Graphics Reference
In-Depth Information
onlinemotion capture, the identities of missingmarkers can thus be predicted by the
body model and filled in with reasonable guesses from inverse kinematics until the
markers are reacquired.
A unique consideration for processing motion capture data is the preservation of
foot contact with the ground, or
footplants
. Filling in occluded foot markers using
the methods described so far can produce a perceptually distracting phenomenon
called
footskate
, in which the feet of the resulting kinematic model do not appear
to be firmly planted on the ground (or worse, appear to penetrate or hover above
the ground in a physically impossible way). Footskate can even result from using
inverse kinematics to fit complete motion capture data, since the kinematic model
is a simplified version of how the human body actually works. Kovar et al. [
254
]
proposed an algorithm for removing footskate artifacts by allowing small changes
to the leg bone lengths in the skeleton. Footplant locations are semiautomatically
identified, and an analytic inverse kinematics algorithm is applied to determine the
skeletal model most similar to the original data that still satisfies the constraints. This
can be viewed as a type of motion editing, which we discuss in the next section.
7.5
MOTION EDITING
The goal of motion capture for visual effects is usually to precisely record a per-
former's action. However, it's often necessary to modify the recorded motion in a
way that preserves the personality of the performance but achieves a space-time goal
for animation. We call these
motion editing
problems. For example, we may need
to stitch together multiple motions from the same performer captured at different
times, such as stringing together separately recorded fighting moves. This is a prob-
lem of
motion blending
or
motion interpolation
. We may instead need to extend or
alter the path of a performer's walk, since the motion capture volume may not match
the environment an animated character must traverse. This is a problem of
motion
path editing
.
In this section, we assume that the rawmotion capture data has been transformed
into a time-varying vector of joint angles by means of an inverse kinematics algo-
rithm, as described in the previous section. That is, a given motion capture clip is
represented as
, where
T
is the number of frames in the clip.
At the simplest level, we can treat each of the time-varying parameters
{
θ
(
t
)
,
t
=
1,
...
,
T
}
as
a one-dimensional signal, and apply any one-dimensional signal processing tech-
nique to it, such as filtering. For example, Witkin and Popovi´c[
550
] discussedmotion
warping using functions of the form
θ
(
t
)
i
θ
i
(
)
=
a
i
(
)
θ
i
(
)
+
b
i
(
)
t
t
t
t
(7.30)
Another simple application is to change the frame rate of themotion by fitting splines
through the samples of the joint angles and resampling.
Bruderlin and Williams [
75
] applied multiresolution filtering to motion capture
signals using a one-dimensional Laplacian pyramid (see Section
3.1.2
). This allows
the modification of frequency bands for individual joints to alter the corresponding
motion. For example, amplifying the middle and high frequencies exaggerates the
recorded motion, making it seemmore cartoonish.
