Time-dependent interactive graphical models forhuman activity analysis

Automatically monitoring and classifying human activities is one of the most challenging problems currently faced in machine learning. In this paper, we propose a statistical model aimed at modeling interactions among human subjects, in particular, conversational audio data is here analyzed. The proposed model, called Coupled Hidden Duration Semi Markov Model, takes inspiration from the large literature on Hidden Markov Models and its variants. The novelty introduced by the model is the capability of dealing with interacting state processes, where 1) states that characterize a single process exhibit different time durations, and 2) different processes involved in an interaction are not synchronized, i.e., their states do not begin/end at the same time instants. Comparative synthetical and real data experiments are presented, showing that the proposed model is able to tackle difficult interactive situations, not otherwise manageable by the state-of-the-art algorithms.