Abstract
In this paper, we form a 3-dimensional hand model with structural constraint and kinematical constraint. By using the constraints, our hand model behaviors more similar to ordinary human hand. In order to estimate human hand state in high degree of freedom space, we proposed separatable state based particle filter (SSBPF). This method tries to separate the states of the variable which we want to estimate into several different parts, and track each part with the particle filter. Combining the hand model and SSBPF, we implement a human hand gesture reconstruction system. Our system may initialize automatically and reconstruct human hand model with a 3D virtual hand in one second with less than 15 degree average error.