Since the motion of the target in the experiment is quite arbitrary, and also the motion learning can not be done within the given time constraint, a very simple dynamic model - the 2-D random walk over shape-space - is chosen with the parameters determined heuristically from the test data. Although the tracker is very much successful in tracking the contour of the target most of the time, it is sometime hesitating to follow some motion due the limitation of the dynamic model. This can be illustrated by figure 4 and figure 5. Figure 4 shows that the motion is reasonably well estimated, while figure 5 shows a bit slow response of the tracker to the given motion.
Figure 4: Samples drifted by the dynamic model -
: Good estimation of the motion
Figure 5: Samples drifted by the dynamic model -
: Delayed estimation of the motion