Extraction of human action elements with transition network of partial time series data modeled by Hidden Markov Model
Kae Doki, Akihiro Torii, Suguru Mototani, Yuki Funabora, Shinji Doki, Kohjiro Hashimoto · 2016
The authors have researched on a method of human action modeling to realize systems such as to support human human operations or watch persons to prevent various kinds of accidents. In order to recognize or support various kinds of human actions, a certain human action model is necessary in these systems. Therefore, we have proposed a modeling method of human actions, which is extracted statistically from enormous data acquired from various kinds of sensors by long-term observation of human actions and situations around persons. However, human action elements composing a model should have been extracted heuristically by a designer. In this paper, an extraction method of human action elements is proposed in order to extract frequent human action elements automatically from acquired data. In the proposed method, a series of acquired time series data is divided into short partial ones, which are modeled by Hidden Markov Models(HMM). Then, the transition network of generated HMMs is constructed based on the likelihood between the original data and each HMM. In the obtained network, a transition sequence with the only one edge is regarded as a frequent human action element. Extraction results with artificial and actual human action data are shown in this paper in order to verify the usefulness of the proposed method.