In the last decade, there have been a great deal of researches on hand gesture recognition and its utilization as a human computer interface. Among them, Korean Sign Language (KSL) recognition has been recognized as a promising interface between the deaf and the people who have little knowledge of KSL. For the recognition, researchers usually gather a huge database of gesture instances from various users and then train a recognizer with them. The constructed model by this approach is called Gesturer-Independent Model (GIM). Recognition using this GIM approach has some disadvantages in case when there is severe inter-person difference in KSL. Although best recognition performance is achieved by using Gesturer-Dependent Model (GDM) con structed from training instances of the specific user, it is hard to implement GDM because construction requires the user to provide too many training instances to the system. This burden can be resolved by applying user adaptation technique whose aim is to tune GIM into the user adapted model by using small user-specific data.
There are several techniques appeared in the lituratures. Among them, Maximum a Posteriori (MAP) is widely used in speak adaptation. This method update or adapt model parameters of the state of HMM by summing statistics of given adaptation data and parameters of previous model. The issue of this update, especially with regard to incremental user adaptation, is that when the unobserved state appears in adaptation data, how we update model parameters of them. Because unobserved states is the states that are not observed in the adptation data, there is none of information given for update. For this matter, conventional method updates the model parameters of unobserved states using adaptation data which belongs to other observed states - those are close in the model parameter space - with the assumption that close states in model parameter space may have similar direction of adaptation. Although conventi...