Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 101

hundreds of particles are sufficient for tracking an articulated model of the human
body with 29 pose parameters. For three-dimensional body tracking, Schmidt et al.
( 2006 ) employ the 'kernel particle filter', which approximates the probability den-
sity in state space by a superposition of Gaussian kernels. They use 150 particles to
track a three-dimensional model of the upper human body defined by 14 pose pa-
rameters, relying on monocular colour images. The particles are weighted by the use
of colour cues which are combined with ridge and edge cues. Due to the monocular
approach, pose ambiguities may be encountered in this framework.
To alleviate the ambiguity of the three-dimensional pose estimation result of the
human body, Pons-Moll et al. ( 2011 ) propose a method which combines image in-
formation with data acquired by inertial sensors attached to the body. The inertial
sensors provide information about the orientation of the body parts to which they
are attached, which are converted into three-dimensional poses based on an inverse
kinematics approach. This technique allows one to reduce the number of pose pa-
rameters of the utilised human body model, corresponding to 31, to an effective
number of 16 parameters when employing five inertial sensors. Three-dimensional
pose hypotheses consistent with the inertial sensor data are compared by Pons-Moll
et al. ( 2011 ) with the observed image features, especially silhouette information,
using a particle filter framework, where the sensor noise is modelled by the von
Mises-Fisher distribution. The experimental evaluation shows a high accuracy and
robustness of the correspondingly obtained three-dimensional pose estimation re-
sults.
At this point it is illustrative to mention methods for three-dimensional pose es-
timation and tracking of the human hand, which also represents a complex articu-
lated object with a large number of degrees of freedom. From the methodical point
of view, methods for hand pose estimation tend to be fairly similar to many of the
previously described full body pose estimation approaches. The extensive survey by
Erol et al. ( 2007 ) provides an overview of hand pose estimation techniques. They
divide methods for hand modelling into geometric techniques, usually involving a
considerable number of pose parameters, and kinematic approaches that learn typi-
cal dynamical patterns of the hand motion, where they point out that in most systems
a manual user-specific calibration of the kinematic hand model is required. Further-
more, they distinguish between two-dimensional and three-dimensional methods for
hand pose estimation, where the three-dimensional techniques may rely on colour,
edges, point correspondences, disparity information, or actively scanned range data.
A further distinction is made by Erol et al. ( 2007 ) between tracking of single hy-
potheses, typically using a Kalman filter, and tracking of multiple hypotheses, e.g.
involving an extended or unscented Kalman filter or a particle filter. For details refer
to Erol et al. ( 2007 ) and references therein.
The work by Stößel ( 2007 ) is one of the few studies that examine the problem of
three-dimensional pose estimation of articulated objects in the context of industrial
quality inspection. The described system performs a three-dimensional pose esti-
mation of objects which consist of several connected rigid parts, termed 'multi-part
assemblies' by Stößel ( 2007 ), relying on a monocular image, where the correspond-
ing articulated object models are characterised by up to 29 pose parameters. The