Much of the time a person's head, feet and hands extend out from the projected shape of their body, forming very salient features in the contour of their silhouette. Consequently, once the initial figure-ground segmentation has been obtained, it is possible to identify these features and thus reliably locate the person's head, hands, and feet.
This is accomplished by ``walking'' around the edge of the segmented foreground region, producing a chain code model of the person's 2D silhouette, and identifying the large protrusions. The geometry of these protrusions is then analyzed using statistically-derived rules to determine which protrusion is the head, which the hands, and so forth [10, 7]. This contour chain-code is also very useful for some applications, because it is a much more compact representation of the person than the segmentation bitmap.