Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 97

model, which are defined by a Gaussian distribution rather than a set of sharply de-
fined quantities. Adapting this 'blurred curve model' (Hanek and Beetz, 2004 )to
the local statistical properties of the pixel grey values yields a large convergence
radius and at the same time a high accuracy of the determined boundary. As a result,
the CCD algorithm is capable of separating objects with ill-defined outlines from a
cluttered background. Section 2.2.3 describes in detail a multiocular variant of the
CCD algorithm developed by Hahn et al. ( 2007 , 2010a ).
Ellenrieder ( 2004 ) proposes a three-dimensional active contour method which in-
corporates a technique similar to the shape from texture approach (Jiang and Bunke,
1997 ) to estimate the normal of a surface displaying a pronounced texture (e.g. the
surface of a tube laminated by textile material) based on spatial variations of the am-
plitude spectrum of the surface texture. In the context of industrial quality inspection
of tubes and cables, Ellenrieder ( 2005 ) introduces a method for three-dimensional
pose estimation of non-rigid objects which is based on the analysis of the contour of
the shadow of the non-rigid object cast on a surface of known shape under known
illumination conditions.
An approach to the computation of the derivatives of the bundle adjustment error
function ( 1.25 ) for non-rigid objects is introduced by Krüger ( 2007 ), who adapts the
model to the image based on a gradient descent scheme. The method relies on the
sign of the gradient of the error function, and its determination reduces to one look-
up per feature for which a correspondence with the image is established in a table
that is computed offline. Once the space of pose parameters is divided into appro-
priate sections, it is sufficient to memorise one bit, denoting the sign of the gradient
of the error function, for each pixel, each pose parameter of the regarded pose es-
timation problem, and each defined section in the pose parameter space. These bit
matrices, which are termed 'gradient sign tables' by Krüger ( 2007 ), have the same
size as the image. The computational complexity of this optimisation approach is
quite low, while its memory demand may become fairly high.
2.2.1.2 Articulated Objects
Most pose estimation approaches regarding articulated objects address the scenario
of human body pose estimation (cf. e.g. Moeslund et al. ( 2006 ) for an introduction
to and overview of the large field of pose estimation and tracking of the human
body).
Many approaches, especially those aiming for gesture recognition in the context
of human-robot interaction, rely on monocular image sequences. A more detailed
overview of such techniques is thus given in Sect. 7.1.2 . As an example, Schmidt
et al. ( 2006 ) adapt a three-dimensional articulated body model consisting of chains
of cylinders to monocular colour images of the person, where the optimisation of
the pose parameters basically relies on skin colour detection as well as on intensity,
edges, and the spatially varying statistical distribution of colour cues. Sminchisescu
( 2008 ) provides a broad discussion of the advantages and limitations resulting from
monocular body pose estimation. Specifically, the problem of ambiguities of the