Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 100

to the approach introduced by Rosenhahn et al. ( 2005 ), combined with a global
stochastic optimisation and filtering stage based on a technique named 'interacting
simulated annealing'.
Hofmann and Gavrila ( 2009 ) suggest an extraction of three-dimensional human
body pose parameters from single images based on a hierarchical matching scheme
similar to the one described by Gavrila and Philomin ( 1999 ) and a simultaneous
adaptation of the parameters of the human body model based on multiple image
sequences acquired from different viewpoints distributed around the observed per-
sons. The results are integrated over time relying on representative sequences of
three-dimensional pose parameters extracted from the acquired image sequences,
which are used to generate a model of the apparent surface texture of the person.
Texture information is used in combination with contour information to arrive at a
final estimate of the three-dimensional pose parameters.
Salzmann and Urtasun ( 2010 ) introduce a method which combines discrimi-
native approaches (such as regression or classification techniques) estimating the
three-dimensional pose of an articulated object and the three-dimensional shape of
an arbitrary surface with the minimisation of a likelihood function depending on the
image information (such as the reprojection error or the distances between edges
generated by the three-dimensional model pose and those observed in the image).
'Distance preservation constraints' are introduced by Salzmann and Urtasun ( 2010 )
into the estimation of the three-dimensional pose parameters. These constraints im-
pose constant distances between reference points, i.e. the joints in the case of an
articulated human body model and the mesh points in the case of a surface model.
Approaches like those introduced by Plänkers and Fua ( 2003 ), Rosenhahn et al.
( 2005 ), and Brox et al. ( 2008 ) determine a single pose which is updated at every
time step. A more refined tracking scheme is proposed by Ziegler et al. ( 2006 ), who
employ an unscented Kalman filter for tracking the pose parameters of the body. An
ICP-based approach to estimate the pose of the upper human body is used, relying
on a comparison of a three-dimensional point cloud generated by the analysis of
several pairs of stereo images with a synthetically rendered depth map, obtained
with a polygonal model of the upper body using the z -buffer technique. The system
of Ziegler et al. ( 2006 ) determines the position of the torso along with the joint
angles of the upper arms at the shoulders and the joint angles characterising the
postures of the forearms relative to the upper arms.
To increase the robustness of tracking, other approaches such as those proposed
by Deutscher et al. ( 2001 ) and Schmidt et al. ( 2006 ) rely on the particle filter ap-
proach introduced by Isard and Blake ( 1998 ) in order to take into account multi-
ple pose hypotheses simultaneously. In this probabilistic framework, the probability
distribution of the parameters to be estimated is modelled by a (typically large) num-
ber of random samples ('particles'). For tracking the human body based on a small
number of particles, Deutscher et al. ( 2001 ) introduce an approach inspired by the
optimisation technique of simulated annealing, termed 'annealed particle filtering',
which allows one to determine the absolute maximum of the (generally multimodal)
probability distribution of the pose parameters by introducing several subsequent
annealing stages. For weighting the particles, edge detection and background sub-
traction are used. A relatively small number of at least 100 and typically a few