Applications to Safe Human–Robot Interaction - 3D Computer Vision: Efficient Methods and Applications - page 292

Fig. 7.8 Typical
three-dimensional pose
estimation results for the head
and shoulders part of a person
using system configuration 1

tion results for the head and shoulders part of a person obtained using a correspond-
ingly adapted articulated object model are shown in Fig. 7.8 .
In all sequences, the background is fairly cluttered, and the contrast between the
persons and the background tends to be low. The persons are wearing various kinds
of clothes with long and short sleeves, with and without work gloves. For each
system configuration, the parameter setting is the same for all sequences, respec-
tively.
The image sequences were acquired with the trinocular colour camera system
shown in Fig. 7.6 , where the horizontal and vertical baseline amounts to 150 mm.
The time interval between subsequent image triples amounts to t = 71 ms. Each

sequence consists of at least 300 image triples. The average distance of the test per-
sons to the camera system varies from 2 . 7mto3 . 3 m. The ground truth data consist
of the coordinates of three reference points in the world coordinate system, which
correspond to the points W p 1 , W p 2 , and W p 3 of the three-dimensional model of
the hand-forearm limb. To extract the ground truth, three markers were attached
to the tracked limb. The positions of the markers in the images were measured
with the chequerboard corner localisation routine by Krüger and Wöhler ( 2011 )
(cf. Sect. 1.4.8 ), and their three-dimensional coordinates were determined based on
bundle adjustment. 1

7.4.3 Fusion of the ICP and MOCCD Poses

The structure of the fusion stage, which combines the ICP and the MOCCD re-
sults, is shown in Fig. 7.9 . First, the initial pose parameters are used for both pose
estimation algorithms to start their respective optimisation. To verify and rate the
calculated pose parameters we use the quality measures described in Sect. 2.2.3.5 .
These three criteria are calculated for the pose results of both algorithms separately.
Afterwards, the weighting values are summed to a weighting factor w ICP for the ICP
result and another weighting factor w MOCCD for the MOCCD result according to

w ICP = σ p | T ICP + σ o | T ICP + σ c | T ICP (7.1)

w MOCCD = σ p | T MOCCD + σ o | T MOCCD + σ c | T MOCCD . (7.2)

1 The image sequences and ground truth data are accessible at http://aiweb.techfak.uni-bielefeld.de/
content/hand-forearm-limb-data-set .