A system for marker-less motion capture

B. Rosenhahn, T. Brox, U. G. Kersting, A. W. Smith, J. K. Gurney, and R. Klette

In this contribution we present a silhouette based human motion capture system. The system components contain silhouette
extraction based on level sets, a correspondence module, which relates image data to model data and a pose estimation
module. Experiments are done in different camera setups and we estimate the model components with 21 degrees of freedom
in up to two frames per second. To evaluate the stability of the proposed algorithm we perform a comparison of the motion
estimation system with a marker based tracking system. The results show the applicability of the system for marker-less
sports movement analysis. We finally present extensions for motion capture in complex environments, with changing lighting
conditions and cluttered background.
This paper is an extended version of [21] which was awarded the DAGM Main Prize on the annual symposium of the German
pattern recognition society (DAGM) in Vienna, 2005.

1 Introduction plexity (21 degrees of freedom). We perform a comparison of

the system with a commercial marker based tracking system [17]
Classical motion capture (or MoCap) comprises techniques for used to analyze sports movements1 . Exercises, such as push ups
recording the movements of real objects, usually humans [25]. or sit ups are analyzed. We further present an approach which
A common application aims to analyze the captured data for allows to track humans in cluttered scenes and changing light-
subsequent motion analysis purposes, e.g., clinical studies, diag- ing conditions. This is an extension of [2], where previously used
nostics of orthopaedic patients or to help athletes to understand rigid objects are replaced with multiple free-form surface patches
and improve their performances. It has also grown increasingly connected by joint axes of various degrees of freedom.
important as a source of motion data for computer animation. The contribution is organized as follows: We will start with
Surveys on existing methods for MoCap can be found in [16, 12]. the basic setup of the motion capture system. Then we will
Well known (commercially available) marker-based tracking sys- continue with the system modules. Here we will briefly describe
tems exist, e.g. provided by Motion Analysis, Vicon or Simi [15]. image segmentation based on level sets, pose estimation and
There are intrinsic problems in using surface markers, e.g. incor- the dynamic occlusion handling to deal with partial occlusion
rect tracking of markers, tracking failures, the need for special in certain frames. The next section presents the experimental
lab environments and lighting conditions and the fact, that peo- results, the quantitative error analysis and motion capture results
ple do not feel comfortable with markers attached to the body. in scenes with cluttered background. The contribution ends with
This often leads to unnatural motion patterns. Marker-based a brief summary.
systems are also designed to track the motion of the markers
themselves and thus it must be assumed that the recorded mo-
tion of the markers is identical to the motion of the underlying 2 The human motion tracking sys-
human segments. Since human segments are not truly rigid,
this assumption may cause problems, especially in highly dy- tem
namic movements typically seen in sporting activities. For these
reasons, marker-less tracking is an important field of research. It A 3D object model builds the a priori knowledge of the system,
requires knowledge in biomechanics, computer vision and com- which is in this case given as two free-form surface patches with
puter graphics. two kinematic chains. Each kinematic chain consists of seven
joints (three for the shoulder, two for the elbow and two for
Typically, researchers working in the area of computer vision
the wrist). Furthermore we added one back segment joint to
prefer simplified human body models, e.g., based on stick, ellip-
the torso surface patch. The estimation procedure is dealing
soidal, cylindrical or skeleton models [1, 14, 11, 13]. In computer
with 21 unknowns, six for the pose parameters (three for ro-
graphics advanced object modelling and texture mapping tech-
tation and three for translation), seven for each arm and one
niques for human motions are well known [24, 8, 6, 26] but the
backbone joint. During correspondence estimation (along four
image processing and pose estimation techniques are often sim-
frames) we collect around 5000 point correspondences (slightly
plified. Therefore we started to combine silhouette based pose
varying dependent on the visible information) and still track in
estimation with more realistic human models: These are repre-
two frames per second for four camera sequences. Using the 3D
sented by free-form surface patches and local morphing along
model and images from a (triggered) calibrated multi-camera
the surface patches is applied to gain a realistic human model
sequence, the motion tracking system consists of three main
within silhouette based MoCap. Our previous works [22, 23] are
now extended to a complete human motion capture system. The 1 Motion Analysis Corporation is one of the leading providers of
system consists of an advanced image segmentation method, dy- optical motion capture systems in entertainment, video-games, film,
namic occlusion handling and kinematic chains of higher com- broadcasting, virtual reality, medicine, sports, and research.

