Pfinder goes beyond those systems by also building statistical models of the person's clothing, head, hands, and feet. This multi-class approach, as described in Chapter 2, provides not only more robust tracking, but also a richer description of the scene as compared to single-sided classification techniques. The statistical, region-based nature of features allows the tracking to proceed, stably, at interactive speeds, without special purpose hardware. Chapter 4 provides some concrete information about estimation stability and execution performance.
The statistical nature of the tracking algorithm makes it possible to include a priori knowledge about the nature of subjects to be tracked. Chapter 3 discusses how this knowledge, as well as heuristic techniques, can be combined to provide automatic, quick, and reliable initialization and error recovery.
Pfinder's performance, and stability, have resulted in it being utilized by a number of whole-body interaction applications. Chapter 5 provides an overview of some of these systems.