketch compu-
tational techniques that are classically related to the
group constructions.
2. Lie Group Representations and Continuous
Wavelet Transforms
This section contains the main purpose of this pa-
per: the construction of group representations for the
analysis of both translational and rotational motion.
These representation provides the related continuous
wavelets which support the motion estimation lter-
ing and tracking. To clarify the construction, a brief
description of Lie group representation theory is rst
provided with some useful references in this eld. In
each case of study, we adopt a similar structure which
proceeds with the construction of Lie groups of inter-
est, their Lie algebras, their group representations and
their associated wavelets. The group construction pro-
gresses by successive extensions and sections. Exten-
sions add new group parameters of interest. Sections
add a constraint on the parameters i.e. diminish the
dimension of the parameter space. For the shake of
clarity, our procedure will start from two well-known
groups: the afne group in multidimensional space, and
the extended Galilei group which introduces the trans-
lational velocity. The construction will proceed as fol-
lows. A section on the extended afne-Galilei group
provides the appropriate representations for image se-
quence processing. The next step consists of an new
extension followed by a section to add the angular po-
sition and velocity on the later version of the afne-
Galilei group. This procedure is fairly general; similar
extension-section procedures enable to incorporate
any additional temporal derivative of translational,
rotational and deformational motion.
2.1. Group Representations
Let S( x. t ) be the spatio-temporal signal under anal-
ysis and +( x. t ) the analyzing function in the Hilbert
space of observation H
o
= L
2
(R
n
R. d
n
xdt ). The
conditions of existence of such an analyzing func-
tion +, called a wavelet, are examined latter in this
section. The spatial dimension is n =2 when con-
tinuous wavelets are discretized to process image
sequences. The continuous wavelet transform W
+
is
dened [10, 20, 21, 26, 30, 31] as a linear map W
+
:
L
2
(R
n
R. d
n
xdt ) L
2
(G. dg) as follows
[W
+
S](g) = !+
g
. S) (1)
=

R
n
R

+
g
( x. t )S( x. t )d
n
x dt (2)
where !.. .) is an inner product which expresses the
correlation between two functions dened in H
o
and
the overbar symbol stands for the complex con-
jugate. Let us denote g G as an element of the Lie
group G that models the motion transformations. The
element g G carries all the motion parameters of in-
terest. Hence, g is precisely the state of the moving sys-
tem. Therefore, H
s
= L
2
(G. dg) is clearly the Hilbert
state space. The family of wavelets +
g
is generated as
an orbit with dense span in H
o
through the action of an
210 Leduc
operator T
g
such that +
g
( x. t ) = [T
g
+]( x. t ). The map
g . T
g
is a group representation i.e. a homomorphism
that maps the group elements g G to operators T
g
acting in H
o
. This mapping is such that T
g
1
g
2
= T
g
1
T
g
2
and T
e
= I
H
o . I
H
o denes the unit operator in H
o
and
e the identity element in G. The symbol stands for
the group composition in G. In the following, all the
group representations T
g
are constructed as UIRs and
calculated by the technique of induced representations
that invokes the Mackeys theory of induced represen-
tations [5, 34, 36, 37] and the Kirillovs method of
orbits [23]. The existence of continuous wavelet trans-
forms is further derived under the condition that the
UIRs be square-integrable. This condition of square-
integrability is equivalently known as the condition of
wavelet admissibility and as the condition for the linear
mapping dened in Eq. (1) to be invertible. This third
point of view is referred as the Cald eron reproducing
formula [39]. The inverse transform of Eq. (1) yields
the perfect reconstructions of the signal. This reads in
terms of operators I
H
o = W
1
+
W
+
and in functional
forms as
S(x) =

G
!T
g
+. S)(T
g
+)(x)dj
l
(g) with x = ( x. t )
(3)
The measure dj
l
(g) is the left-invariant Haar measure
on G. Let us remark that x = ( x. t ) may alternatively
be dened in the Fourier domain as k = (

k. ) where
the variables

k and are the dual of the translations
i.e. respectively the spatial and temporal frequencies.
In this brief introduction, we have seen how the linear
mapping dened in Eq. (1) denes an isometry between
the Hilbert space H
o
and a sub-space H
s
r
of the Hilbert
space H
s
. Let us notice that H
s
r
is a reproducing kernel
space [2]. In fact, if the signal is a wide-sense stationary
stochastic process, the isometry operator provides a
Wolddecompositionwhere the signal, anunpredictable
process, is decomposed into a regular process and an
orthogonal predictable process [15]. The predictable
belongs to H
s
r
and the former process belongs to the
complementary Hilbert space (H
s
r
)

.
Let us now restart with the construction of left-
regular UIRs. They read in H
o
[T
g
+](x) = (x. g)
12
+(g
1
x) with f
+
H
o
(4)
The function (x. g)
12
is usually a one-dimensional
representation that contains the characters and the nor-
malization factor that provides unitarity. These UIRs
[T
g
+](x) are usually calculated by the technique of in-
duced representations [23, 34]. The representation T
g
is
indeed induced from the representation L
K
of a closed
subgroup K of G in the Hilbert space H
o
. G is a locally
compact group and K G. Let us dene X as the ho-
mogeneous space of left cosets X =G`K ={gK. g
G} and denote j a quasi-invariant measure in X. The
group representations constructed with the technique
referenced here above are derived in a dual space
(Appendix 1). The dual space

X is a Fourier domain
known as the phase space made of patial and temporal
frequencies,

k and , as respective duals of the spatio-
temporal translations. When the Lie group admits a
non-trivial central extension [40], the phase space also
contains the associated dual parameter which is de-
noted m in the sequel and known as the dual of the
phase in the group. Eventually, Eq. (4) reads as the
following mapping
IND
G
K
L
K
: g [T(g)

+](k)
=
12
(g)(b. k)L
K

+(g
1
k) with
k = (

k. . m)

X and b = (

b. . ) X (5)
which denes an induced representation IND
G
K
of G
by L
K
in L
2
(

X. dj). In this equation, (g) =
j(g
1
k)
j(k)
is
the Radon-Nikodym derivative continuous on G X,
K is a stability sub-group (also called little group) of
G whose representation is L
K
, (b. k) are the char-
acters related to the dual group

X, g
1
k is the left
group action on X expressed by the co-adjoint action
of the group [23]. Therefore, L
2
(

X. dj) = H
o
. We
also know that H
o
on space and time corresponds to
the Fourier space. If there exists a system of imprim-
itivity dened as the triple (T
g
. X. j) and if K G
is a commutative stability subgroup, then for any UIR
L
K
, (IND
G
K
L
K
) is also UIR [34]. It turns out that all
semi-direct products G = NS where N is commu-
tative, have a system of imprimitivity [5]. The group
representations treated in this paper belong to this
category.
The additional condition for unitary irreducible rep-
resentations to fulll square-integrability provides a set
of important properties that are not rooted in standard
unitary irreducible representations [2]. For instance,
the condition of square-integrability the condition for
continuous wavelet admissibility, sets up the isometry
as dened above in Eq. (1) with a reproducing ker-
nel sub-space of H
s
r
= L
2
(G. dj
l
(g)). The projection
from L
2
[G. dj
l
(g)]) onto H
o
is the integral operator
with reproducing kernel K(g. g
/
) = !+
g
. +
g
/ ), that is,
A Group-Theoretic Construction with Spatiotemporal Wavelets 211
the autocorrelation function of +. For a given non-zero

+ H
o
, the condition reads
C(

+. ) =

G
[!T
g

+. )[
2
dj
l
(g) - H
o
(6)
where dj
l
(g) is the left invariant Haar measure on G
and C(

+. ) is some nite real constant. The functions

+ that fulll the condition are said to be admissible as


a mother wavelet. As a consequence of irreducibility,

+
g
denes an orbit with dense span in H
o
. The orbit
is dened as S
g
= {

+
g
= T
g

+ [ g G;

+.

+
g
H
o
}
i.e. a family of continuous wavelets indexed by the
motion parameters in g. Hence, square-integrability is
a well-dened property of a pair (T
g
. O), where T
g
is a representation and O is an orbit. The condition
of wavelet admissibility is calculated from Eq. 6 by
integrating over G with dg an invariant Haar measure
for G. This construction yields the state space as a sub-
space of L
2
(G. dg) with reproducing kernels [2], and
the map W

with all the properties of the isometry i.e.


the perfect reconstruction [2], the MMSE estimations
[30]. In Section 6, the Morlet wavelets will be chosen
as admissible rotational wavelets

+(

k. ) to perform
the numerical simulations. The subsequent sections are
devoted to the characterization of the parameter sets
g, the calculation of the UIRs and the conditions of
wavelet admissibility.
2.2. The Multidimensional Afne Group
The n-dimensional afne group is similar to the one-
dimensional case. In n-dimensional space, a rotation
matrix can be introduced to provide an additional
parameter for the estimation of spatial orientation.
This generalizes the one dimensional case of reec-
tion which is never used for the practical purpose of
motion analysis. As a result of the small amount of
parameters carried by this group, the afne group is
a best place to start and illustrate the technique. Let
us also remark that the rotations addressed in this pa-
per primarily consist of circular rotations. Hyperbolic
rotations also t to this approach and will be briey
discussed in this paper. To distinguish among the suc-
cessive Lie groups and related Lie algebras to be pre-
sented, we adopt the generic notation G
i
and G
i
re-
spectively for each group and algebra. The integer
i 1 determines the position the sequence of group
denitions.
The n-dimensional afne group, denoted here G
1
=
SIM(n), is an ordered 3-tuple of elements
g = {

b. a. R} (7)
where the parameters

b R
n
, a R

`{0} and R
SO(n) stand respectively for translation in space (the
Cartesian position), dilation (the scale) and rotation
in n-dimensional space (the angular orientation). This
group is a subgroup of GL(n 1. R) with the following
matrix representation
g =

aR

b
0 1

(8)
Let g = {

b. a. R} and g
/
= {

b
/
. a
/
. R
/
} be elements of
G
1
. Consequently, the associative law of composition
and the inverse element may be immediately identied
from matrix multiplication and inversion. These are as
follows
g g
/
= {

b aR

b
/
. aa
/
. RR
/
} (9)
and
g
1
= {a
1
R
1

b. a
1
. R
1
} (10)
Let us notice that the Lie group G
1
has the following
semi-direct group structure
G
1
= R
n
[R

SO(n)] (11)
where the symbols and stand respectively for the
direct and the semi-direct group products.
The left and right Haar invariant measures, respec-
tively dj
l
(g) and dj
r
(g) for G
1
are easily computed
through innitesimal variations of volumes in the mul-
tidimensional space of group parameters {g
i
}
n
i =1
and
are given for the left-invariant measure by the general
form
dj
l
(g) =

det

g g
/
g
/

1
g
/
=e
dg
1
dg
2
dg
n
(12)
where is the exterior product, and for G
1
, by
dj
l
(g) =
da d

b dm(R)
[a
n1
[
(13)
212 Leduc
The right-invariant measure is given in its general form
as
dj
r
(g) =

det

g
/
g
g
/

1
g
/
=e
dg
1
dg
2
dg
n
(14)
and for G
1
as
dj
r
(g) =
da d

b dm(R)
[a[
(15)
where dm(R) is the left and right Haar measure for
SO(n). Equations (12) and (14) are easily demonstrated
see in [5, 21].
Let us now derive the Lie algebra related to G
1
=
SIM(n). Let {J
i
}
n(n1)
2
i =1
be the standard generators for
so(n), the Lie algebra of SO(n). Then, the Lie alge-
bra G
1
of G
1
is generated by the elements {P
j
. Q. J
i
}
where j = 1. . . . . n and i = 1. . . . .
n(n1)
2
and the
commutator relations are given as follows
[P
k
. P
j
] =0. [Q. P
j
] = P
j
. [J
i
. P
j
] = c
k
i j
P
k
.
[Q. J
i
] =0 i. j. k (16)
where c
k
i j
is the antisymmetric tensor. It is easy to verify
that these generators forma Lie algebra i.e. all the com-
mutators fulll anti-symmetry and Jacobi relations. Let
us recall that the Lie algebra is the vector space tangent
to the group at identity and endowed by the Lie product.
G
1
is generated by the set of innitesimal generators
{P
j
. Q. J
i
}. The Lie product is the commutator prod-
uct [A. B] = AB BA where A and B are any two
generators. See [5, 17] to read all the major properties
of Lie algebras. It is clear that the Lie algebra has matrix
representations (Ados theorem [5]). We can now ob-
tain the generic group element g = {

b. a. R} G
1
and
the composition law by exponentiating a point Y G
1
.
It is known that the exponential map is a local diffeo-
morphism of G onto G. We may denote an element of
G by
X =

bP s Q r J = {

b. s. r} G. (17)
Thereafter, with g = E(X) and E the exponential map,
we have
E(X) = exp(

bP) exp(s Q) exp(r J) (18)


with a = e
s
R

`{0}, R = exp(r) SO(n).