Page 1
 Input Image Extracted Silhouettes functional that is sought to be minimized:
Z
`
E(Φ, p1 , p2 ) = − H(Φ) log p1 + (1 − H(Φ)) log p2
Ω
´
+ν|∇H(Φ)| dx (1)

where ν > 0 is a weighting parameter and H(s) is a regularized

version of the Heaviside function, e.g., the error function. Min-
Pose result Correspondences imization with respect to the region boundary represented by Φ
can be performed according to the gradient descent equation
„ „ ««
p1 ∇Φ
∂t Φ = H ′ (Φ) log + ν div (2)
p2 |∇Φ|
where H ′ (s) is the derivative of H(s) with respect to its argu-
ment. The probability densities pi are estimated according to
the EM principle: having the level set function initialized with
some contour, the probability densities within the two regions
are estimated by the gray value histograms smoothed by a Gaus-
sian kernel Kσ with standard deviation σ. The density estimates
Figure 1: The capture system consists of iterating the following are used in the gradient descent and the process is iterated until
steps: Segmentation, correspondence estimation, pose estima- it converges.
tion. This rather simple and fast approach is sufficient for our
laboratory set-up. Figure 2 shows an example image and the
contour evolution over time. Obviously, the body silhouette is
components, namely silhouette extraction, matching and pose well extracted, besides some deviations in the head region caused
estimation. All components are iterated to stabilize segmenta- by the dark hair. Such inaccuracies, however, can be handled by
tion on the one hand and pose estimation on the other hand. the pose estimation procedure. Moreover, we have a tracking
assumption. Therefore, we can initialize the silhouette with the
pose of the last frame which greatly reduces the number of
iterations needed. This implementation is fast; the segmentation
2.1 Image segmentation algorithm needs 50 ms per frame, i.e., 200 ms in a four-camera
setup.
In more complex scenes, the segmentation is susceptible to
distracting artifacts like shadows, texture, or background clut-
ter. A grossly bad contour, however, makes the pose estimation
procedure fail. In [2] we therefore presented an approach for
iterative segmentation and pose estimation, where the shape
knowledge supports the segmentation procedure. The energy in
(1) is extended by an additional term
Z
EShape (Φ, ξ) = λ (Φ − Φ0 (ξ))2 dx (3)
Ω

Figure 2: Silhouette extraction based on level set functions. that penalises deviations of the contour from the projected
Left: Initial segmentation. Right: Segmentation result. model surface with the pose parameters ξ given as level set
function Φ0 . This term is minimized by drawing the contour
We refer to image segmentation as the estimation of the towards the projected surface on one hand, but also by adapting
closed object contour in the image. This task can become very the pose parameters such that the projected surface better fits
difficult, since noise, shading, occlusions, inhomogeneous ob- to the extracted contour. The energy functional therefore repre-
jects, or background clutter may distort the segmentation. We sents a variational model for the joint estimation of the contour
apply the level set technique [19, 7, 9, 3], in which a level set and the pose parameters.
function Φ ∈ Ω 7→ R splits the image domain Ω into two regions Local optimization is achieved by alternating the minimiza-
Ω1 and Ω2 , the object and the background. It holds Φ(x) > 0 if tion with respect to the contour for fixed pose parameters
x ∈ Ω1 and Φ(x) < 0 if x ∈ Ω2 . The zero-level line thus marks
„ „ ««
p1 ∇Φ
the boundary between the two regions. The segmentation is ∂t Φ = H ′ (Φ) log + ν div
p2 |∇Φ|
sought to maximize the total a-posteriori probability given the
+ 2λ (Φ0 − Φ). (4)
probability densities p1 and p2 of Ω1 and Ω2 , i.e., pixels are as-
signed to the most probable region according to the Bayes rule.
Additionally, the boundary between both regions should be as and the minimization with respect to the pose parameters for a
small as possible. This can be expressed by the following energy fixed contour with the method described in section 2.3.

