Skip to Main Content
We present an agent-based framework for human hand and head tracking. The benefits of this architecture are that it simplifies system organization, facilitates flexible extension, and supports hierarchical abstraction. The framework employs low-level processing agents that computes motion-vector clusters, skin-color patches and edge. Mid-level agents fuse the features computed into hand and head candidates. Finally, high-level agents select among these candidates to yield consistent interpretations of hand and head groupings. Hence more domain-specific knowledge is encoded into the agents as we ascend the agent hierarchy. We demonstrate the effectiveness of our system by defining a set of gestures (move, zoom, and rotate) that were recognized and tracked.