Motion Capture - Computer Vision for Visual Effects - page 272

A

B

(a) (b)

Figure 7.12. (a) Attaching tapered cylinders to the skeleton of a kinematic model results in a
solid model of the body. (b) The solid model in a given pose can be used to predict which markers
will be occluded from a given viewpoint. A marker will not be visible if either its normal points
away from the camera (e.g., point A) or a body solid lies between it and the camera (e.g., point B).

from a boxer is used to fit the online behavior of a gymnast). Finally, if our goal is to
capture the subtle mannerisms of a skilled performer, a model-based approach may
too strongly bias the result toward the library of precaptured motion, smoothing out
the subtlety that makes the individual performance unique.

7.4.4 Marker Cleanup with a Kinematic Model

Once we can map a set of 3D marker locations to the joints of kinematic model
and vice versa, the kinematic model can be used to predict which markers will be
occluded from a given pose and viewpoint. For example, we can put “flesh” on the
kinematic skeleton tomake a crude bodymodel out of tapered cylinders or ellipsoids,
as illustrated in Figure 7.12 a. 12 If either the normal vector to the marker points away
from the camera or the line connecting the marker and the camera center passes
throughone of thebody solids, themarkerwill not be visible fromthe givenviewpoint,
as illustrated in Figure 7.12 b. The solid model can also be used to rule out infeasible
configurations in which the limbs inter-penetrate.
Herda et al. [ 194 ] showed how a skeleton and body model could be used to predict
the location of missing markers. The user associates the detected markers to body
parts and joints based on a calibration motion that exercises all the joints. The bone
lengths of the skeleton and the distance between markers and underlying joints are
then fit to the observed marker trajectories using a least-squares approach. During

12 This can be viewed as a crude skinning of the skeleton, that is, a mapping from points on a
skeleton to the surface of a 3D character model [ 29 ]. We mentionmore sophisticated body models
in Section 7.7 and Chapter 8 .