Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 124

iteration step the pose parameters are updated while at the same time some of the
scene points are assigned to the model, based on the distance to the model and the
similarity between the directions between neighbouring points for the data and the
model curve, such that outliers in the data are not taken into account for matching.
The ICP algorithm yields those measured points that can be assigned to the model,
i.e. a scene segmentation, along with the pose parameters.
2.3.1.6 Photogrammetric Approaches
In the domain of photogrammetry, an important application which requires the seg-
mentation of three-dimensional point clouds is the three-dimensional reconstruction
of buildings, mainly from airborne or terrestrial laser scanner data. Rottensteiner
( 2006 ) states that polyhedral models are appropriate for modelling buildings on spa-
tial scales at which topographic mapping is performed, where symmetry constraints
or a priori knowledge, e.g. about the orthogonality of walls or the orientation of parts
of the building, can be exploited. Rottensteiner ( 2006 ) defines one group of com-
mon approaches to building extraction as bottom-up methods, where the geomet-
ric primitives of the building are constructed from the measured three-dimensional
point cloud and are then used to build up a model of the building. This approach
is used by Rottensteiner et al. ( 2005 ) to determine the roof planes of buildings in
three-dimensional point clouds acquired by a laser scanner based on statistical con-
siderations. The second basic class of methods discussed by Rottensteiner ( 2006 )is
top-down reconstruction, where given geometric primitives are adapted to parts of
the three-dimensional point cloud. A priori knowledge can either be incorporated by
'hard constraints', which are always exactly satisfied by the inferred model, or by
'soft constraints', which allow a residual deviation of the adapted model from the
three-dimensional point cloud data.
2.3.2 Mean-Shift Tracking of Human Body Parts
This section describes the mean-shift-based approach of Hahn et al. ( 2010b )to
tracking human body parts, which extracts all moving objects or object parts from
the scene based on a simple ellipsoid model. The presentation in this section is
adopted from that work. Further details are provided by Hahn ( 2011 ).
2.3.2.1 Clustering and Object Detection
Object detection and three-dimensional tracking are based on a scene flow field. We
use a combination of dense optical flow and sparse correlation-based stereo. The
dense optical flow algorithm described by Wedel et al. ( 2008a , 2008b )isusedto