Segmentation is accomplished by classification of each pixel into
one of several models. The models consists of a static world modeled
as a texture surface and a dynamic user modeled as a collection of
gaussian blobs. The Mahalanobis
distance to each class is computed for each pixel. This is a
measure of the likelihood that a pixel can be explained by a class.
Then pixels are assigned to the most likely class. This process produces
a suppoert map:
This support map is then used to update the class models in a robust estimation framework.
Pfinder takes special care to identify and compensate for shadows cast by the user.