Abstract: On the one hand, the unscented Kalman filter （UKF） is an algorithm for recursive state estimation in nonlinear systems by transforming approximations of the distributions through the nonlinear system and observation functions. This transformation is used to compute predictions for the state and observation variables in the standard Kalman filter. In this approach, the distribution is represented by a set of deterministically chosen points, which are called sigma points. These points capture the mean and covariance of the random variables and are propagated through the nonlinear system. On the other hand, interactive multiple model （IMM） filter can deal with system parameter uncertainties and obtain better precision motions. In order to combine UKF and IMM and absorb their primes, starting with both the inherent mechanism of UKF and dynamic state models of human hand, and aiming at improving accurateness of human hand tracking, some theoretical problems unsolved in UKF are firstly discussed and a novel improved UKF based on double unscented transformation （UKFDUT） is put forward. Subsequently, IMM is modified and changed into multiple model（MM）. The research results show that sigma points take on many wonderful features through which some novel approaches can be explored to improve tracking precision, and that using MM for state prediction can reach higher precision than using IMM lonely. The experimental results also demonstrate the effectiveness and satisfactory tracking results.