The UIRs T
g
of G
1
in the Hilbert space L
2
(R
n
. d
n

k)
can be thoroughly calculated using the machinery de-
veloped in [4, 6, 23, 24, 37, 43]. This computation
requires to proceed through a set of new denitions.
The procedure is sketched in Appendix 1 for this case
which requires the smallest amount of computations.
The UIRs of G
1
= SIM(n) are eventually given in the
forms of Eq. (5) in the Fourier space L
2
(R
n
. d
n

k) as
[T
g

+](

k) = a
n
2
e
i

+(

k
/
) (19)
with

k = aR
1

k (20)
By construction (see Appendix 1, Eq. (99)), the param-
eter

k stands for the dual parameter of x in the duality
dened in Eq. (99). The parameter

k has to be interpre-
tated as a Fourier variable for the spatial frequency [2].
In this representation, it is easy to see that the normaliz-
ing factor a
n
2
provides unitarity, and the characters e
i

k
of R
n
dene the Fourier kernel. The change of variable
to

k
/
stems from the action of the group on

X = {

k}
dual of X = { x}, this is clearly g
1

X.
At this stage, the function

+(

k) L
2
(R
n
. d
n

k) can
be considered as a template to be used as a matched
lter but the function

+(

k) requires more to be admitted


as a continuous wavelet. To dene a matched lter for
a signal

S(

k) L
2
(R
n
. d
n

k), the function T


g

+(

k) has
just to fulll the following well-known condition

+ =

S to yield the equality in Cauchy-Schwarz inequality.


In this approach, we have dened the cross-correlation
function on the group as [C
+
S](g) a linear map, i.e. the
following function g
[C
+
S](g) = !T
g

+.

S) (21)
In order to obtain the continuous wavelet transform
from this cross-correlation function, the linear 21 map
has to become an isometry. According to Eq. (6), the
condition of square-integrability reads in this case

R
n

[!T
g

+.

+)[
2
da d
n

b
a
n1
- (22)
and, eventually after some computations and changes
of variables, continuous wavelets do exist under the
condition

+ L
2
(R
n
. d
n

k) such that

+(

)[
2
d
n

[
= C

- (23)
A Group-Theoretic Construction with Spatiotemporal Wavelets 213
and the representation T
g

+ is bounded for all g. The
variable

R
n
is still a Fourier variable.
Let us remark that the Weyl-Heisenberg group is an-
other important group which is not related to motion
analysis but can be treated in a similar way. The inter-
ested reader may nd the corresponding calculations
in [2, 20].
2.3. The Afne-Galilei Group
The Galilei group, referred here as G
2
= GAL(n),
is the group of physics that describes the motion of
rigid objects with constant velocity :, in the Euclidean
spacetime. An element, g of the Galilei group is dened
to be an ordered 4-tuple of elements, g = {

b. . :. R};
R SO(n);

b. : R
n
, R realized a subgroup of
SL(n 2. R) according to this matrix representation
g = {

b. . :. R} =

R :

b
0 1
0 0 1

SL(n 2. R).
(24)
Let us now combine the multidimensional afne group
G
1
= SIM(n) and the Galilei group G
2
= GAL(n)
to dene the afne-Galilei group G
3
=aff-GAL(n) as
an ordered 6-tuple of elements g = {

b. . :. a. a
0
. R}
with R SO(n);

b. : R
n
, R, a. a
0

R

`{0}. The associated Lie Algebra G


3
is a real vec-
tor space of dimension
n(n1)
2
2n 3 spanned by
the elements {P
i
. H. K
i
. Q. Q
0
. J
j
; i =1. . . . . n; j =
1. . . . .
n(n1)
2
} where the innitesimal generators are
denedinthe respective order of the parameters dened
in g. In this case, we have
exp(r) = (

0. 0;

0. 1. 1. R)
exp

i
b
i
P
i

= (

b. 0;

0. 1. 1. I )
exp

i
:
i
K
i

= (

0. 0; :. 1. 1. I ) (25)
exp( H) = (

0. ;

0. 1. 1. R)
exp(s Q) = (

0. 0;

0. a. 1. I )
exp(s
0
Q
0
) = (

0. 0; :. 1. a
0
. I )
where r =

i

i
J
i
so(n), e
s
= a R

`{0} is
the spatial dilation and e
s
0
= a
0
R

`{0} is the tem-


poral dilation. By matrix multiplication, we can check
that
g = {

b. . :. a. a
0
. R} E(X)
= exp(

bP) exp( H) exp( :K) exp(s Q)


exp(s
0
Q
0
) exp(r) (26)
where Y G
3
.
The attempt to construct continuous wavelets on
the afne-Galilei group G
3
=aff-Gal(n) fails in the
case n >1 as a result of the divergence of square-
integrability. Fortunately, the afne-Galilei group may
still be naturally extended with a central parameter .
By denition, this additional parameter is such that
its innitesimal generator M commutes with all the
other generators of the afne-Galilei algebra. This
extension, called a central extension, is not rooted
in the afne group. This new property will allow
our construction to solve the problem of wavelet ex-
istence in the n-dimensional afne-Galilei case. In
fact, a dimension-matching change of variables from
(a
0
. a. :) to (m
/
.
/
.

k
/
) can now be applied to yield the
condition of admissibility (see Theorem 1 below and
Appendix 1). The central parameter is a phase on
the group; its dual m can be interpretated exactly as
the uncertainty parameter between position and veloc-
ity sharing a whole theoretical similarity with the un-
certainty in Weyl-Heisenberg group (for instance the
construction of the harmonic oscillator).
The structure of the extended afne-Galilei group,
referred to as G
4
, is a semi-direct product of the form
G = NS with N = R
n2
, S = R
n
[R


SO(n)] where R

means R

`{0}. A typical element,


x N is of the form x = (.

b. ;

0; 1. 1. I
n
) and a
typical element s S is s = (0.

0. 0; :; a. a
0
. R).
The extended afne-Galilei algebra G
4
= {G
3
. M} is
dened by the innitesimal generators as follows
[M. Z] = 0 Z G
3
[K
i
. P
j
] =
i j
M [K. Z] = 0
[K
i
. P
j
] =
i j
M [J
i
. P
j
] = c
k
i j
P
k
[J
i
. K
j
] = c
k
i j
P
k
[H. K
i
] = P
i
[Q
0
. H] = H [Q
0
. K
i
] = K
i
[Q
0
. M] = M [Q. M] = 2M [Q. P
i
] = P
i
[Q. K
i
] = K
i
[J
i
. J
j
] = c
k
i j
J
k
(27)
Let us remark that the commutator [H. K
i
] = P
i
between time and position generate velocity in form

b = :t with an uncertainty generated by the com-


mutator [K
i
. P
j
] =
i j
M. This structure will be ex-
ploited below to generate the angular velocity from the
214 Leduc
angular position in whole similarity. All the other com-
mutators not mentioned in this collection are equal to
zero. Then, this set of generators denes a Lie algebra.
The afne-Galilei group G
3
consists of all elements
of the form E(X) where X G
3
. These elements are
g = {.

b. ; :; a. a
0
. R} with the parameter of cen-
tral extension. The lawof composition of the group can
be computed from the Lie algebra by exponentiating
the product of two elements X. X
/
G
3
i.e. computing
gg
/
= E(X)E(X
/
) G
3
and repeatedly applying the
Baker-Campbell-Hausdorff formula [44]. An example
of how to treat this kind of computation is provided in
[33]. A matrix representation for the extended afne-
Galilei group G
4
is still available and reads
g = {

b. . :. a. a
0
. R}
=

a
2
a
0
:
T a
a
0
R | :|
2 1
2a
0

0 aR :

b
0

0
T
a
0

0 0 0 1

(28)
and allows the computation of the associative law in
an easier way. The law of composition in the extended
afne-Galilei group G
4
is therefore given as
g g
/
=


a
2
a
0

/
:
T

:
/
2a
0

a
a
0
R

b
/

b aR

b
/
a
/
0
:
/
. a
0

/
;
: aa
1
0
R :
/
. aa
/
. aa
/
0
. RR
/

(29)
and the inverse element as
g
1
=

a
0
a
2
a
2
:

a
0
:
2

.
a
1
R
1
(

b a
0
:). a
1
0
;
a
0
a
1
R
1
:. a
1
. a
1
0
. R
1

(30)
The left and right-invariant Haar measure of G are
given by
dj
l
= a
n1
0
a
(2n3)
d dm(R) d

b d : d da da
0
(31)
and
dj
r
= a
1
0
a
1
d dm(R) d

b d : d da da
0
(32)
We see that the group is not unimodular since the left
and right-invariant measures are not equal. In fact, it is
easy to see that the dilation parameters stretch neigh-
borhoods of the identity differently on the left and on
the right which causes unimodularity to fail.
The UIRs of G
4
, that is the the extended afne-
Galilei group, arise in L
2
(R
n2
. d
n

kddm) using the


method of orbits and the induced representation tech-
nique in the extended spatio-temporal space of dual
variables

X = (m.

k. ). The variable stands for the


temporal frequency dual of the parameter . The UIRs
of G
4
read as
[T
g

+](m.

k. ) = a
n2
2
e
i (m

k)

+(m
/
.

k
/
.
/
)
(33)
with
m
/
=
a
2
a
0
m

k
/
= aR
1
(

k m :) (34)

/
= a
0


m| :|
2
2
:

for

+ L
2
(X. dmd

kd). Let us remark that the factor


a
n2
2
provides unitarity; it arises as a Radon-Nikodym
derivative which can be equivalently calculated from
either the co-adjunct action or the normalizing factor.
The factor e
i (m

k)
comes fromthe monomial rep-
resentation which induces the UIRs and plays the role
of characters. The variable m measures the Heisenberg
uncertainty between velocity and position. m is a non-
zeroreal constant whose effect canbe actuallyobserved
in image sequences [31].
Theorem 1. The UIRs of G
4
the extended afne-
Galilei group. in the space L
2
(R
n2
. dmd

kd)
which are given by the mapping g T
g
are square-
integrable. For a given non-zero

+ L
2
(R
n2
. dmd

k
d). the condition of square-integrability is equiva-
lent to the statement that the following two integrals be
nite

R
n2
[

+(m
/
.

k
/
.
/
)[
2
m[[

k[
2
2m]
2
dm d

k d - (35)
and

R
n2
[

+(m.

k. )[
2
dm d

k d
(m)
n
[[

k[
2
2m]
3
- (36)
A Group-Theoretic Construction with Spatiotemporal Wavelets 215
This theorem is proved in Appendix 2. Eventually, a
section is performed in the dual space

X = (m. .

k)
on m xing uncertainty and on the parameter space on
a
0
xing the time scale as follows
m = m
0
,= 0 a
0
= 1 (37)
This section addresses the applications in image se-
quence processing and fullls the condition for wavelet
admissibility with a dimensional-matching change of
variables from (a. :) to (
/
.

k
/
). Let us now apply on
G
4
the section referred in Eq. (37) to dene a group G
5
that provides representations in L
2
(R
n1
. d
n

kd) with
n = 2 for image sequences. These UIRs are

T
g

+
m
0

k. ) = a
n
2
e
i (

k)

+(

k
/
.
/
) (38)
with

k
/
= aR
1
(

k m
0
:)
(39)

/
=


m
0
| :|
2
2
:

Theorem 2. The UIRs of G


5
(the group obtained
by sectioning the extended afne-Galilei group) in
the space L
2
(R
n1
. d
n

kd) which are given by the


mapping g T
g
are square-integrable. For a given
non-zero

+ L
2
(R
n1
. d

kd). the condition of


square-integrability is equivalent to the statement that
the following two integrals be nite
c
+
=