Page 2
 The model in [2] further extends the image gray value by space. To reconstruct a group action M ∈ SE(3) from a given
color and texture features from [4] in order to allow for tex- ˆ = M ∈ SE(3) must be
twist, the exponential function exp(θξ)
tured objects. These additional features can be integrated into computed. This can be done efficiently by using the Rodriguez
the segmentation model via the probability density models [3]. formula [18],
Whereas in [2] the model has been restricted to rigid objects,
(I − exp(ω̂θ))(ω × v) + ωω T vθ
„ «
section 3.1 presents first results for motion capture of human ˆ = exp(θω̂)
exp(ξθ) ,
upper torsos with cluttered background. Due to its iterative 01×3 1
nature, the extended algorithm is considerably slower than the for ω 6= 0 (8)
basic one, and real-time performance is currently not available.
with exp(θω̂) computed by calculating
2.2 Correspondence estimation
exp(θω̂) = I + ω̂ sin(θ) + ω̂ 2 (1 − cos(θ)). (9)
After image segmentation, correspondences between the object
model and the extracted silhouettes have to be established. To Note that only sine and cosine functions of real numbers need
this end, we follow a modified version of an ICP algorithm [22] to be computed.
and use a voting method to decide whether a point belongs to For pose estimation we combine the reconstructed Plücker
the torso or to one of the arms. These correspondences are used lines with the screw representation for rigid motions and apply
by the pose estimation module described in the subsequent sec- a gradient descent method: Incidence of the transformed 3D
tion to determine new pose parameters. The correspondences point Xi with the 3D ray Li = (ni , mi ) can be expressed as
are iteratively updated according to the new pose until the over-
all pose converges. ˆ i )3×1 × ni − mi
(exp(θξ)X = 0. (10)

Indeed, Xi is a homogeneous 4D vector and after multiplication

2.3 Pose estimation with the 4×4 matrix exp(θξ)ˆ we neglect the homogeneous com-
For pose estimation we assume a set of point correspondences ponent (which is 1) to evaluate the cross product with ni . We
(Xi , xi ), with 4D (homogeneous) model points Xi and 3D (ho- now linearize the equation by using exp(θξ)ˆ = P∞ (θξ̂)k ≈
k=0 k!
mogeneous) image points xi . Each image point defines a 3D ˆ with I as identity matrix. This results in
I + θξ,
Plücker line Li = (ni , mi ) (projective ray), with a (unit) direc-
tion ni and moment mi [18]. ˆ i )3×1 × ni − mi = 0
((I + θξ)X (11)
Every 3D rigid motion can be represented in an exponential
form and can be re-ordered into an equation of the form Aξ = b.
„ « Collecting a set of such equations (each is of rank two) leads
ˆ = exp ω̂ v
M = exp(θξ) (5) to an overdetermined system of equations, which can be solved
03×1 0
using, for example, the Householder algorithm. The Rodriguez
formula can be applied to reconstruct the group action M from
where θξˆ is the matrix representation of a twist ξ ∈ se(3) = the estimated twist ξ. Then the 3D points can be transformed
{(v, ω̂)|v ∈ R3 , ω̂ ∈ so(3)}, with so(3) = {A ∈ R3×3 |A = and the process is iterated until the gradient descent approach
−AT }. The Lie algebra so(3) is the tangential space of the converges.
3D rotations. Its elements are (scaled) rotation axes, which can Joints are expressed as special screws with no pitch of the
either be represented as 3D vector form θj ξˆj with known ξˆj (the location of the rotation axes as
0 1 part of the model representation) and unknown joint angle θj .
ω1
The constraint equation of a jth joint has the form
θω = θ @ ω2 A , with kωk2 = 1 (6)
ω3
(exp(θj ξˆj ) . . . exp(θ1 ξˆ1 ) exp(θξ)X
ˆ i )3×1 × ni − mi = 0
or as screw symmetric matrix (12)

