Statistical modeling method of human actions expressed by multi-dimentional time series data with Hidden Markov Model
Kae Doki, Takahito Hirai, Akihiro Torii, Kohjiro Hashimoto, Shinji Doki · 2015
In this paper, a modeling method of human actions is proposed in order to realize such systems as to assist human operations have been desired, which must have a certain human action model to recognize or support various kinds of human actions. In the proposed method, a human action model is extracted statistically from enormous data obtained by long-term observation of human actions with sensors, which means only frequent human actions are modeled in this method. In addition, the human action model obtained by the proposed method has high readability, which makes human action analysis much easier. In order to generate a human action model with the previous two features, a human action and a situation around a person are modeled as time series data expressed by Hidden Markov Model(HMM). This is because HMM can efficiently model a time series data with temporal and spatial redundancy. In addition, the relationship between a situation and a human action modeled by HMMs is expressed by If-Then-Rule style explicitly.