Applications to Safe Human–Robot Interaction - 3D Computer Vision: Efficient Methods and Applications - page 301

Fig. 7.18 Positional error of the mean-shift tracking stage (Euclidean distance with respect to
ground truth data (GT)) for the nine test sequences. The '100 %' labels indicate that the method is
able to track the hand performing the working actions completely across all test sequences
requires detailed person-specific modelling, more cameras, much wider baselines,
and less complex backgrounds than the small-baseline approaches proposed in this
section.
7.4.7 Evaluation of the Three-Dimensional Mean-Shift Tracking
Stage
This section describes the evaluation of the mean-shift tracking method described in
Sect. 2.3.2 on the same data set as the one described in Sect. 7.4.2 , where only the
greyscale versions of the images are used. The presentation is adopted from Hahn
et al. ( 2010b ).
In the test sequences, an average number of 6 . 3 objects are tracked simultane-
ously by the mean-shift method. These objects always comprise the right hand of the
person (which performs the working actions). The ellipsoid associated with the right
hand is indicated manually once for the first image of the sequence. When tracking
fails and the hand gets lost, the corresponding ellipsoid is re-initialised based on
the ellipsoid located closest to the last known hand position. The average Euclidean
distances between the estimated hand position and the ground truth data (here: the
coordinates of the wrist point) along with the corresponding standard deviations are
depicted for each test sequence in Fig. 7.18 . Due to the re-initialisation step, our
system is able to track the hand performing the working actions completely across
all test sequences, as indicated by the labels on top of the error bars in Fig. 7.18 .The
average Euclidean distance corresponds to 45-90 mm, the standard deviation to 16-
50 mm. The metric accuracy of the hand position estimated by the three-dimensional
mean-shift approach is thus comparable to the accuracy of the wrist point estimated
by the SF method on the same data set. In the latter work, however, the full three-
dimensional pose of the articulated hand-forearm limb, including internal degrees
of freedom, is determined along with its temporal derivative.
Figure 7.19 illustrates the results of the mean-shift tracking approach for four
test sequences. In part, large values of the Euclidean distance between the estimated
hand position and the ground truth data may result from the fact that the centre of