Triangulation-Based Approaches to Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 78

evaluation in Sect. 1.6.4 will show that a matching error rate of the stereo algorithm
of 50-70 % still allows a good estimate.
1.6.2 Multiple-plane Hand-Arm Model
To show that the framework proposed in this study is also able to cope with com-
plex objects that may have internal degrees of freedom, we also regard the problem
of three-dimensional pose estimation of parts of the human body. This application
is closely related to the scenario of safe human-robot interaction (Schmidt et al.,
2007 ): In an industrial safety system, e.g. monitoring the collaboration between
a human worker and an industrial robot, false correspondences due to repetitive
structures are rather unfavourable as they cause spurious objects that may result
in an emergency stop of the robot. As an example, we adapt the articulated three-
dimensional model introduced by Hahn et al. ( 2007 , 2010a ), which consists of a
kinematic chain of five truncated cones and one complete cone (cf. Sect. 2.2.3 ), to
a motion-attributed three-dimensional point cloud according to the method by Bar-
rois and Wöhler ( 2008 ) (cf. Sect. 2.3.3 ). Their three-dimensional pose estimation
approach is inspired by the iterative closest point (ICP) algorithm introduced by
Besl and McKay ( 1992 ).
It is assumed that the arm is moving in the scene. The spacetime stereo approach
described by Schmidt et al. ( 2007 ) then yields a motion-attributed three-dimensional
point cloud. This point cloud is clustered in the space spanned by the four coordi-
nates x , y , z , and the epipolar velocity, and the hand-arm model is adapted to the
largest moving cluster (here one might of course use more sophisticated cluster se-
lection techniques). To reduce the complexity of our further analysis, the fairly com-
plex hand-arm model is then approximated by three connected planes representing
the fingers, the hand, and the arm, respectively. These three connected planes still
provide a reasonable representation of the correct disparities in the image regions
associated with the hand-arm limb. The approximation of the full model by a set of
connected planes is acceptable, because the deviation between the simplified model
and the true object in disparity space is negligible compared to the disparity errors
occurring due to the repetitive structures (cf. Sect. 1.6.4 ). In the three-dimensional
reconstruction context regarded in this study, models composed of several connected
partial planes are a fairly general, approximative “tessellated” description of more
complex objects.
1.6.3 Decision Feedback
The initial correspondence analysis is based on the matrix E SSD . In the presence
of repetitive structures the SSD values are small and quite similar for all possible
correspondences, resulting in a large number of false correspondences. Based on the