which is linearized in the same way as the rigid body motion

0 1
0 −ω3 ω2
θω̂ = θ @ ω3 0 −ω1 A (7) itself. It leads to three linear equations with the six unknown
−ω2 ω1 0 pose parameters and j unknown joint angles. Collecting a suf-
ficient number of equations leads to an overdetermined system
In fact, M is an element of the one-parametric Lie group SE(3), of equations.
known as the group of direct affine isometries. A main result Note, that since we work with reconstructed 3D lines we can
of Lie theory is that to each Lie group there exists a Lie alge- gain equations from different cameras (calibrated with respect
bra which can be found in its tangential space, by derivation to the same global coordinate system) and merge them into one
and evaluation at its origin; see [18] for more details. The cor- system of equations and solve them simultaneously. This is the
responding Lie algebra to SE(3) is denoted as se(3). A twist key idea to deal with partial occlusions: A joint that is not visible
contains six parameters and can be scaled to θξ with a unit in one camera must be visible in another one to get a solvable
vector ω. The parameter θ ∈ R corresponds to the motion system of equations. A set of four cameras around the subject
velocity (i.e., the rotation velocity and pitch). For varying θ, covers a large range and allows the analysis of quite complex
the motion can be identified as screw motion around an axis in motion patterns.

Page 3
 In order to deal with larger motions during pose tracking
we use a sampling method that applies the pose estimation al-
gorithm for different neighboring starting positions. From all
results we then choose the one with the minimum error between
the extracted silhouette and the projected surface mesh. This is
important to avoid local minima during tracking.

3 Experiments
Figure 3 shows pose results of a 3-camera sequence. For visual-
ization, the surface patches are transformed (with the estimated
pose and joint angles) and projected onto the images. In this
sequence the subject performs a 360 degrees rotation. It results
in partial occlusions which can be handled from the algorithm.
A lack of many studies is that only a visual feedback about
the pose result is given, by overlaying the pose result with the
image data, e.g. [22, 14]. To enable a quantitative error Figure 5: The coordinate systems of the markers in the lab
setup. Within the Solver Interface 2.0, the markers are (manu-
ally) connected to a skeleton model.

with an 8-Falcon-camera system. For data capture we use the

Eva 3.2.1 software and the Motion Analysis Solver Interface 2.0
for inverse kinematics computing [17]. The solver interface ex-
tracts Euler angles which are transferred to angles representing
the kinematic chain structure of the used surface model. In this
system a human has to wear a body suit and retro-flective mark-
ers are attached to it. Around each camera is a strobe light led
ring and a red-filter is in front of each lens. This gives very strong
image signals of the markers in each camera. These are treated
as point markers which are reconstructed in the eight-camera
system. The system is calibrated by using a wand-calibration
method. Due to the filter in front of the images we had to use
a second camera set-up which provides real image data. This
camera system is calibrated by using a calibration cube. After
Figure 3: A 360 degrees rotation during a 3-Camera sequence. calibration, the offsets and rotations between both world coor-
The algorithm is able to handle partial occlusions. dinate systems are calculated. Then we generate a stick-model
from the point markers including joint centers and orientations.
This results in a complete calibrated set-up we use for a system
comparison. It is visualized in figure 4. Some reconstructed
markers of the Motion Analysis system are shown in figure 5.
The skeletons are connected manually in Eva 3.2.1 by connect-
ing the reconstructed points. Using this setup we then grabbed
Basler Camera a series of test sequences.
Surface model coordinate system
Falcon Camera
Surface model The images in the upper left of figure 1 show the body-suit
Lab coordinate system
with the attached markers. These lead to minor errors during
Stick figure model silhouette extraction, which are omitted here. Figure 6 shows
Stick model coordinate system
the first test sequence, where the subject is just moving the arms
forwards and backwards. The diagram shows the estimated an-
gles of the right elbow. The marker results are given as dotted
lines and the silhouette results in solid lines. The overall er-
ror between both angles diagrams is 2.3 degrees, including the
tracking failure between frames 200 till 250.
Figure 7 shows the second test sequence, where the subject
Figure 4: The coordinate systems in the lab setup. is performing a series of push-ups. Here the elbow angles are
much more characteristic and also well comparable. The overall
analysis, we use a commercial marker based tracking system for error is 1.7 degrees. Both sequences contain partial occlusions
comparison. Here, we use the Motion Analysis software [17], in certain frames. But this can be handled by the algorithm.

