Finding meaningful robust chunks from driving behavior based on double articulation analyzer

The estimation of human intention is essential to realize intelligent vehicle systems which interact and assist humans to accomplish their tasks. In this paper, we propose a novel method for finding meaningful segments from driving behavior which are important for intelligent vehicle systems that act on human intentions. We assume that contextual information of driving behavior has a double articulation structure and develop a novel method to find meaningful segments. The double articulation analyzer consists of the sticky HDP-HMM which can encode multivariate time series data into sequence of labels and the nested Pitman-Yor language model which analyze sentences written in unknown language morphologically. Effectiveness of our method was evaluated based on real driving data by comparing robust chunks with outside environmental information. It was observed that the extracted robust chunks reflected outside information influential for driving intentions.