R
n1


+
m
0
(

k. )

2
I
m
0
(

k. ) d

k d - (40)
where
I
m
0
(

k. ) =

R
n1


+
m
0
(

k
/
.
/
)

[[

k[
2
2m
0
(
/
)]
n1
[

k[
2n
m
n1
0

k
/
d
/
- (41)
This demonstration theorem is sketched at the end
of Appendix 2 as a variant of Theorem 1. At this
stage, a comprehensive construction of Lie groups,
Lie algebras and associated continuous wavelets has
been completed for the description of translational mo-
tion in image sequences. Appendix 3 presents an al-
ternative way of computing the wavelet admissibility
through Calder on reproducing formula. With the tech-
nique developed so far and Appendixes 13, we are
fully equipped to proceed to the main construction of
this section, i.e., Lie groups that includes rotational
motion.
2.4. Groups for Rotational
and Translational Motion
This section proceeds one step further to extend the
afne-Galilei group G
5
to take into account rotational
motion. In this work, we focus on image process-
ing. Therefore, in the sequel, we set n =2. The trans-
formation of rotational motion is described as fol-
lows in R
2
R applied on a spatial function f [ x]:
R
2
R. We also suppose that the function f ( x. t ) is
C

is compactly supported in the spacetime: f ( x. t )


L
2
(R
n
R. d xdt ). Let us apply the rotational transfor-
mation to the function f ( x. t ) that is characterized by
a spin around the origin at uniform angular velocity

1
R. Potential angular accelerations
n
R; n > 2
can also be taken into account. The function f ( x) trans-
forms as

f ( x. t ) = f [R() x. t ] (42)
R may be expanded into a temporal Taylor series
as
(t ) =
0

1
t
k=

k=2

k
k!
t
k
(43)
where
0
is the orientation in space. The construc-
tion of the Lie group, to be referred to as G
6
, for
rotational motion with angular position
0
and uni-
form angular velocity
1
requires a trick to provide
the correct law of composition with the property
to be associative, to t within rotations as in ma-
trix RSO(2) and to provide a physical law of
composition for the angular coordinates in the form

0

/
0

1

/
which is similar to the translational
version without scaling. The construction proceeds
from the afne-Galilei group G
3
=aff-GAL(n) in two
steps: a group extension with new parameters
0
and
1
different from R(A) followed by a section
that re-establishes the rotational nature of the new
parameters.
The rotational extensionof the afne-Galilei groupis
denoted G
6
=aff-ROT(2) is dened by ordered 7-tuple
216 Leduc
of elements, g, and the matrix representation
g = {

b. ; :.
0
.
1
. R[A]. a}
=

1

0
T

1

0

0 aR[A] :

b
0 0 1
0 0 0 1

(SL(5). R). (44)


where A [0. 2), and
0
.
1
R. The associated Lie
algebra is given by
G
6
= {P. H; K; J
0
. J
1
. J. S} (45)
where the innitesimal generators are dened in the
respective order of the elements in g. The additional
commutators of the Lie algebra G
6
fulll
[J
1
. H] = J
0
[J
0
. J
1
] = 0 [H. J
0
] = 0
(46)
[P. H] =0 [J
0
. P
1
] = P
2
[J
0
. P
2
] = P
1
where the commutator [J
1
. H] = J
0
acts as the com-
mutator [K. H] = P to generate the velocity as in the
Galilei group. The angular velocity is related to the an-
gular position according to the usual law
0
=
1
. Ex-
ponentiation from this Lie algebra determines that the
elements in G
6
=aff-ROT(2) compose in the following
manner
g g
/
={

b aR[A]

b
/
:
/
.
/
; : aR[A] :
/
;

/
0

1

/
.
1

/
1
. R[AA
/
]. aa
/
}.
(47)
The inverse element reads
g
1
= {a
1
R[A](

b a
0
:). ; a
1
R[A] :;

0
(1
1
).
1
. R[A]. a
1
} (48)
The group structure is a semi-direct product NS with
N = R
n
Rand S = RRSO(2) R

. After this
extension, a section on the group parameters space
A = [
0

1
]mod 2 (49)
leads to the appropriate group of rotational motion. Let
us remark that this construction clearly differs from a
one-parameter subgroup since
0
,
1
and have their
own Lie generator. The one-parameter construction
would consist in generating a trajectory as function of
a dummy variable t out of one single parameter
0
.
A central extension can be dened on the afne-
rotational group G
6
=aff-ROT(2) to generate an ex-
tended afne-rotational group, referred to as G
7
in the
sequence. This extension is totally similar to that from
G
3
=aff-GAL(n) to G
4
, the central extension of aff-
GAL(n). The Lie algebra G
7
is given by elements of the
form G
7
= {M. P. H; K; J
0
. J
1
. J. S}, and the group
elements G
7
compose as
g g
/
=


a
2
a
0

/
:
T

:
/
2a
0

a
a
0
R[
0

1
]

b
/

b aR[
0

1
]

b
/
:
/
.
/
; : aR[
0

1
] :
/
;

/
0

/
.
1

/
1
. R[
0

1

/
0

/
1

/
]. aa
/

.
(50)
The left and right-invariant Haar measures on G
7
are
given by
dj
l
= a
n1
0
a
(2n3)
d dm(
0
) dm(
1
) d

b d : d da
(51)
and
dj
r
= a
1
0
a
1
d dm(
0
) dm(
1
) d

b d : d da (52)
The group is not unimodular.
If we proceed with sections on the extended afne-
rotational group G
7
that are similar to those described
in Eqs. (37) and (49), we get the group G
8
with UIRs
in L
2
(R
2
R. d

kd) expressed as

T(g)

+
m
0

k. ) = a
n2
e
i [R
1
[
1
]

k]

+
m
0
(

k
/
.
/
)
n = 2 for image sequence (53)
where

k
/
and
/
represent the same group actions
as those computed in Eq. (39). As rotations are of
concern, it is natural to rewrite these IURs in po-
lar coordinates. In this transformation, we denote
O =

1
, n =
1
,

k = (k. ),

b = (r. ), =
and

2
= n , successively. We further let : =

0 and m
0
= 0, to get
[T(g)

+](k. . O) = a
1
e
i O(

2
)
e
i [Okr sin ]


+(ak.
0
.
1
O) (54)
with the polar decomposition

b = (r. ),

k = (k. ).
The character e
i [R
1
(
1
)

k]
of the UIRs in Eq. (53)
introduces a special function derived by integration on
A Group-Theoretic Construction with Spatiotemporal Wavelets 217
. In this case, we reach the following form
J
O
(k) =
1
2

2
0
e
i [Ouk sin u]
du (55)
which is not a Bessel function except for O Z. In-
deed, the Bessel function has an analytic extension for
n C [19, 22, 25, 45] given by
Z
n
(z) =
1
2

e
i (nz sin )
d

sin n


0
e
(nt z sinh t )
dt.
For n Z, J
n
(z) coincides to Z
n
. The complexication
of u i y in Eq. (55) gives rise to hyperbolic rotations
and the special function
H
O
(k) =
i
2


0
e
(Oyk sinh y)
dy (56)
A theorem of additivity for these special functions can
be deduced from the composition of the translations.
In this case, it reads after some computations

J
[O
1
l]
(kr
1
)J
[lO
2
]
(kr
2
)dl = J
[O
1
O
2
]
[k(r
1
r
2
)]
(57)
The special functions dened in Eq. (55) further lead to
integral transforms where the special function is taken
as kernel; the rst example is
[H
O
f ](k) =


0
f (r) J
O
(k r) rdr n = 2 (58)
Further properties involving those special functions
and integral transforms will be considered in Section 3.
Let us rst formally suppose that

+[aR
1
(
1
)

k. ] =

+[aR(
1
)

k. ] =
1

1
[F
O
+](

k). By
similar arguments as those developed for the proof of
Theorem 2 in Appendix 2, Eq. (127), we reduce the
condition of square-integrability to showing that there
is some non-zero + L
2
(R T. d
2
xdt ) such that

R
4

F
O
+
m
0

k)

2
da d : dm(
1
)
a
3
- (59)
Let us remark that the integration on dm(
1
) requires
that the limit of the integrand be zero when
1
0.
This limit corresponds to O . This requires that
F
O
+
m
0
= O([
1
[
1c
) be of compact support on
1
to
avoid the integration in Eq. (59) to diverge. This state-
ment remains a formal wish. The condition to achieve
convergence [39] is reached if we consider the con-
vergence of an (inner) product of F
O
with a compact
supported function. This leads to a weak topology of
convergence and reconstruction and therefore to a weak
denition of continuous wavelet transforms.
Theorem 3. The UIRs of G
8
in the space
L
2
(R
21
. d
2

kd) which are given by the mapping g


T
g
are square-integrable. The condition of square-
integrability requires that the following integral be
nite
c
+
=

R
n1

F
O
+
m
0

k)

2
I
m
0
(

k. O)d

k dO -
(60)
where
I
m
0
(

k. O) =

R
n1

F
O
+
m
0

k
/
)

[[

k[
2
2m
0
(OO
/
]
n1
[

k[
2n
m
n1
0

k
/
dO
/
- (61)
This demonstration theorem is as sketched in
Appendix 1 as a variant of Theorem 2. A similar dis-
cussion can be conducted for hyperbolic rotations and
the related wavelets. In this case, the convergence of
the integral is neatly obtained through the negative real
exponential function.
2.5. Further Extensions to New Kinematics
Two avenues for further extensions and sections on the
afne-Galilei group have been opened in this paper.
These potentialities demonstrate the general founda-
tion of the approach developed in this paper.
The rst direction concerns all the variant of exten-
sions with the triple [R(A),
0
,
1
; R SO(n)] dened
in Eq. (44) where the matrix R is a matrix of spatial
transformation,
0
and
1
are destined to correspond
to the 0th and 1st order coefcient of the transforma-
tion after the appropriate section. The matrix R which
has been considered as the group SO(2) can be re-
placed by another group of spatial transformation. For
instance, let us rst mention the complexication of A
in R(A) SO(2) which leads to hyperbolic rotations
218 Leduc
as dened in Section 1. Asecond example is to consider
a deformation matrix of the form

A 0
0 A

(62)
and the section A = a
0
a
1
which describes a scaling
action with temporal evolution. A third example is to
take the matrix of shear of the form

1 S
0 1

(63)
and the section S = s
0
s
1
. Each case of study leads to
its own group representation, special functions related
to the characters, a specic harmonic analysis with
PDEs and ODEs, and a complete discussion about
the existence of continuous wavelets suited for image
sequence processing.
The second path for extensions is to add columns in
the matrix of Eq. (44) to introduce further orders of time
derivative i.e. the angular accelerations
m
, associated
with the translational acceleration
m
, and the temporal
translations of higher order
m
; m > 1. For instance,
the rst acceleration will yield the following matrix
representation
g = {

b. .
2
; :.
1
.
0
.
1
. R[A]. a. a
0
}
=

1

0
T

2

1

0

0 aR[A]
2
:

b
0 0 1 0
0 0 2 1
2
0 0 0 0 1

SL(n 4). R).


