Motion Capture - Computer Animation: Algorithms and Techniques - page 188

triangulated surface model of the performer). The first requires some basic image-processing tech-
niques, simple logic, usually some user input, and sometimes a bit of luck. The second requires camera
calibration and enough care to overcome numerical inaccuracies. The third requires satisfying con-
straints between relative marker positions.
Optical markers can be fashioned from table tennis balls and coated to make them stand out in the
video imagery. They can be attached to the figure using Velcro strips or some other suitable method.
Colored tape can also be used. The markers are usually positioned close to joints since these are the
structures of interest in animating a figure. The difference between the position of the marker and that
of the joint is one source of error in motion capture systems. This is further complicated if the marker
moves relative to the joint during the motion of the figure. Once the video is digitized, it is simply a
matter of scanning each video image for evidence of the optical markers. If the background image is
static, then it can be subtracted out to simplify the processing. Finding the position of a single marker in
a video frame in which the marker is visible is the easy part. This step gets messier the more markers
there are, the more often some of the markers are occluded, the more often the markers overlap in an
image, and the more the markers change position relative to one another. With multiple-marker sys-
tems, the task is not only to isolate the occurrence of a marker in a video frame but also to track a
specific marker over a number of frames even when the marker may not always be visible.
Once all of the visible markers have been extracted from the video, each individual marker must be
tracked over the video sequence. Sometimes this can be done with simple domain knowledge. For exam-
ple, if the motion is constrained to be normal walking, then the ankle markers (or foot markers, if present)
can be assumed to always be the lowest markers, and they can be assumed to always be within some small
distance from the bottom of the image. Frame-to-frame coherence can be employed to track markers by
making use of the position of a marker in a previous frame and knowing something about the limits of its
velocity and acceleration. For example, knowing that the markers are on a walking figure and knowing
something about the camera position relative to the figure, one can estimate the maximum number of
pixels that a marker might travel from one frame to the next and thus help track it over time.
Unfortunately, one of the realities of optical motion capture systems is that periodically one or more
of the markers are occluded. In situations in which several markers are used, this can cause problems in
successfully tracking a marker throughout the sequence. Some simple heuristics can be used to track
markers that drop from view for a few frames and that do not change their velocity much over that time.
But these heuristics are not foolproof (and is, of course, why they are called heuristics). The result of
failed heuristics can be markers swapping paths in mid-sequence or the introduction of a new marker
when, in fact, a marker is simply reappearing again. Marker swapping can happen even when markers
are not occluded. If markers pass within a small distance of each other they may swap paths because of
numerical inaccuracies of the system. Sometimes these problems can be resolved when the three-
dimensional positions of markers are constructed. At other times user intervention is required to resolve
ambiguous cases. See Figure 6.1 for an example of mocap data processing.
As a side note, there are optical systems that use active markers. The markers are light-emitting
diodes (LEDs) that flash their own unique identification code. There is no chance of marker swapping
in this case, but this system has its own limitations. The LEDs are not very bright and cannot be used in
bright environments. Because each marker has to take the time to flash its own ID, the system captures
the motion at a relatively slow rate. Finally, there is a certain delay between the measurements of
markers, so the positions of each marker are not recorded at exactly the same moment, which may
present problems in animating the synthetic model.