Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 99

time step, the three-dimensional pose of the body model which is consistent with
the optical flow can be predicted for the subsequent time step. A priori knowledge
about likely and unlikely configurations of joint angles is used to impose constraints
on the pose by learning the probability distribution of the joint angles from ex-
amples and incorporating it as an a priori probability into the Bayesian inference
scheme.
Rosenhahn et al. ( 2008a ) propose a method for tracking the motion of dressed
persons by integrating a kinematic simulation of clothes covering parts of the body
into a silhouette-based system for body pose estimation. The body pose is deter-
mined by minimising an appropriately defined error function, where the correspon-
dences between the observed and modelled silhouettes, the parameters of the kine-
matic chain defining the body model, the appearance of the clothes on the body, and
the forces exerted on the clothes are taken into account. A fairly detailed modelling
is performed, since parameters of the clothes of the person such as the length of
a skirt are extracted during the optimisation process, and physical kinematic mod-
elling of the motion of the clothes resulting e.g. from wind is performed. A quan-
titative evaluation demonstrates that despite the fact that parts of the tracked body
are occluded by clothes, the error of the proposed method is less than one degree
higher than typical inaccuracies of tracking systems relying on markers attached to
the person.
Grest and Koch ( 2008 ) adapt a three-dimensional body model consisting of rigid
fixed body parts to a three-dimensional point cloud extracted from a pair of stereo
images with a dynamic programming-based dense stereo technique. A maximum
number of 28 pose parameters is estimated for the human body model using a
3D-3D pose estimation technique based on the ICP algorithm. The Gauß-Newton,
gradient descent, and stochastic meta-descent optimisation methods are compared
with respect to their convergence behaviour, where the Gauß-Newton method is
found to be the superior approach.
A markerless system for three-dimensional body pose estimation specifically de-
signed for the distinction between normal and pathological motion behaviour of
a person is described by Mündermann et al. ( 2008 ). It is based on an ICP tech-
nique involving an articulated surface model designed such that the exact positions
at which the articulated motion is performed within the joints can be refined (within
certain limits) during the model adaptation procedure, where multiple (between 4
and 64) images of the scene acquired from viewpoints distributed around the person
are used. Segmentation of the person from the background is performed by apply-
ing an intensity and colour threshold to the background-subtracted images, which
yields a three-dimensional visual hull of the person to which the articulated model is
adapted. A direct comparison to a marker-based body pose estimation system yields
accuracies of 10 . 6

± 7 . 8 mm, 11 . 3 ± 6 . 3 mm, and 35 . 6 ± 67 . 0 mm for the full body

and 8 . 7 7 . 6 mm for the lower limbs of the

human body for 64, 8, and 4 cameras, respectively.
Gall et al. ( 2009 ) rely on silhouette and colour features and assume the existence
of an accurate three-dimensional model of the analysed human body. They utilise a
local optimisation technique for three-dimensional pose estimation which is similar

± 2 . 2 mm, 10 . 8 ± 3 . 4 mm, and 14 . 3 ±