Applications to Safe Human–Robot Interaction - 3D Computer Vision: Efficient Methods and Applications - page 303

7.4.8 Discussion
In this section the experimental investigations by Hahn et al. ( 2010a , 2010b )have
been described, which were performed on real-world image sequences displaying
several test persons performing different working actions typically occurring in an
industrial production scenario. For evaluation, independently obtained ground truth
data are used. Mainly due to the cluttered background and fast discontinuous move-
ments of the test persons, the individual algorithms presented in this study show a
good performance only on some of the test sequences, encountering difficulties on
others. However, due to their orthogonal properties, a fusion of the algorithms has
been shown to be favourable. The proposed system with the highest performance
takes advantage of a fusion between the ICP algorithm, the MOCCD algorithm, and
the SF algorithm. It is able to track the hand-forearm limb across all images of all
our small-baseline test sequences with an average positional error of 40-100 mm
with standard deviations between 20 and 50 mm and standard deviations of the
component-wise velocity errors of 6-10 mm per time step. The evaluation of the
mean-shift tracking method, which estimates the positions of moving objects in the
scene without providing detailed pose information, on the same image sequences
has demonstrated that it is of comparable positional accuracy.
The reasonable accuracy and high robustness are obtained even though (i) the
image sequences used for evaluation all display a cluttered background, (ii) the sys-
tem has to adapt itself to several test persons wearing various kinds of clothes, and
(iii) the movements of the test persons may be rapid and discontinuous. Hence, the
described three-dimensional pose estimation and tracking framework is a step to-
wards collaborative working environments involving close interaction between hu-
mans and machines.
7.5 Recognition of Working Actions in an Industrial
Environment
This section describes the methods developed by Hahn et al. ( 2008a , 2009 , 2010b )
for the recognition of working actions in a realistic industrial environment based on
the combination of the MOCCD and the shape flow algorithm (Hahn et al., 2010a )
or relying on the three-dimensional mean-shift tracking stage (Hahn et al., 2010b ).
An experimental evaluation on a limited but realistic data set is also performed.
Further details are provided in Hahn ( 2011 ).
In the system described in Hahn et al. ( 2008a ), trajectories are generated with the
MOCCD-based approach of Hahn et al. ( 2007 , 2010a ) (cf. Sects. 2.2.1.2 and 7.4 ).
After a normalisation procedure with respect to the length of the trajectory and
its position and orientation in space, classification is performed with a nearest-
neighbour approach. Furthermore, the classification result is used for a prediction of
the hand position based on the corresponding reference trajectory, which is demon-
strated to be significantly more accurate than a Kalman filter-based prediction.