The visual tracking has been studied much because it has numerous application in the real-time computer vision system, for instance, human computer interaction(HCI), surveillance and entertainment. In this report, CONDENSATION-based visual tracking algorithm, which has been considered to perform the best in the visual clutter scene, will be investigated, and its simple implementation and the related issues will be discussed.
The tracking problem itself can be defined in various ways depending
on what kind of features are to be tracked and where its result can be
applied. In this report, the visual tracking means Model-based Visual
Object Tracking, and especially the Contour of an Object is
the feature-to-be-tracked. In this setting, Kalman filtering based method
has been preferred over the other methods, and it performs quite well in
less visual clutter scene. In the severe visual clutter scene, however,
the performance of Kalman filtering based method is not satisfactory since
the observation uncertainty about the current configuration
is not unimodal, that is, there can be multiple hypothesis. Since Kalman
filtering carries only the first two moments of the probability density,
it can only work well with Gaussian density, which is unimodal.
CONDENSATION(Conditional Density Propagation) algorithm provides an
alternative approach to this problem. Instead of propagating the first
two moment of the probability density, it propagates the whole probability
density. The obvious question is how to efficiently propagate the density
function when it does not have a closed form representation. CONDENSATION
algorithm uses ``factored sampling'' iteratively to do this. In Section
2, the details of thie algorithm
will be explained.