(64)
Rigorously, according to the extension technique that
has been developed, the parameter
n
; n > 1 should
be thought rst as an independent parameter which
is sectioned afterwards as =
n
where is the usual
temporal translation. Once again, each additional or-
der of time derivative leads to its own group represen-
tation, special functions related to the characters, the
associated harmonic analysis for PDEs and ODEs,
and, again, a complete discussion about the existence
of continuous wavelets for image sequence process-
ing. Moreover, the rst and the second paths can be
combined one with the other in a similar discussion.
Our technique of construction reaches in this manner a
complete description of all the observable kinematics
in the Euclidean spacetime.
3. Special Functions and Symmetries
To get more insight on the special functions related to
motion, let us revisit the Fourier transform of a func-
tion

f ( x. t ), x R
n
rotating with the angular velocity

1
in a two-dimensional plane. If we rst consider the
group representation in Eq. (53) with : =

0, a = 1,
a
0
= 1,
0
= 0 and extend the functions to distribu-
tions to take the + function as a Dirac measure along a
trajectory from time t = 0 to t = 1, the integration on
the character leads to the special function in Eq. (55).
Let us further assume without loss of generality
that the functions are separable in terms of magnitude
and orientation i.e. f ( x. t ) = f
r
(r. t ) f
s
( x
u
. t ) where we
consider the polar coordinates inR
n
r = [ x[ is the mag-
nitude in R
n
and x
u
S
n1
belongs to the unit sphere
in R
n
, S
n1
= { x
u
R
n
: [ x
u
[ = 1}. We then observe
that

f (

k. ) =

R
n
T
dt d
n
x f [R(
1
t ) x]e
i (

k xt )
=

R
n
T
dt d
n
x
/
f ( x
/
) e
i [R(
1
t )

k x
/
t ]

f (k.

k
u
. ) =

S
n1
T
dr d( x
/
u
) dt r
n1
f
r
(r)
f
s
( x
/
u
) e
i [kr R(
1
t )

k
u
x
/
u
t ]
=

S
n1
d( x
/
u
) f
s
( x
/
u
)


0
dr r
n1
f
r
(r)

2
0
du e
i

1
)ukr R(u)

k
u
x
/
u

f (k.

k
u
. O) =
1

S
n1
d( x
/
u
) f
s
( x
/
u
)


0
dr r
n1
f
r
(r)J
O
(kr

k
u
x
/
u
)
=
1

S
n1
d( x
/
u
) f
s
( x
/
u
)H
O
[ f
r
. k

k
u
x
/
u
]
=
1

1
F
O
[ f
s
. f
r
; k

k
u
] (65)
where d is the Lebesgue measure on the sphere S
n1
.
In this calculation, we have successively applied a
change of variable x
/
= R(
1
t ) x i.e. x = R(
1
t )
T
x
/
, a
change from Cartesian R
n
T to general polar coordi-
nates S
n1
Tin the space domain. The spatial Fourier
domain is changed to polar coordinates with k R

A Group-Theoretic Construction with Spatiotemporal Wavelets 219


and

k
u
R
n
, [

k
u
[ = 1. We also let O =

1
. Then, an in-
tegration is taken on u that generates special functions
of real index J
O
. The integration on variable r modied
the Fourier transform into an integral transform H
O
. In
turn, this transform becomes the kernel of an integra-
tion on the sphere S
n1
. Clearly, f ( x. t ) C

c
(R
n
T)
is compactly supported by assumption and we have
F
O
[ f
s
. f
r
;

k]

2
n2
I(n2)
sup
x
/
u
S
n1
f
s
( x
/
u
)H
O
[ f
r
.

k x
/
u
] - (66)
where
2
n2
I(n2)
= j(S
n1
). The construction of ODEs
and PDEs as equations of the wavelet motion is an
additional topic related to the group representation.
Reference [27] presents one way to derive and com-
pute those equations that play an important role for
motion tracking.
The derivation of new motion-based special func-
tions as substitute of the Fourier kernel requires to
revisit the sampling theorem. It is clear indeed that
the sampling of rotational motion [22] requires to state
some new remarks related to symmetries. Without any
particular symmetry, the critical sampling is obtained
at = i.e.
1
= rad/image. At this rate the
object is orbiting at one half-round turn per consec-
utive image. If the object is shaped with one or several
axes of symmetry, then the nite group of symmetries
[43] brings its characters on the functions f ( x. t ), J
O
and H
O
and involutions that reduce the periodicity at
sampling. One may easily verify that the rst com-
ponent of the expansion of the spectrum along the
-axis is located at 2m
1
with m axes of symmetry.
In the case of a square, the integration in Eq. (65) is
performed from

4
to

4
. The critical sampling cor-
responds to an angular velocity of
1
=

4
rad/image.
This reaches the Shannon sampling bound. Similarly,
a spinning 2m-gons reach the bound at

2m
rad/image.
If m tends to , the polygon tends to a circle whose
shape is rotational invariant. Hence, any angular ve-
locity is potentially admissible and motion becomes
ambiguous.
4. Rotational Motion and Harmonic Oscillator
It is instructive to examine how the rotational mo-
tion is intimately connected with the harmonic oscilla-
tor [46, 47], the groups SU(2) and SL(2. C). In fact,
the existence of an harmonic oscillator may be evi-
denced in the extended Galilei group when consid-
ering its Lie algebra restricted to position, velocity
and central extension. Simple analytic computations
lead then to an analysis which is step by step similar
to the classic theory developed for the harmonic os-
cillator in the context of the Weyl-Heisenberg group
[7]. Numerous Lie groups lead to harmonic oscillators
[46, 47]. In the particular case of performing rotational
motion analysis with angular parameters dened on
the real line R, we show from the Hamiltonian that
SU(2) is appropriate to describe the spinning or or-
biting in a plane around the orthogonal axis. SU(2)
is locally isomorphic to the group SO(3) of the ro-
tations in three-dimensional space (Euler angles not
motion). In fact, both algebras su(2) and so(3) are iso-
morphic but the group SO(3) is the quotient of SU(2)
by its center Z
2
= {I. I } i.e. SO(3) SU(2)Z
2
.
SU(2) is the group of the 2 2 unimodular, unitary
matrices i.e.
g =

a b

b a

with [a[
2
[b[
2
= 1 (67)
SU(2) is a three-parameter group, a UIRis given below
in Eq. (75).
Let us consider the Hamiltonian of the one-
dimensional harmonic oscillator [3]
H =
p
2
2m

1
2
m
2
1
q
2
(68)
where
1
=

km, k >0 is usually known as the


spring constant, m is the mass of the system, and p
and q are momentum and position respectively. This
Hamiltonian gives rise to the equation of motion i.e.
the ODE for the position of the system, q, which
reads
q
2
1
q = 0 (69)
In the phase space of coordinates (q. p), classi-
cal mechanics shows how the time evolution op-
erator can be obtained from the Poisson brackets
[3]. The standard theory [3] in n dimensions fur-
nishes the correspondence between smooth functions
A( p. q) of p and q and the differential operator as
follows:
A O
A
=
n

i =1

A
p
i

q
i

A
q
i

p
i

(70)
220 Leduc
In this case, the evolution operator O
H
related to the
Hamiltonian H in Eq. (68) reads
O = i 2
1
J
2
with J
2
= i

x

p
p

x

2.
(71)
Let us restate the dual space as a two-dimensional com-
plex plane (z
1
. z
2
) = (q. p) and dene the operators
X
i j
= z
i
(

z
i
) with i. j (1. 2). Hence, the operators
dened by the following relations
J
1
= (X
12
X
21
)2; J
2
= (X
12
X
21
)2;
J
3
= (X
11
X
22
)2 (72)
i.e. the standard commutation relations read [J
i
J
j
] =
c
i j k
J
k
, with c
i j k
the anti-symmetric tensor. The above
Lie algebra generated by J
i
; i = 1. 2. 3 is the algebra
of so(3). However, since SO(3) and SU(2) are locally
isomorphic about the identity, we immediately see that
J
i
; i = 1. 2. 3 is the Lie algebra of SU(2). This is easily
seen if the J
i
are re-written in terms of matrices
J
1
=
i
2

0 1
1 0

. J
2
=
1
2

0 1
1 0

. J
3
=
i
2

1 0
0 1

(73)
Let us consider the one-parameter subgroup of
SU(2) generated by the ow of the innitesimal gen-
erator J
3
to represent the rotation about the z-axis or-
thogonal to the sensor plane xy. The exponential map
yields the sub-group G
[3.su(2)]
SU(2) generated by
J
3
as the set of matrices
exp[i
1
J
3
] =

e
i
1
0
0 e
i
1

(74)
Hence, the basic harmonic oscillator of interest is
a group G
[3.SU(2)]
generated by J
3
as being the set
{e
i t J
3
[ - t - }.
The next problem is to nd the appropriate UIRs of
SU(2) in L
2
(CR. dj). The group SU(2) is a subset
of SL(2. C) which is the group of the 2 2 matrices
g =

SL(2. C)
i.e. with det =1 and . . . C. The UIRs of
SL(2. C) are indexed by pairs (k. i :) with k Z and
: R. The representations in L
2
(C R. dj) is given
by [24]
[T
g
+](z. t ) = [z [
2i :

z
[z [

k
+

z
z
. t

(75)
with z C. SU(2) is part of the bilinear transforma-
tionthat describes the stereographic projectionof three-
dimensional rotations in the two-dimensional plane. If
we let = e
j
, = e
j
, = = 0, we effec-
tively yield the UIRs of G
[3.SU(2)]
in L
2
(C R. dj).
As a compact group, L
2
[SU(2)] possesses a countable
orthonormal basis according to Peter-Weyl theorem
[9] made of Bessel functions generated by the char-
acters. Let us remark that the bilinear transformations
in Eq. (75) can be built as self-maps of the com-
plex disk (i.e. a M obius transformation with a rota-
tion). This link leads to considering Bergman spaces
of complex functions in a spacetime approach. We al-
ready know the existence of continuous wavelets in
these spaces. The harmonic oscillator generalizes in
n-dimensions, n > 2, by taking m n independent
copies of the preceding version. For example, let us
consider n = 3 involving three orthogonal axes of ro-
tation: G
t
= SU(2) SU(2) SU(2). The function
space of these oscillators is the tensor product of the
original space
V
n
=
n

i = 1
V.
The resulting group is still compact and corresponds to
the direct group product
G
t
=
n

i = 1
SU(2).
In fact, this mathematical structure corresponds to
the usual addition of angular momenta. The group
representation is the product of the characters of
the different SU(2)s. Moreover, these multidimen-
sional isotropic harmonic oscillators seem to be rich
in structure since there exist many distinct complete
admissible sets of commuting/non-commuting oper-
ators whose eigenvectors span the tensor product
space.
A Group-Theoretic Construction with Spatiotemporal Wavelets 221
5. Tracking Rotational Motion
It has been shown in [30, 31] that the Galilean wavelets
perform MMSE motion estimations. This property es-
sential to track the velocity along a trajectory. The same
properties holdfor acceleratedwavelets onacceleration
parameters under the condition that the velocity param-
eter be xed i.e. be previously estimated [27]. In this
approach, translational motion estimation starts with
the spatio-temporal location (i.e. the spacetime trans-
lation parameters in the afne wavelets) and proceeds
with the velocity (Galilean wavelets). On enough small
spatio-temporal neighborhoods, motion is assumed to
be linear and Galilean wavelets are always appropriate.
The estimation process proceeds with the estimation of
the acceleration parameters (accelerated wavelets) on
a larger span of scene. This leads to the construction of
optimal trajectories that can be buit with algorithms
derived from dynamic programming and Bellmans
principle of optimality (Viterbi algorithms). The as-
pects of this construction dealing with Lagrangians
and dynamic programming have already been exam-
ined in [31]. In this section, we develop the specic
part that performs the separate tracking of the rotational
motion.
The technique developed in this section is based on
Kalman lters, an alternative point of view of Lie alge-
bra, and rotational continuous wavelet transforms. The
Kalman lter is used as a recursive minimum variance
estimator which introduces an alternative approach to
the formulation and solution of linear MMSE estima-
tion problems. This alternative is particularly useful to
deal with non-stationary processes and to approach the
estimation of dynamic motion problems. We adopt a
structure of Lie algebras in forms of a set of differen-
tial operators that describe the motion. These opera-
tors are derived from the equation of motion i.e. the
equation that relates the kinematical parameters of po-
sition, velocity and accelerations. In this application,
the motion equation superposes rotational motion on a
translational carrier trajectory.
The Kalman lter is based on two equations, the state
equation and the observation equation, along with an
optimization algorithm. The state equation is a linear
predictor which provides a prediction of the state of
the sytem at time t from the state of the sytem at
time t . The state q of the system is composed of all the
relevant motion parameters that exhaustively explain
the local behavior on the trajectory. These parameters
under study are known to be position, velocity, and ac-
celerations for translational and rotational motion. As
a matter of fact, the innitesimal local evolution of the
system is linear and described by
d q
dt
. Expanding the
total derivative into its partial derivatives associated to
the equation of motion leads to a set of differential op-
erators whose closure denes a Lie algebra. The state
equation can be seen as a one-parameter sub-group cal-
culated from the Lie algebra which provides the lo-
cal transformation in terms of an evolution operator.
At each step of this prediction, an optimization takes
place after the prediction to optimally estimate the ac-
tual parameters. The difference between predicted and
estimated parameters provides the innovation of the
system. This step is implemented with either gradient
or dynamic programming algorithms. The observation
equation may have two implementations. It is either
a convolutional ltering of the image sequence with
continuous wavelets tuned to the optimum parame-
ters or a frame-based reconstruction technique based
on the inverse wavelet transform (see Appendix 3).
The former technique selectively captures and extracts
the moving pattern of interest moving along its tra-
jectory in form of an image sequence. The later tech-
nique reconstructs a motion compensated sequence
where motion has been frozen or cancelled. or a recon-
struction out of the inverse wavelet transform. Partial
selective reconstruction is achieved from the wavelet
transform that performs a motion-selective band-pass
ltering on the entire scene. The wavelet family is
generated by the action of a group representation
that describes the global motion transformation under
estimation.
5.1. State Equation and Lie Algebras
Let us consider the innitesimal transformation of the
Lie algebra given by
q
/
=
d q
dt
= A q (76)
where A is the matrix representation of a Lie algebra
of differential operators and q is the state vector of
the system. Some examples of matrix A are computed
below in this section and in Appendix 4. One can also
see that the state equation that updates the state of the
Kalman lter is a one-parameter subgroup obtained
by exponentiation of operators in the Lie algebra. The
algebra is of nite dimension, the operator O
t
of the
state space equation is the matrix A in Eq. (76) and is
222 Leduc
dened as
q(t ) = e
O
t
q(t ) = e
A
T
q(t ) (77)
The goal is now to build the Lie algebras, dene the
states and construct the representation A from given
equation of motion.
The trajectories can be described in terms of Taylor
expansions. The coefcients of the expansions are the
parameters of motion. Let x(t ) be the position at
time t , the trajectory can be expressed as
x(t ) =

