Page 20 - Human Inspired Dexterity in Robotic Manipulation

16 Human Inspired Dexterity in Robotic Manipulation

the positions of joint centers on the two-dimensional image. In these
methods, joint-center position in the normal direction to the image may
not be accurate, because it cannot be measured directly from the image,
and thus it is estimated from a predefined regression model [12].
In the medical images-based methods, multiple MRI images or CT
images are acquired from the same subject in different postures (joint angles),
and bone surfaces geometries are extracted from these images by applying
image segmentation. The procedure to compute the joint centers is as fol-
lows. The proximal bone of the joint target in different postures are regis-
tered first, then transformation matrices that describe relative movement of
the distal bone with respect to the proximal bone are computed, and finally,
a joint center for this joint is determined by finding an immovable point of
the transformation.
The MoCap data based methods utilize hand motions collected using a
MoCap system that measures three-dimensional coordinates of retro-
reflective spherical markers. At least three markers attached to each link
are enough to describe relative movements of the distal bone with respect
to the proximal bone. However, this is not a practical solution, because more
than fifty markers are necessary to capture the entire movement of the link
structure. Usually, to minimize the number of markers attached to the hand,
only one marker at the joint is attached, and the position of the marker is
measured while the subject is performing the calibration motion for each
joint to estimate the joint centers as an immovable point regardless of the
joint movement.
The link length is calculated as the distance between two adjacent joint
centers. The link length of the most distal segment of the finger is calculated
as the distance between the fingertip and the DIP joint center. In the digital-
hand model, the link length is always assumed to be constant.
In terms of reliability, the second approach generates the best link struc-
ture, because the medical images have three-dimensional information, and
unlike the third approach, it is not affected by skin movement artifact [11].
However, the third approach is the most common, because the MoCap is
also used in the mechanical analysis for synthesizing hand posture and
motion after creating the link and joint model.

2.2.3 Surface Mesh of a Digital Hand
Like the second approach in determining joint centers, a surface mesh of a
digital-hand model is usually created by applying image segmentation to