Normalized scoring of Hidden Markov Models by on-line learning and its application to gesture-sequence perception

A normalized scoring algorithm has been developed for hidden Markov models (HMMs) to establish independent individual-model evaluation of each input sequence. Using this model, it has become possible for each trained HMM to judge if an input sequence is classified to the category of the model by a simple thresholding operation without referring to other models. Such evaluation has been enabled by creating a self model for each input sequence by on-line learning. As a result, a long action sequence composed of unit-motions can be recognized using multiple models each trained for each unit-motion. The algorithm was evaluated in 120 test sessions and the recognition rates of average 92.3% and 85.3% for unit motion detection and entire sequence recognition, respectively, have been demonstrated.