k=0

k
k!

d
k
x()
d
k

=t
(78)
In this method, we also assume to know how many
orders in the expansion are relevant to describe the lo-
cal motion transformation. Let us remark that the mo-
tion parameter estimation performed by the continuous
wavelets enables the determination of how many rele-
vant parameters exhaust the expansion [27]. Newtons
equation of classical dynamics limits the expansion to
the rst-order of acceleration
1
. This is no longer true
for motion taking place within two-dimensional signal
generated from the projection onto the sensor plane.
The expansion may need higher orders of acceleration.
These aspects are treated in [28]. Hence, the equation
of motion captured in the camera can be stated as

n
(t ) = F[x(t ). :(t ).
1
(t ). . . . .
n1
(t ); t ] (79)
where :(t ) =
dx
dt
,
1
(t ) =
d:
dt
, and
n1
(t ) =
d
n2
dt
.
From Eqs. (77) and (78), it becomes possible to char-
acterize the differential operator O as an innitesimal
generator that produces the evolution operator
e
O
t
=

k=0

k
k!

O
k
t
(80)
and the operator O
t
is clearly a time differential op-
erator whose expansion into the partial derivatives is
calculated from the equation of motion (79)
O
t
=
d
dt
= :(t )

x

1
(t )

:

n
(t )

n1


t
(81)
It is now possible to break O
t
into a sum of elemen-
tary differential operators and to derive a Lie algebra.
Since any Lie algebra is closed under the Lie multi-
plication, new generators arise to yield the algebraic
closure.
By Ados theorem [5], Lie algebras have matrix rep-
resentations in gl(n. R) meaning that it exists a homo-
morphism between any Lie algebra and the matrices
A of n n dimension. The matrix A of the evo-
lution operator is a representation of the differen-
tial operator O
t
dened in the algebra gl(n. R). Let
us now consider two examples of equation of mo-
tion where a harmonic oscillator is superposed on
a carrier trajectory F
1
[x(t ). :(t ); t ] and F
2
[x(t ). :(t ).

1
(t ); t ]
F
1
:
1
(t ) = a
1
x(t ) a
2
:(t ) b
1
cos(
1
t ) (82)
F
2
:
2
(t ) = a
1
x(t ) a
2
:(t ) a
3

1
(t )
b
1
cos(
1
t ) b
2
cos

2
2
t
2

(83)
These examples are thoroughly treated in Appendix 4.
As an illustration of how to proceed, the appendix
presents the corresponding operators O
t
, the innites-
imal Lie generators X
i j
involved in O
t
. The closure
of the Lie algebra of operator O
t
is calculated under
the Lie multiplication. We adopt the operator nota-
tion X
i j
= q
i

q
j
. The new closure operators are de-
rived from visiting all the commutation relations in the
Jacobi equations. The non-zero commutators are even-
tually presented in Appendix 4 respectively for tra-
jectory F
1
and F
2
. Interesting cyclic properties of the
Lie multiplication can be evidenced in both cases as
a result of the derivation properties: [X
i j
. X
j k
] = X
i k
and [X
i j
. X
kl
] = 0 for any {i j } ,= {kl}. As a further
consequence of cyclicity, all the Jacobi equations are
also satised. Finally, the matrix A is dened in both
cases F
1
and F
2
and belong respectively to gl(4. R) and
gl(7. R). The corresponding Lie groups are GL(4. R)
and GL(7. R).
Let us remark that there is no actual obstacle to con-
sider multidimensional spaces, this just requests an in-
crease of size of the Lie algebra. The differential oper-
ator O in Eq. (81) reads now
O
t
= :
x

1

:

n


n1


t
(84)
and both equations of trajectory 78 and of motion 79
become vector equations.
A Group-Theoretic Construction with Spatiotemporal Wavelets 223
5.2. Diagonalization and Reducibility
of the State Equation
In this section, we showthe reducibility of the Lie alge-
bra representations presented in the previous section.
It turns out that the representations are reducible into
subspaces one for the carrier trajectory and one for
each rotational parameter. It would be interesting to ex-
amine this property with more details since it enables
splitting estimation and tracking into several dedicated
parts. The splitting results from diagonalization and
reducibility as follows:
1. The diagonalization of matrix A. The matrix sup-
ports a (n m)-dimensional vector space made of
n m eigen-values and eigen-vectors. This means
that A
T
= P
1
P where =diag [
1
. . . . .
nm
].
It is also clear that diagonalizing matrix A leads to
diagonalizing the state equation e
A
T
= P
1
e

P
2. The matrix representation A
T
may be written in
forms of
A
T
=

A
1.1
A
2.1
A
3.1
. . . A
m1.1
0 A
2.2
0 . . . 0
0 0 A
3.3
. . . 0
0 0 0 . . . A
m1.m1

.
where A
1.1
is a n n matrix, A
2.1
, A
3.1
. . . . . A
m1.1
are 2 n matrices, A
2.2
, A
3.3
. . . . . A
m1.m1
are
2 2 matrices. The matrix A
T
is reducible into
m 1 complementary closed sub-spaces. The rst
subspace is of dimension n and corresponds to the
parameters of the carrier trajectory x, :,
i
; i =
1. . . . . m 2. The others are of dimension 2 and
correspond to rotational motion. The structure of re-
ducibility in the representation provides the motion
analysis with a hierarchical structure of estimation
and tracking. This approach splits the carrier trajec-
tory from its rotational motions (velocity and ac-
celerations). In the following, we will consider the
expansion up to rotational acceleration of order 2.
As a consequence of items 1 and 2, we can decom-
pose the matrix P in Q and R as follows
A
T
= (QR)
1
QR with Q =

U 0 0
0 V 0
0 0 W

;
R =

I
n
C
1
C
2
0 I
2
0
0 0 I
2

(85)
where I
n
is the n-dimensional unit matrix. We let P =
QR and P
1
= R
1
Q
1
, then
P =

U UC
1
UC
2
0 V 0
0 0 W

.
P
1
=

U
1
C
1
V
1
C
2
W
1
0 V
1
0
0 0 W
1

.
=

1
0 0
0
2
0
0 0
3

. (86)
Matrix U contains n eigen-vectors of the sub-matrix
A
11
since A
11
= U
1

1
U,
1
= diag [
1
. . . . .
n
].
Matrix V contains two eigen-vectors, those of
sub-matrix A
2.2
since A
2.2
=V
1

2
V, with
2
=
diag [
n1
.
n2
],
n1
=i
1
,
n2
=i
1
, and
V =
1

1 i
1 i

.
Similarly, matrix W contains two eigen-vectors, those
of sub-matrix A
3.3
since A
3.3
=W
1

3
W,

3
=diag [
n3
.
n4
].
n1
=i
2
2.
n2
=i
2
2.
Let the state equation of the system be given by
u(t ) =

x(t ). :(t ).
1
(t ). cos[
1
t ]. sin[
1
t ].
cos

2
t
2
2

. sin

2
t
2
2

T
.
The evolution is given by the prediction equation u(t
) = e
O
t
u(t ) = e
A
T
u(t ). We have also
u(t ) =

x(t ). :(t ).
1
(t ).
cos [
1
(t )]. sin[
1
(t )].
cos

2
(t )
2
2

. sin

2
(t )
2
2

T
.
As a result of the diagonalization of matrix e
A
T
,
we yield an interesting closed form of the prediction
equation that splits carrier trajectory from rotational
224 Leduc
motion. The derivation of the one-parameter subgroup
expression with Y
i
= e

i
; i = 1. 2. 3 and Y
4
= e

3
leads to three relations

x(t )

1
(t )
:(t )

= (U
1
Y
1
U)

x(t )

1
(t )
:(t )

[(U
1
Y
1
U)C
1
C
1
(V
1
Y
2
V)]

cos[
1
t ]
sin[
1
t ]

[(U
1
Y
1
U)C
2
C
2
(W
1
Y
3
W)]

cos

2
t
2
2

sin

2
t
2
2

(87)
and also

cos[
1
(t )]
sin[
1
(t )]

=(V
1

2
V)

cos[
1
t ]
sin[
1
t ]

cos

2
(t )
2
2

sin

2
(t )
2
2

=(W
1
Y
3
W)(W
1
Y
4
W) (88)

cos

2
t
2
2

sin

2
t
2
2

For quite evident reasons described in [27], the restora-


tion of the double product in (t )
2
comes from a
contribution of the exponent of the rst order in .
The tracking strategy is based on combining both
Kalman lters and wavelet transform. The state of the
system denes the state of the Kalman lter. This state
is given by the set of the wavelet parameters. These
parameters dened in Sections 2.3 and 2.4 are nothing
but the group parameters. Without any major change
in the reasoning, the scale parameter can be added in
the state vector of Eq. (77). Awhole tracking algorithm
is described in [31] that exploits the wavelet transform
as state parameter estimator and the Bellman theory
of dynamic programming to build the trajectory on the
span of the scene.
5.3. Observation Equation and Wavelet Transform
The estimation of the motion parameters i.e. the group
elements, and the motion-based feature extraction and
reconstruction are based on two concepts: a variational
principle that provides an optimal estimation of motion
and a motion-based ltering that extracts the pattern
with motion of interest. These processes are part of
the Kalman lter and are iteratively performed along
a trajectory. At step 0, all the inital conditions i.e. the
motion parameters are provided by an external device
(RADAR or satellite). This allows the Kalman lter to
start computing.
The observation equation is based on the wavelet
transform where the optimum wavelet T
g
opt
+( x. t ) per-
forms the motion-based extraction of the object in mo-
tion. g
opt
corresponds to the state parameters at step n,
it is estimated by a gradient algorithmin L
2
(G. dj
l
(g))
to yield
g
opt
= arg
g
max [!+
g
[ S)[
2
(89)
or by a dynamic programming approach [31]. The
wavelet +
g
opt
tuned to the estimated state parameters
act as a motion-oriented band-pass lter. The contin-
uous wavelet transform discretized on the grid of the
signal captures and isolates the selected objects from
the scene S to provide a display I ,
I (

b. ) =

+
g
opt

S V

. (90)
I is the segmented image of the selected object, dis-
played alone at its correct location; S is the original sig-
nal under analysis, and V(

b. ) is the noise produced by


the optical sensors. The estimation of state parameters
in the next section is performed with Morlet wavelets.
An alternative way of reconstructing a relevant in-
formation from the inverse wavelet transform consists
of generating a motion-compensated sequence S( x. t )
S( x. t ) =