Page 4
 0 2

-10
Right Elbow Right Elbow
1.5

-20
1
-30
Angle (Degrees)

Angle (Degrees)
0.5
-40

-50 0

-60
-0.5

-70
-1
Left Elbow
-80

-1.5
-90

-100 Frame -2 Frame

0 50 100 150 200 250 300 0 20 40 60 80 100 120 140

Figure 6: Tracked arms: The angle diagrams show the elbow Figure 7: Tracked Push-ups: The angle diagrams show the elbow
values of the Motion analysis system (dotted) and the silhouette values of the Motion analysis system (dotted) and the silhouette
system (solid). system (solid).

3.1 Tracking in complex environments son of the marker-less approach with a commercial marker based
tracking system. In [20] eight bio-mechanical measurement sys-
While the basic segmentation and pose estimation model is suf- tems are compared (including the Motion Analysis system) show-
ficient to capture the human upper torso under controlled lab ing that the root mean square (RMS) errors are typically within
conditions nearly in real-time, more challenging scenes with clut- three degrees. The errors we achieve with our system fit in this
tered background need the enhanced model with joint segmenta- range quite well. The method thereby achieves almost real-time
tion and pose estimation to yield reliable tracking results. Figure performance.
8 shows an example stereo frame and highlights some proper- We have further presented an extended version of the system
ties of the sequence which are challenging during tracking: the that allows for human motion tracking in more complex environ-
subject is wearing a loose shirt which causes problems in the seg- ments with textured objects, shadows and cluttered background.
mentation and does not fit accurately to the surface model which This shows that our work in [2] can be extended to kinematic
is generated from a person wearing a slim body suit. The hands chains and that we have the foundations for motion capturing
in complex outdoor environments. Marker-less human motion
are sometimes difficult to separate from the background and tracking is highly challenging for sports, exercise, and clinical
clutter offers many possibilities for the contour to be distracted. analysis. The system evaluation of our approach revealed very
Finally, light from the right leads to a brightly illuminated face promising results. Future works will continue with experiments
on one side, while it is dark on the other. Pose results of the on high-speed sports movement tracking in outdoor scenes.
sequence are shown in figure 9. The person is moving in the
scene and shaking his arms. The upper images show the pose
results overlaid with the image data and the lower images show
the segmentation results. The experiment indicates the possi-
bility to track humans even under more difficult conditions with
cluttered background, which is required, for instance, in outdoor
scenes.

4 Summary
The contribution has presented a human motion estimation sys-
tem. It extracts silhouettes by using level set functions. The
silhouettes are used for model fitting with 21 degrees of freedom
in a four-camera set-up. We have further performed a compari- Figure 8: Tracking with cluttered background.

