Applications to Safe Human–Robot Interaction - 3D Computer Vision: Efficient Methods and Applications - page 297

Fig. 7.13 Spatio-temporal three-dimensional pose estimation results for configuration 2 using
colour images
Fig. 7.14 Spatio-temporal three-dimensional pose estimation results for configuration 3 using
greyscale images
tion 2, the fraction of images in which the object is tracked is lower for sequences 1,
2, 5, and 6 but higher for the more 'difficult' sequences 7-9. The standard deviations
of the component-wise lateral velocity errors are lower than for configuration 2 and
correspond to 7-10 mm per time step for all sequences. However, configuration 3
does not determine the velocity component along the depth axis.
A tracking of the object across all images of all test sequences is achieved by
configuration 4 (fusion of ICP and MOCCD, cf. Fig. 7.15 ). The average positional
error amounts to 40-70 mm for sequences 1-5 and 80-100 mm for sequences 6-9
with standard deviations between 20 and 50 mm. The standard deviations of the
component-wise lateral velocity errors correspond to 6-10 mm per time step. Con-
figuration 4 does not estimate the velocity component along the depth axis.
The accuracy achieved by configuration 5 (fusion of ICP and SF) is very simi-
lar to that of configuration 4 (cf. Fig. 7.16 ). Again, the object is tracked across all
images of all test sequences. As additional information, configuration 5 reveals the
velocity component along the depth axis, which is determined at an accuracy of
8-15 mm per time step.