G
!S [ T
g
+)(T
g
+)( x. t ) dj
l
(g) (91)
Appendix 3 shows howthis technique compensates the
motion to produce a frozen object.
6. Applications and Simulation Results
The applications presented in this paper are based on
Morlet wavelets. An anisotropic Morlet wavelet is ad-
missible as an continuous wavelet in the rotational and
translational family. In this case, the wavelet is struc-
turally a non-separable lter meaning that it can not be
completely factorized in terms of the spatial and tempo-
ral variable. The numerical computation is performed
in the Fourier domain. The Morlet wavelet reads in the
Fourier domain
+(k) = [detC[
12

1
2
!(kk
0
) [ C(kk
0
))
e

1
2
!k
0
[ Ck
0
)
e

1
2
!k [ Ck)

(92)
A Group-Theoretic Construction with Spatiotemporal Wavelets 225
i.e. a Gaussian function shifted to the point of coordi-
nates k
0
=(

k
0
.
0
). Moreover, k=(

k. ) R
2
Rand
C is a positive-denite matrix. For two-dimensional
spatio-temporal signals, we have

C =

c
k
x
0 0
0 c
k
y
0
0 0 c

where the c factors introduce anisotropy in the


wavelet shape. Figures 13 presents three congura-
tions of four-symmetrical rotational Morlet wavelets.
The Shannonsamplingboundis at

4
rad/image. The in-
terested reader should read [26, 31, 35] to compare the
spectrum of the Galilean wavelets in G
4
and compare
to the spectrum of pure rotational motion presented in
these gures.
Figure 4presents the sequence calledCaltrain with
a ball simultaneously in translation and rotation along
a track. Detection and estimation of the angular ve-
locity and acceleration is made possible by the texture
on the rotating ball made of four white dots acting as
the corner of a square in rotation. Figure 4 presents
the sequence at images 1 and 32. The scene to be
analyzed is made of 32 images, each of 256 256
4
2
0
2
4
4
2
0
2
4
0
2
4
6
8
10
12
14
16
18
omega (temporal frequency)
SPECTRUM OF ROTATIONAL WAVELET
k1 (horizontal frequency)
Figure 1. Four symmetrical rotational Morlet wavelet functions close to
1
t = (4)t , and : = (0. 0) pix/image: square modulus in plane of
the (. k
x
) axes i.e. k
y
= 0. This yields critical temporal sampling.
pixels. The ball is traveling from right to left; the cen-
ter is located at x 150 in image 1 and at x 95 in
image 32.
Figure 5 presents the global analysis of velocity per-
formed with Galilean wavelets fromgroup G
4
in a win-
dow B of size 60 60 centered on the ball at image
18. The square modulus of the wavelet transform is
computed on the window as a function of :
x
. :
y
f (:
x
. :
y
) =

bB.=18

+
g
1

2
g
20
. with
g
1
= { :.
0
. a.

b. }. g
20
= {a = a
0
.
0
=
00
}
(93)
provides the analysis with a representation of the lo-
cal spatio-temporal content in terms of its velocity
content. Figure 6 presents the estimation of the mean
angular velocityandthe scale of the ball at velocitywith
rotational wavelet tuned to velocity : = (1.7. 0.1).
Figure 7 proceeds to the estimation of the angular ac-
celeration, the measurement shows a deceleration of
0.0028 rad/image
2
. Figure 8 presents the tracking
of the ball in the scene and the carrier trajectory. The
algorithm exploits the Galilean wavelets from group
G
4
. The square modulus of the wavelet transform is
226 Leduc
3 2 1 0 1 2 3
3
2
1
0
1
2
3
SPECTRUM CONTOUR OF ROTATIONAL WAVELET
k
2

(
v
e
r
t
i
c
a
l

f
r
e
q
u
e
n
c
y
)
omega (temporal frequency)
Figure 2. Four symmetrical rotational Morlet wavelet function at
1
t = (10)t , and : = (0. 0) pix/image: contours of the square modulus
in plane of the (. k
y
) axes i.e. k
y
= 0. This yields 4 10 on the temporal frequency axis. The Fourier signature of this rotational motion
is a ball of energy located on the frequency axis. The Fourier signature of a translational motion is located on a so-called velocity plane. As
described in [26, 31], this plane is orthogonal to the velocity vector v = ( :. 1).
3 2 1 0 1 2 3
3
2
1
0
1
2
3
SPECTRUM CONTOUR OF ROTATIONAL WAVELET
k
1

(
h
o
r
i
z
o
n
t
a
l

f
r
e
q
u
e
n
c
y
)
omega (temporal frequency)
Figure 3. Four symmetrical rotational Morlet wavelet functions at
1
t = (8)t , and : = (1. 0) pix/image: square modulus in plane of the
(. k
x
) axes i.e. k
y
= 0. In this sketch of the energy, translational motion and rotational motion compose each to the other. The Fourier signature
is two distinct components. One is odd in the velocity plane around the origin and the other is even on the temporal frequency axis tilded by the
velocity. Each signature may be estimated independently of the other.
A Group-Theoretic Construction with Spatiotemporal Wavelets 227
Figure 4. Image 1 and 32 in the Caltrain sequence ([256 256] 32 images): motion analysis of the ball, more precisely of four white spots,
which are spinning, and moving on a quasi-horizontal trajectory. There is a deceleration fromimage and moving on a quasi-horizontal trajectory:
deceleration from images 1 to 14 followed by acceleration from images 15 to 32.
3
2
1
0
1
2
3
3
2
1
0
1
2
3
0
50
100
150
200
250
300
350
400
vx
ENERGY OF GALILEAN WAVELET TRANSFORM
vy
Figure 5. Analysis of translational velocities contained in the ball at image 18 on a window that contains the ball: the square modulus of the
wavelet transform (the energy) is computed using the Galilean wavelet of group G
4
. The energy density is integrated on a window at image
18. Two signatures are visible; two symmetric peaks result from the rotational motion estimated with group G
4
which is suited for translational
motion alone and a domed wall testies of the spreading of translational velocities which results from accelerated motion. This treatment is
instructive for the shape of the Fourier signature but not well suited for precise estimations.
successively maximized on the velocity as processed
in [30, 31]. The gravity center of the ball is computed
from the maxima of the square modulus of the wavelet
transform within a window B of size 60 60 which
is sliding according to the velocity estimations. These
maxima are located along the contours of the ball in
translational motion. The correspondingreconstruction
is presented in Fig. 9.
228 Leduc
0
0.02
0.04
0.06
0.08
0.1
1
2
3
4
5
0
1
2
3
4
5
x 10
11
angular velocity: theta (radian/image)
ESTIMATION OF SCALE AND ANGULAR VELOCITY
scale a
Figure 6. Estimation of the angular velocity and the scale of the rotating ball in Caltrain sequence with rotational wavelets. The diagram
sketches the square modulus of the wavelet transform at : = (1.7. 0.1) pix/image. The component at
1
= 0 stands for the non-rotating
structures. The component at
1
= 0.045 rad/image, a = 3.3 is the actual ball contribution which is observed rotating of 2 over 32 images.
The component at
1
= 0.09 rad/image is the rst harmonic.
0.02 0.015 0.01 0.005 0 0.005 0.01 0.015 0.02
20
22
24
26
28
30
32
34
s
q
u
a
r
e
d

m
o
d
u
l
u
s

o
f

t
h
e

w
a
v
e
l
e
t

t
r
a
n
s
f
o
r
m

|
<


|

s
>
|
2
angular acceleration
2
Figure 7. Estimation of the angular acceleration of the rotating ball in Caltrain sequence on image 4 with rotational wavelets from an extended
version of group G
7
which contains
2
. The diagram sketches the square modulus of the wavelet transform at : = {1.7. 0.1} pixel/image,
a = 3.3. The maximum at
2
= 0.003 rad/image
2
corresponds to the angular deceleration. The ball is indeed observed decreasing speed and
rotational motion from image 1 to 14 in the tracking presented in image 7.
A Group-Theoretic Construction with Spatiotemporal Wavelets 229
0 5 10 15 20 25 30 35
90
100
110
120
130
140
150
160
TRAJECTORY TRACKING OF THE BALL
x

c
o
o
r
d
i
n
a
t
e
s


image number
Figure 8. Tracking the moving-ball position in the Caltrain sequence with Galilean wavelets from group G
4
. At image 15, the ball is pushed
and keeps constant speed. The x-coordinates decrease as a result of the motion steering to the left.
Figure 9. Reconstruction of the rotating ball at image 4 in the Caltrain sequence; the scene is ltered by rotational wavelets from group G
7
traveling at velocity : = (1.7. 0.1) pix/image, the scale a = 0.5 to trace all the contours with ne resolution, c
t
= 10 to be more accurate in
position and
1
= 0.045 rad/image. As a result of the uncertainty between velocity and angular orientation, the intensity of the ball boundaries
depends upon the orientation. This explains the fading of the edge intensities on the upper and lower arcs. The isolation of the moving object of
interest is eventually performed by a Gaussian lter of standard radius of 50 (a
ball
= 3.3) centered on the maximum intensity.
230 Leduc
7. Conclusions
In this paper, we have developed new families of
continuous wavelets to analyze rotational motions (an-
gular velocity and accelerations) in spatio-temporal
digital signals. It has been shown how rotational pa-
rameters superpose and combine with translational pa-
rameters when traveling along the carrier trajectory and
how the motion estimation and tracking of parameters
can be performed quasi-independently. Simulations re-
sults have presented the practical capabilities of these
wavelet estimation process.
Appendix 1
This appendix provides some guidelines and outlined
ideas on howLie group representations have been com-
puted in this paper. As an example, the multidimen-
sional afne group SIM(n) is sketched here below. The
reason of this choice is the small amount of compu-
tations involved for this case. The Galilei and rota-
tional cases proceed exactly on the same sketch but
involve more machinery. For all denitions, theorems,
properties, the interested reader is invited to consult
the following references [4, 6, 9, 14, 23, 24, 34, 37].
The construction of UIRs may proceed from two ap-
proaches namelyMackeys Methodof Induction[5, 34]
and the Kirillovs Method of Orbits [23]. Both methods
can also be intertwined.
In the case where G is structured as a semi-direct
product G = HK where K and H are subgroups.
Let be a group homorphism from K into AUT(H).
G is the set of the ordered pairs (k. k) with h H
and k K. We have now the following strategy to
determine all the unitary representations of G.
1. Determine the set H

of all the irreducible charac-


ters of H.
2. Search for the orbits O
K
(

) in the dual group H

. Classify all the orbits.


3. From each orbit, select an element
0
O
K
(

)
and determine the form of the stability subgroup
K(
0
) K.
4. Take a representation L
K
of K(
0
) and induce the
representation T
g
as IND
G
K
L
K
in Eq. (5).
The interested reader should read Glimms theorem
[18]. The importance of this theoremlay in its ability to
guarantee the irreducibility of induced representations.
Basically, Glimms theorem allows us an easy way to
check for the existence of a system of imprimitivity for
a representation [5, 34]. Let G = HK, and let the
orbits of G in H

satisfy the conclusions of Glimms


theorem. Then
1. For each orbit O
K
(

) H

, and for each UIR L


K
of the stability subgroup, there exists an induced
representation for G dened as IND
G
K
L
K
in
Eq. (5).
2. The carrier space of the induced UIR IND
G
K
L
K
is the Hilbert space H
IND
G
K
L
K
= L
2
(O
K
. dj).
The form of these induced representations has been
presented in Eq. (5) and their computation required to
derive the adjoint and coadjoint actions. Let us treat
the case G
1
= SIM(n) for the shake of simplicity. If g,
g
/
G
1
, then the conjugation on the group is given by
gg
/
g
1
= (

b aR

b
/
a
/
RR
/
R
1

b. a
/
. RR
/
R
1
)
(94)
and, thus, the adjoint action of G on G (computed from
gg
/
g
1
g
/ [
g
/
=e
is given by
Ad
g
X = X
/
= (

b
/
. s
/
. r
/
). X. X
/
G (95)
with

b
/
= aR

b (sI Rr R
1
)

b (96)
s
/
= s (97)
r
/
= Rr R
1
(98)
We further consider G

to be the algebraic dual of G


in the usual manner, and let X

to be X

=
(p

. s

. r

) with p

=

k dened in Eq. (5). The duality
between G and G

is expressed by
!X. X

) = p p

ss

tr(rr

) (99)
where tr is the trace. We can proceed to the coadjoint
action of G on G

which is computed as Ad

g
= Ad
T
g
and is given by
Ad

g
X

=

X

= (

p
/
. s
/
. r
/
). X

.