Page 5
 [5] Campbell R.J. and Flynn P.J. A survey of free-form object rep-
resentation and recognition techniques. Computer Vision and
Image Understanding (CVIU), Vol. 81, pp. 166-210, 2001.
[6] Carranza, J., Theobalt, C., Magnor, M. A., and Seidel, H.-
P. Free-viewpoint video of human actors. In Proceedings of
SIGGRAPH 2003, ACM Press / ACM SIGGRAPH, Computer
Graphics Proceedings, Annual Conference Series, ACM, 569–
577, 2003.
[7] Caselles V., Catté F., Coll T. and Dibos F. A geometric model for
active contours in image processing. Numerische Mathematik,
66:1–31, 1993.
[8] Chadwick J.E., Haumann D.R. and Parent R.E. Layered con-
struction for deformable animated characters Computer Graph-
ics, Vol. 23, No. 3, pp. 243-252, 1989.
[9] Chan T. and Vese L. An active contour model without edges.
In M. Nielsen, P. Johansen, O. F. Olsen, and J. Weickert, edi-
tors, Scale-Space Theories in Computer Vision, volume 1682 of
Lecture Notes in Computer Science, pages 141–151. Springer,
1999.
[10] Cremers D., Kohlberger T., and Schnörr Ch. Shape statistics in
kernel space for variational image segmentation. Pattern Recog-
nition, No. 36, Vol. 9, pp. 1929-1943, 2003.
[11] Fua P., Plänkers R., and Thalmann D. Tracking and modeling
people in video sequences. Computer Vision and Image Under-
standing, Vol. 81, No. 3, pp.285-302, March 2001.
[12] Gavrilla D.M. The visual analysis of human movement: A survey
Computer Vision and Image Understanding, Vol. 73 No. 1, pp.
82-92, 1999.
[13] Herda, L., Urtasun, R., and Fua, P. Implicit surface joint limits
to constrain video-based motion capture. In Proc. 8th European
Conference on Computer Vision, Springer, Prague, T. Pajdla and
J. Matas, Eds., vol. 3022 of Lecture Notes in Computer Science,
405–418, 2004.
[14] Mikic I., Trivedi M, Hunter E, and Cosman P. Human body
model acquisition and tracking using voxel data International
Journal of Computer Vision (IJCV), Vol. 53, Nr. 3, pp. 199–223,
2003.
Figure 9: Pose and segmentation results.
[15] MoCap-Systems Motion analysis, vicon, simi: Marker based
tracking systems. www.motionanalysis.com, www.vicon.com,
Acknowledgments www.simi.com/en/, June, 2005.
We gratefully acknowledge funding by the German Research [16] Moeslund, T., and Granum, E. A survey of computer vision
Foundation (RO 2497/1-1 and RO 2497/1-2) and the Max based human motion capture. Computer Vision and Image Un-
Planck Center for Visual Computing and Communication. derstanding 81, 3, 231–268, 2001.
[17] Motion Analysis Corporation www.motionanalysis.com last
accessed February 2005.
References [18] Murray R.M., Li Z. and Sastry S.S. A Mathematical Introduction
[1] Bregler C. and Malik J. Tracking people with twists and to Robotic Manipulation. CRC Press, 1994.
exponential maps. IEEE Computer Society Conference on Com- [19] Osher S. and Sethian J. Fronts propagating with curvature-
puter Vision and Pattern Recognition, Santa Barbara, California, dependent speed: Algorithms based on Hamilton–Jacobi formu-
pp. 8-15, 1998. lations. Journal of Computational Physics, Vol.79, pp. 12-49,
[2] Brox T., Rosenhahn B. and Weickert J. Three-Dimensional 1988.
Shape Knowledge for Joint Image Segmentation and Pose Es- [20] Richards J. The measurement of human motion: A comparison
timation. Pattern Recognition 2005, DAGM, W. Kropatsch, R. of commercially available systems Human Movement Science,
Sablatnig and A. Hanbury (Eds.), Springer-Verlag, Berling Hei- Vol. 18, pp. 589-602, 1999.
delberg, LNCS 3663, pp. 109-116, Wien, 2005. [21] Rosenhahn B., Kersting U., Smith A., Gurney J., Brox T. and
[3] Brox T., Rousson M., Deriche R., Weickert J. Unsupervised Klette R. A system for marker-less human motion estimation
segmentation incorporating colour, texture, and motion In Com- Pattern Recognition 2005, DAGM, W. Kropatsch, R. Sablat-
puter Analysis of Images and Patterns, Springer LNCS 2756, nig and A. Hanbury (Eds.), Springer-Verlag, Berling Heidelberg,
N.Petkov, M.A.Westenberg (Eds.), pp. 353-360, Proc. CAIP LNCS 3663, pp. 230-237, Wien, 2005.
2003 Conference, Groningen, The Netherlands, 2003. [22] Rosenhahn B., Klette R. and Sommer G. Silhouette based
[4] Brox T. and Weickert J. A TV flow based local scale measure human motion estimation. Pattern Recognition 2004, 26th
for texture discrimination. In T. Pajdla and J. Matas, editors, DAGM-symposium, Tübingen, Germany, C.E. Rasmussen, H.H.
Computer Vision - Proc. 8th European Conference on Computer Bülthoff, M.A. Giese, B. Schölkopf (Eds), LNCS 3175, pp 294-
Vision, LNCS 3022, pages 578–590. Springer, May 2004. 301, 2004, Springer-Verlag Berlin Heidelberg.

