Construction of General HMMs from a Few Hand Motions for Sign Language Word Recognition

We propose a method to construct a Hidden Markov Model (HMM) for sign language recognition with a topology which is suitable for a variety of hand motions. First, candidate HMMs are generated from sub-motions extracted from training samples. If we have many and various samples of motions, an optimal HMM can be selected from candidates by the maximum likelihood (ML) method. However, it is difficult to collect many real samples and the ML method with a small number of samples may select a HMM too much specialized for the training samples. The proposed method selects the best HMM for each word by evaluating the performance using real samples and virtual samples generated by an HMM made from real training samples. Virtual samples are generated from a HMM estimated from given real samples. On evaluation of HMMs, they bring a HMM that accepts not only given real samples but also their variations sufficiently similar to the real samples. With experiments, we show the effectiveness of the proposed method.