X
/
G

(100)
with

p
/
= aR
1
p

(101)
s
/
= s



b p

(102)
r
/
= R
1
r

R R
1

b r
1
p

(103)
A Group-Theoretic Construction with Spatiotemporal Wavelets 231
Let G(k) = HH(k) denote the stabilizer of point k
in H

where H(k) is called the little group (stability


subgroup). For G = SIM(n), V

. R
n
. Let us x the
point e = (

0. 1) in R
n
(

0 R
n1
). The stabilizer of e
is G( e) = SIM( e) = R
n
H( e) = R
n
SO(n 1). In
this example, the little group is H( e) =SO(n 1) and
then
L
H
(s) = D
[ j ]
(s). s SO(n 1) (104)
where the D
[ j ]
are the representations of SO(n 1) [5,
16]. The Radon-Nikodym derivative is computed from
the coadjoint action det

aR
1

= a
n
. For applications
in signal processing, we let D
[ j ]
(s) = 1 = L
H
(s) (no
spinor representations are required). Following Eq. (5),
the UIRs T
g
of G
1
in the space L
2
(R
n
. dx) are then
given by
[T
g

+]( p) = a
n
2
e
i

b p

+( p
/
) (105)
with
p
/
= aR
1
p (106)
These UIRs are unitary and irreducible.
Appendix 2
Theorem 1. The UIRs of G
3
in the space L
2
(R
n2
.
dx) given by g T
g
are square-integrable.
Proof: Let us recall that the UIRs g T
g
are square-
integrable if and only if

G
[!T
g
+. +)[
2
dj
l
(g) - (107)
Let us denote the element (m.

k. ) by and denote
the measure dm d

k d by d. Then, we may interpret


as an element of

N, the dual vector space of N, and
we may consider the integration over

N in the inner
product in Eq. (107) to be an inverse Fourier Transform.
We let
d x = d d

b d. ds =
d : da da
0
dm(R)
a
2n3
a
(n1)
0
(108)
so that dj
l
(g) = d xds.
The integration in Eq. (107) proceeds as follows

G
[!T
g
+. +)[
2
dg
=

N
[T
g

+]()

+()d

2
d xds (109)
=

S
[I
N
] ds (110)
Moreover, we let

V() = [T
g

+]()

+() (111)
and calculate I
N
as follows (the symbols and stand
respectively for direct and inverse Fourier versions).
Since

V() contains the e
i x
in T
g

+, the integral on

N
is an actual inverse Fourier transform. Therefore,
I
N
=

N
[

V(x)[
2
dx (112)
= |

V(x)|
2
L
2
(N.dx)
(113)
= |

V()|
2
L
2
(

N.d)
(114)
=

N
[

V()[
2
d (115)
=

N
[[T
g

+]()[
2
[

+()[
2
d (116)
At this point, the Fubinis theorem applied to Eq. (110)
grants

S
[I
N
] ds =

N
[

+()[
2

S
[T
g

+]()[
2
ds

d
(117)
Thus, if there exists a non-zero + L
2
(R
n2
. d x) so
that
c
+
=

N
[

+()[
2

S
[[T
g

+]()[
2
ds

d -
(118)
the theoremis proved. Let us rst observe that SO(n) is
a compact group, the integration over R can be pushed
outside of Eq. (118). So as long as the remaining inte-
grals are nite, c
+
will also be nite. Next, Eq. (118)
may be rewritten as
c
+
=

R
n2
[

+(m.

k. )[
2
I
1
dm d

k d - .
(119)
232 Leduc
where
I
1
=

R
n2
[

+(m
/
.

k
/
.
/
)[
2
d : da da
0
a
n1
a
(n1)
0
- .
(120)
At this stage, the change of variable dm
/
d

k
/
d
/
=
J(d : da da
0
) is required to eliminate the parameters
from S. The Jacobian is given by
J =

det

m
/
a
0

k
/
a
0

/
a
0
m
/
a

k
/
a

/
a
m
/
:

k
/
:

/
:

(121)
=

det

a
2
a
2
0

0
T
a
0

/
2
a
a
0
[a
1

k
/
]
T
0

0 amI
n
a
0
a
1
R

k
/

(122)
=
a
n1
m
n1
a
0
[[

k
/
[
2
2m
/

/
]. (123)
Thus, I
1
yields after the change of variables
I
1
=

R
n2
[

+(m
/
.

k
/
.
/
)[
2
m dm
/
d

k
/
d
/

a
2n
m
n
a
n
0

a
4
[[

k
/
[
2
2m
/

/
]
(124)
But
a
2
=
[

k
/
[
2
2m
/

/
[

k[
2
2m
. m
/
=
a
2
m
a
0
(125)
Therefore, Eq. (124) may be rewritten as

R
n2
[

+(m
/
.

k
/
.
/
[
2
m[[

k[
2
2m]
2
(m
/
)
n
[[

k
/
[
2
2m
/

/
]
3
dm
/
d

k
/
d
/
(126)
Hence, c
+
- in Eq. (118) if the following pair of
integrals is nite

R
n2
[

+(m
/
.

k
/
.
/
)[
2
m[[

k[
2
2m]
2
dm d

k d -
(127)
and

R
n2
[

+(m.

k. )[
2
dm d

k d
(m)
n
[[

k[
2
2m]
3
-
(128)
Since there exist non-zero functions in L
2
(R
n2
. d x)
which fulll these two conditions, the theorem is
proven.
Theorem 2. The UIRs of G
4
in the space
L
2
(R
n1
. d
n

kd) which are given by the mapping


g T
g
are square-integrable.
Proof: In this case, following the same path as in
Theorem 1, we want to examine the boundness of
I
2
=

R
n1


+
m
0
(

k
/
.
/
)

2
da d :
a
n1
(129)
Therefore, we reach a change of variables from (a. :)
to (

k
/
.
/
) given by d

k
/
d
/
= J da d :, where
J =

det

k
/
a

/
a

k
/
:

/
:

(130)
=

det

a
1
(

k
/
)
T
0
am
0
R
1
a
1
R

k
/

(131)
= a
n3
m
n1
0
[

k
/
[
2
(132)
Thus, Eq. (129) becomes
I
2
=

R
n1


+
m
0
(

k
/
.
/
)

2
d

k
/
d
/
a
2n2
m
n1
0
[

k
/
[
2
(133)
Therefore, we obtain the condition of square-
integrability (129) which requires that the following
integral be nite
c
+
=

R
n1


+
m
0
(

k. )

2
I
m
0
(

k. ) d

k d -
(134)
where
I
m
0
(

k. ) =

R
n1


+
m
0
(

k
/
.
/
)

[[

k[
2
2m
0
(
/
)]
n1
[

k[
2n
m
n1
0

k
/
d
/
-
(135)
There are many functions for which the integral will
be nite. For c >0, the following

+ L
2
(R R)
will be admissible as Galilean wavelets:

+(

. ) =
A Group-Theoretic Construction with Spatiotemporal Wavelets 233
O([

[
(3n2c)
). [

[ . R,

+(

. ) =
O([

[
nc
). [

[ 0. R,

+(

. ) =
O([[
(2n1c)
). [[ .

R
n
,

+(

. ) =
O([[
1c
). [[ 0.

R
n
.
Appendix 3
In this appendix, the condition of wavelet admissibility
is examined with Calder ons point of view. In this ap-
proach, we look for the condition to yield an invertible
wavelet transform i.e. an isometry. The example to be
considered is the Galilei group with g = {b. . :. a}
in a one-dimensional spacetime; so we may consider
m
0
= 0. To illustrate this technique developed by
Calder on, we will refer to the spacetime version of the
group representation in (38). The inner product
!F. T
g
+) =

RR
F

a
12

+

1
a
[y b :( )]

dy d
is the wavelet transformdened in Section 2. The func-
tional version of the operator statement of the isometry
I
H
o = W
1
+
W
+
reads as follows
!F. T
g
+) =

RR
F

a
12

+

1
a
[y b :( )]

dy d
read as follows
F( x. t ) =

G
!F. T
g
+)(T
g
+)(X)dj
l
(g)
=

RR
F

a
12

+

1
a
[y b :( )]

dy d

a
12
+

1
a
[x b :(t )]
t

db d d: da
a
3
=

RR
F

R
a
1
+

1
a
[x b :( )]

a
1

+

1
a
[y b :(t )]
t

dy d db d

d: da
a
2
=

RR
F

R
+
a

x y u :(t n)
t n

+
a

u :n
n

dy d du dn

d: da
a
2
=

RR
F

(+
a


+
a
)

(x y) :(t )
t

dy d

d: da
a
2
=

{F (+
a


+
a
)}

x
t

d: da
a
2
where we have let = n, b = y u and also

+(x. t ) = +(x. t ) and +


a
(x. t ) = a
1
+(
x
a
. t ).
Let us also remark that the introduction of the spatio-
temporal convolution denoted . Spatial and temporal
translations are related to each other through the veloc-
ity: alonga Galileanframe movingat constant velocity
:. In this case, the convolution dened on the Galilei
group is twisted and performed along this displacement
(i.e. along the trajectory) in order to reconstruct the
still signal F(x. t ). This is a reminiscence of the tech-
nique of motion-compensated ltering. Eventually, we
moved to the Fourier domain to retrieve the condition
of admissibility of the Galilean wavelet as described
in [26]. We have introduced in the second equation the
left-invariant Haar measure dj
l
(g) which satises by
denition

G
F(g
1
h) dj
l
(h) =

G
F(h) dj
l
(h) g. h G
(136)
The left-invariant Haar measure can be calculated from
the law of composition as an innitesimal variation of
volume in the parameter space
dj
l
(g) =

det

g
e
g
/
e
g
/
e

1
g
/
e
=e
db d d: da (137)
=
1
a
3
db d d: da (138)
234 Leduc
Proceeding further with F( x. t ) in the Fourier domain,
we get

F(

k. ) =

F(k. )

+(a k. k:)[
2
d: da
a
2
(139)
which leads to the usual condition of square-
integrability of the Galilean wavelet in one-
dimensional spacetime

R
[

+(k. )[
2
[k[
2
dk d = 1 (140)
We have illustrated an alternative way of determining
the condition of wavelet admissibility. Taking into ac-
count the spacetime domain, we have learned that the
motion-based wavelet analysis performs the motion-
compensated ltering.
Appendix 4
This appendix illustrates the way of howto compute the
Lie algebras G and the matrix representation A dened
in Eqs. (76) and (77). We proceed successively from
equations of motion F
1
and F
2
.
Equation of motion F
1
is given by

1
(t ) = a
1
x(t ) a
2
:(t ) b cos[
1
t ] (141)
The time differential operator reads
O
1
= :

x
a
1
x

:
a
2
:

:
b cos[
1
t ]

:


t
(142)
The state of the dynamical system is composed as fol-
lows: q
1
= x; q
2
= :; q
3
= cos
1
t ; q
4
= sin
1
t . The
innitesimal Lie generators in O
1
are given by
X
12
= x

:
; X
21
= :

x
; X
22
= :

:
;
X
32
= cos[
1
t ]

:
; X
43
=
1

t
The additional Lie generators for closure are as follows
X
11
= x

x
; X
31
= cos[
1
t ]

x
; X
41
= sin[
1
t ]

x
;
X
42
= sin[
1
t ]

:
A generic element Y G becomes
Y = aX
11
bX
21
cX
22
d X
32
eX
43
f X
31
gX
41
hX
42
The representation in the parameter state reads
A = X
21
a
1
X
12
a
2
X
22
bX
32

1
(X
34
X
43
)
(143)
Eventually, matrix A reads

0 a
1
0 0
1 a
2
0 0
0 b 0
1
0 0
1
0

(144)
The construction of the Lie algebra G and the matrix
representation A from equation of motion F
2
proceeds
as follows. The equation of motion F
2
is given by