Page 6
[23] Rosenhahn B. and Klette R. Geometric algebra for pose estima- Bild Dr. Uwe G. Kersting is a lecturer in biomechan-
tion and surface morphing in human motion estimation Tenth In- ics at the University of Auckland (New Zealand).
ternational Workshop on Combinatorial Image Analysis (IWCIA), He has been Director of the Biomechanics Lab-
R. Klette and J. Zunic (Eds.), LNCS 3322, pp. 583-596, 2004, oratory at Tamaki Campus for the previous two
Springer-Verlag Berlin Heidelberg. Auckland, New Zealand, years. He has published in various fields ranging
from sports biomechanics to tissue adaptation.
[24] Magnenat-Thalmann, N., Seo, H., and Cordier, F. Automatic
modeling of virtual humans and body clothing. Journal of Com- He has a broad experience in video based move-
puter Science and Technology No. 19 vol. 5, 575–584, 2004. ment analysis techniques and is currently running
several projects which require three dimensional
[25] Wikipedia Motion capture. motion capture.
http://en.wikipedia.org/wiki/Motion capture, September,
2005.
[26] You, L., and Zhang, J. J. Fast generation of 3d deformable
Bild Dr. Andrew W. Smith obtained his PhD in
moving surfaces. IEEE Transaction on Systems, Man, and Cy-
Biomechanics from the University of Waterloo in
bernetics, Part B: Cybernetics Vol. 33, No. 4, 616–615, 2003.
Canada in 1986. He is currently Senior Lecturer
[27] Zang Z. Iterative point matching for registration of free-form in the Department of Sport and Exercise Science
curves and surfaces. International Journal of Computer Vision, at the University of Auckland, New Zealand. His
Vol. 13, No. 2, pp. 119-152, 1999. research interests are in three-dimensional move-
ment analysis, particularly in human gait and bal-
ance.
Kontakt
Bodo Rosenhahn
Max-Planck-Center Saarbruecken
Stuhlsatzenhausenweg 85
D-66123 Saarbruecken
Germany Bild Jason Gurney received the BSc (Hons) and is cur-
email: [email protected] rently completing honours. He is currently work-
WWW: http://www.mpi-sb.mpg.de/ rosenhahn/ ing as researcher and technician in the Biome-
Thomas Brox chanics Lab, University of Auckland.
Computer Vision and Pattern Recognition Group
Institute for Computer Science, University of Bonn
Roemerstr. 164, 53117 Bonn, Germany
email: [email protected]
Reinhard Klette
Computer Science Department
The University of Auckland
Private Bag 92019, New Zealand
e-mail: [email protected]
Bild Dr. Reinhard Klette is professor of information
WWW: http://www.citr.auckland.ac.nz/ reinhard/
technology in the department of computer sci-
Uwe G. Kersting, Andrew W. Smith, Jason K. Gurney
ence at the university of Auckland (New Zealand)
Department of Sport and Exercise Science
and director of CITR Tamaki (Centre for Image
The University of Auckland, Private Bag 92019
Technology and Robotics). His research interests
New Zealand
are directed on theoretical and applied subjects in
WWW: http://www2.auckland.ac.nz/tmk/ses/
image data computing, pattern recognition, im-
Bild Dr. Bodo Rosenhahn gained his Ph.D. 2003 age analysis, and image understanding. He has
at the Cognitive Systems Group, Institute of published about 300 journal and conference pa-
Computer Science, Christian-Albrechts Univer- pers on different topics within computer science,
sity Kiel, Germany. From 2003-2005 he was and books about parallel processing, image pro-
(DFG-)PostDoc at the University of Auckland. cessing, shape recovery based on visual informa-
Since November 2005 he is senior researcher at tion, and digital geometry.
the Max-Planck center in Saarbücken, Germany.
He is working on markerless motion capture, hu-
man model generation and animation, and image
segmentation.

Bild Dr. Thomas Brox received his M.Sc. degree

in computer engineering from the University of
Mannheim, Germany in 2002 and the Ph.D. de-
gree in computer science from the Saarland Uni-
versity, Saarbrücken, Germany in 2005. Since
September 2005, he has been working with the
Computer Vision and Pattern Recognition Group
at the University of Bonn, Germany. His re-
search interests include image segmentation, op-
tic flow estimation, texture analysis, shape mod-
elling, and pose estimation.

Page 7