Pose clusters are learned from a silhouette manifold for a variety of highly articulated human poses. A pose transistion model is used for tracking. See the 2007 ICCV paper Detection and Tracking of Multiple Humans with Extensive Pose Articulation for more details.
Body parts are detected frame by frame. The relationship between parts is based on the differences in position, size and appearance. These are combined using the average affinity of all common visible parts after using occlusion analysis to identify visible parts. A motion trajectory is initialized when enough evidence is obtained from the detectors and terminated when the object is not detected for some time. See Human Pose Tracking, Results and the 2009 PAMI paper Human Pose Tracking in Monocular Sequence Using Multilevel Structured Models
The tracking system alone is not the final result, it is important in its use for Event Recognition. More examples are results are given in the Event Recognition sections.