2
(t ) = a
1
x(t ) a
2
:(t ) a
3
(t ) b
1
cos[
1
t ]
b
2
cos

2
t
2
2

(145)
The time differential operator is
O
2
= :

:
a
1
x

a
2
:

a
3

b cos[
1
t ]


t
(146)
The state of the dynamical system is as follows:
q
1
=x; q
2
=:, q
3
= ; q
4
= cos
1
t ; q
5
= sin
1
t ; q
6
=
cos
2
t
2
2
; q
7
= sin
2
t
2
2
. The innitesimal Lie genera-
tors in O
2
are given by
X
32
=

:
; X
21
= :

x
; X
13
= x

; X
23
= :

;
X
33
=

; X
42
= cos[
1
t ]

; X
54
=
1

t
X
76
=
2

t
2
A Group-Theoretic Construction with Spatiotemporal Wavelets 235
The additional Lie generators for closure follow as
X
11
= x

x
; X
12
= x

:
; X
22
= :

:
; X
23
= :

;
X
31
=

x
; X
41
= cos[
1
t ]

x
; X
42
= cos[
1
t ]

:
;
X
51
= sin[
1
t ]

x
; X
52
= sin[
1
t ]

:
;
X
53
= sin[
1
t ]

; X
61
= cos

2
2
t
2


x
;
X
62
= cos

2
2
t
2


:
; X
71
= sin

2
2
t
2


x
;
X
72
= sin

2
2
t
2


:
; X
73
= sin

2
2
t
2

A general element Y G reads


Y = aX
11
bX
12
cX
22
d X
23
eX
31
f X
41
gX
42
hX
51
X
52
j X
53
The representation in the parameter state is given by
A = X
21
X
32
a
1
X
13
a
2
X
23
a
3
X
33
bX
42

1
(X
45
X
54
)

2
2
(X
67
X
76
)
Eventually, matrix A comes as follows

0 0 a
1
0 0 0 0
1 0 a
2
0 0 0 0
0 1 a
3
0 0 0 0
0 0 b
1
0
1
0 0
0 0 0
1
0 0 0
0 0 b
2
0 0 0

2
2
0 0 0 0 0

1
2
0

(147)
Acknowledgments
The author wishes to express his gratitude to the anony-
mous reviewer who provided substantial and construc-
tive details to improve the paper presentation, and to
Prof. B. Blank and J. Corbett, PhD student, both in the
Mathematics Department of Washington University in
Saint Louis for helpful discussions on the topic. The au-
thor also thanks Prof. B.K. Ghosh in the Department of
Science Systems and Mathematics, Washington Uni-
versity in Saint Louis for providing the computing
support.
This material is based upon work supported by the
US Air Force under grant F49620-99-1-0068.
References
1. R. Abraham and J. Marsden, Foundations of Mechanics,
The Benjamin/Cummings Publishing Company: Reading, MA,
1978.
2. S.T. Ali, J.-P. Antoine, J.-P. Gazeau, and U.A. Mueller, Coher-
ent states and their generalizations: A mathematical overview,
Review of Mathematical Physics, Vol. 7, pp. 10131065, 1995.
3. V.I. Arnold, Mathematical Methods of Classical Mechanics,
Springer-Verlag, New York, 1978.
4. A.O. Barut, Localized rotating wavelets with half integer spin,
Physics Letters A, Vol. 189, pp. 277280, 1994.
5. A.O. Barut and R. Raczka, Theory of Group Representations
and Applications, PWNPolish Scientic Publishers, 1985.
6. P. Bernat, N. Conze, M. Duo, M. L evy-Nahas, M. Rais, P.
Renouard, and M. Vergne, Repr esentations des groupes de Lie
r esolubles, Dunod: Paris, 1972.
7. R.E. Blahut, W. Miller, and C. Wilcox, Radar and Sonar, Part I,
Springer Verlag: New York, 1991.
8. C. Bretin and Gazeau, Generalized bessel series and multiplic-
ity problem in complex semisimple lie algebra theory, Journal
of Mathematical Physics, Vol. 22, No. 10, 1981.
9. Th. Br ocker and T. Dieck, Representations of Compact Lie
Groups, Springer-Verlag: New York, 1985.
10. J.S. Byrnes et al., Wavelets and their Applications, Kluwer
Academic Publishers, 1994.
11. T. Caelli, W. Hoffman, and H. Lindman, Subjective Lorentz
transformations and the perception of motion, Journal of the
Optical Society of America, Vol. 68, No. 3, pp. 402411, 1978.
12. M. Duval-Destin and R. Murenzi, Spatio-temporal wavelets:
Application to the analysis of moving patterns, Progress in
Wavelet Analysis and Applications (Proc. Toulouse 1992), Y.
Meyer and S. Roques (Eds.), Fronti` eres: Gif-sur-Yvette, France,
1993, pp. 399408.
13. D.J. Fleet and A.D. Jepson, Hierarchical construction of ori-
entation and velocity selective lters, IEEE Transactions on
Pattern Analysis and Machine Intelligence, Vol. 11, No. 3, pp.
315325, 1989.
14. G.B. Folland, A Course in Abstract Harmonic Analysis, CRC
Press: Boca Raton, 1995.
15. W.A. Gardner, Introduction to Random Processes, Macmillan
Publishing Company, 1986.
16. I.M. Gelfand, R.A. Minlos, and Z.Ya. Shapiro, Representation
of the Rotation and Lorentz Groups and their Applications, The
Macmillan Company: New York, 1963, part I, pp. 1153.
17. R. Gilmore, Lie Groups, Lie Algebras, and Some of Their Ap-
plications, John Wiley: New York, 1974.
18. J. Glimm, Locally compact transformation groups, Transac-
tions of the American Mathematical Society, Vol. 101, No. 1, pp.
124138, 1961.
19. K.I. Gross, Bessel functions and representation theory I, Jour-
nal of Functional Analysis, Vol. 22, pp. 73105, 1976,.
20. C. Heil and D. Walnut, Continuous and discrete wavelet trans-
forms, SIAM Review, Vol. 31, No. 4, pp. 628666, 1996.
21. M. Holschneider, Wavelets: An Analysis Tool, Clarendon Press:
Oxford, England, 1995.
236 Leduc
22. A. Jerri and I. Joslin, Truncation error for the generalized bessel
type samplingseries, Journal of FranklinInstitute, Vol. 314, No.
5, pp. 323328, 1982.
23. A.A. Kirillov, Elements of the Theory of the Representations,
Springer Verlag: New York, 1976.
24. A.W. Knapp, Representation Theory of Semisimple Groups,
Princeton University Press: Princeton, 1986.
25. N. Lebedev, Special Functions and their Applications, Trans-
lated by R. Silverman, Dover Publications: New York, USA,
1972.
26. J.-P. Leduc, Spatio-temporal wavelet transforms for digital sig-
nal analysis, Signal Processing, Elsevier, Vol. 60, No. 1, pp.
2341, 1997.
27. J.-P. Leduc, Harmonic analysis associated with spatio-temporal
transformations, in Proceedings of ICASSP-2000, Istanbul,
Turkey, Vol. 4, 2000, pp. 22982301.
28. J.-P. Leduc, Continuous and discrete wavelets for acceleration
kinematics: A motion estimation based on Lie groups and vari-
ational principles, Journal of Applied and Computational Har-
monic Analysis, Vol. 11, 2001, pp. 347386.
29. J.-P. Leduc, J. Corbett, and V. Wickerhauser, Rotational wavelet
transforms for motion analysis, estimation and tracking, in Pro-
ceedings of ICIP-98, Chicago, 1998, p. 4.
30. J.-P. Leduc, F. Mujica, R. Murenzi, and M.J.S. Smith, Spatio-
temporal wavelet transforms for motion tracking, in Proceed-
ings of ICASSP-97, Munich, Germany, Vol. 4, 1997, pp. 3013
3017.
31. J.-P. Leduc, F. Mujica, R. Murenzi, and M. Smith, Spatio-
temporal wavelets: A group-theoretic construction for motion
estimation and tracking, SIAMJournal of Applied Mathematics,
Vol. 61, No., pp. 596632, 2001.
32. R. Lenz and K. Homma, Constructing iterative matching algo-
rithms with the use of Lie theory: Three-dimensional orientation
example, Journal of the Optical Society of America A (Optics,
Image Science and Vision), Vol. 14, No. 8, pp. 17341741, 1997.
33. J.-M. L evy-Leblond, Galilei group and Galilean invariance, in
Group Theory and its Applications, Vol. II, E.M. Loebl (Ed.),
Academic Press: New York, 1971, pp. 221296.
34. G.W. Mackey, The Theory of Group Representations, Depart-
ment of Mathematics, University of Chicago, 1955.
35. F. Mujica, J.-P. Leduc, R. Murenzi, and M. Smith, A new pa-
rameter estimation algorithm based on the continuous wavelet
transform, IEEETransactions on Image Processing, Vol. 9, No.
5, pp. 873888, 2000.
36. A.M. Perelomov, Coherent states for arbitrary lie group, Com-
munications in Mathematical Physics, Vol. 26, pp. 222236,
1972.
37. S.R. Quint, Representations of Solvable Lie Groups-Lecture
Notes, Department of Mathematics, University of California at
Berkeley, 1972.
38. J. Rubinstein, J. Segman, and Y.Y. Zeevi, Recognition of dis-
torted patterns by invariance kernels, in Proceedings of the 10th
International Conference on Pattern Recognition, Vol. 2, 1990,
pp. 159164.
39. S. Saeki, On the reproducing formula of calderon, The Journal
of Fourier Analysis and Applications, Vol. 2, No. 1, pp. 1528,
1995.
40. H. Samelson, Notes on Lie Algebras, Springer Verlag: NewYork,
1998.
41. I. Segall and Y.Y. Zeevi, Application of wavelet-type functions
in image processing, in Proceedings of the SPIE-The Interna-
tional Society for Optical Engineering, Vol. 1, 1991, pp. 1048
1058.
42. J. Segman and Y.Y. Zeevi, Image analysis by wavelet-
type transforms: Group theoretic approach, Journal of Math-
ematical Imaging and Vision, Vol. 3, No. 1, pp. 5177,
1993.
43. J.-P. Serre, Linear Representations of Finite Groups, Springer
Verlag: New York, 1977.
44. V.S. Varadarajan, Lie groups, Lie Algebras and their Represen-
tations, Springer Verlag: New York, 1984.
45. N.J. Vilenkin, Special Functions and the Theory of Group Rep-
resentations, American Mathematical Society: Providence, RI,
1968.
46. K. Wodkiewicz and J.H. Eberly, Coherent states, squeezed uc-
tuations, and the SU(2) and SU(1. 1) groups in quantum-optics
applications, Journal of the Optical Society of America B, Vol.
2, No. 3, 1985.
47. C.E. Wulfman and B.G. Wybourne, The Lie group of New-
tons and Lagranges equations for the harmonic oscillator,
Journal of Physics A: Math. Gen., Vol. 9, No. 4, pp. 507518,
1976.
Jean-Pierre Leduc received the degree of Electrical Engineer at
the Facult e Polytechnique de Mons (Mons, Belgium) in 1978. He re-
ceived his Master degree at Columbia University (NewYork) in 1987
and his Doctorate degree at the Universit e Catholique de Louvain
(Louvain-la-Neuve, Belgium) in 1993, both in Electrical Engineer-
ing (Signal Processing and Telecommunications). He also earned
a degree in Operations Research and in Theoretical Physics from
the Facult e Polytechnique de Mons and the Universit e Catholique de
Louvain, respectively. He has been successively at the Laboratoire de
Telecommunications et de t el ed etection of the Universit e Catholique
de Louvain, from 1988 until 1993, at the Institut de Recherche en
Informatique et Syst` emes Al eatoires in Rennes, France, from 1993
until 1996, at the Georgia Institute of Technology from1996 to 1997,
and at Washington University in Saint Louis (Math dept.) from 1997
to 2000. Since September 2000, he is with the Mathematics Depart-
ment of the University of Maryland, College Park.
His present research interests include spatio-temporal signal pro-
cessing, discrete and continuous wavelet transforms, group repre-
sentation theory and optimal control. He is also interested in motion
analysis, detection and tracking.

Das könnte Ihnen auch gefallen