Human Hand Model Calibration Using 3D Reconstruction From Multiple Silhouette Views

As real-time performance of 3D graphics machines is achieved the ability to interact more naturally with synthetic 3D scenes is needed. Humans manipulate objects in a spatial environment most easily with their hands and the same should be true for synthetic 3D scene manipulation. Thus, a system is needed for recognition of human hand position and configuration which does not encumber hand movement. This paper describes a system which computes a 3D reconstruction from multiple camera views of a human hand. .A polyhedron synthetic model of the human hand is then modified to most closely match the size and shape of the reconstructed human hand. Ulti-mately, the calibrated hand model will be used in a real-time model-based hand configuration recognition system. This system will include multiple cameras and back-lighting surrounding a convenient hand working space of perhaps one cubic meter. In order to manage the combinatorial explosion of all possible hand configurations, the hand model is divided into components (fingers, thumb and palm) and all possible configurations of the individual components are stored. Hand configuration recognition will be achieved by hierarchical partial matches of hand model components to the reconstructed user's hand. The search for the best matching hand model configuration will be limited by assuming a maximum possible range of hand component movements from frame to frame. Preliminary results from model-based hand configuration recognition experiments will be shown and methods for real-time implementation of the system will be discussed.