Applications to Safe Human–Robot Interaction - 3D Computer Vision: Efficient Methods and Applications - page 296

Fig. 7.11 Three-dimensional
pose estimation results for
configuration 1 using
greyscale images. The labels
above the error bars denote
the fraction of images on
which the object is tracked
truth positions of the reference points amounts to 50-100 mm for sequences 1-5
and to 120-250 mm for sequences 6-9. The standard deviations correspond to 50-
100 mm. For several sequences the tracking process is interrupted and the object is
lost rather early, as illustrated by the percentages indicated on the labels above the
error bars.
The average positional error of configuration 2 (SF algorithm) on greyscale im-
ages (cf. Fig. 7.12 ) is somewhat lower than that of configuration 1 for all sequences,
and the fraction of images in which the object is tracked is always higher or com-
parable. The component-wise average velocity errors of configuration 2 show no
systematic offsets. The standard deviations of the velocity errors are slightly lower
for sequences 1-5 than for sequences 6-9, corresponding to about 10 and 15 mm per
time step, respectively. When colour images are used instead of greyscale images,
the average positional error decreases to below 50 mm for sequences 1-5, where the
object is tracked across all images, but remains largely unchanged for sequences 6-
9(cf.Fig. 7.13 ). The standard deviations of the component-wise velocity errors are
also similar to those obtained for greyscale images.
For configuration 3 (ICP algorithm with constraint line-based motion analysis)
the average positional errors obtain values between 50 and 200 mm with standard
deviations between 50 and 100 mm (cf. Fig. 7.14 ). In comparison with configura-
Fig. 7.12 Spatio-temporal three-dimensional pose estimation results for configuration 2 using
greyscale images