Conclusion - 3D Computer Vision: Efficient Methods and Applications - page 355

yield rotational pose parameters of an accuracy of a few tenths of a degree. It has
been shown that these accuracies are significantly higher than those obtained with
previously existing methods that mostly rely on edge information. This favourable
behaviour can be attributed to the inclusion of photopolarimetric information. The
accuracy of the determined lateral object position is largely given by the pixel reso-
lution of the image. The depth accuracy obtained from defocus information comes
close to the accuracies reported in the literature for systems performing pose esti-
mation in stereo image pairs.
The three-dimensional pose estimation of non-rigid parts (cables and tubes) has
been demonstrated based on the multiocular ziplock ribbon snake method. It has
been employed in a robot-based system in which the result of pose estimation is
utilised for gripping a non-rigid object. The localisation accuracy of this approach
has been shown to be comparable to the pixel resolution of the image.
The integrated frameworks involving the combination of shadow and shading
features, shape from photopolarimetric reflectance and depth, and specular stereo
have been applied to the three-dimensional reconstruction of rough metallic sur-
faces. The accuracy of depth differences on the surface is typically better than the
lateral pixel resolution of the utilised images. For poorly known reflectance param-
eters, a graceful degradation of the reconstruction accuracy is observed. For the
combination of intensity-based methods with depth data acquired by active range
scanning devices, similar accuracies have been obtained.
A different application scenario is safe human-robot interaction in the industrial
production environment. Existing camera-based safety systems which are either still
under investigation or already commercially available have been discussed. Further-
more, an overview of vision-based gesture recognition methods for human-robot
interaction has been provided. This discussion has led to the conclusion that most of
the existing approaches are not well suitable for safety systems. Their performance
tends to decrease in the presence of a cluttered background, many of them depend
on skin colour cues, which are unsuitable for detecting humans in the industrial pro-
duction environment, and a high accuracy of the localisation and pose estimation
of the hand is often only obtained if a user-specific model is available. Hence, we
have evaluated the proposed methods for segmentation of three-dimensional point
clouds and model-based object localisation and pose estimation in the context of
human-robot interaction. Merely relying on coarse object models, the motion of a
human through an industrial workspace as well as the motion of the hand-forearm
limb through a scene with a cluttered background can be reliably tracked. It is not
necessary to adapt these systems to the individual human being regarded. Similarly,
encouraging results have been obtained for the MOCCD and the shape flow algo-
rithm, which are used for tracking the hand-forearm limb through highly complex
cluttered scenes. Based on the obtained three-dimensional reconstruction results, it
has been demonstrated that a robust automatic recognition of a sequence of working
actions is possible in a realistic industrial environment.
The third application scenario refers to the generation of topographic maps in
the domain of lunar remote sensing. Methods based on radar and laser altime-
try, shadow length measurements, stereo and multi-image photogrammetry, and