Learning Human Activity Containing Sparse Irrelevant Events in Long Sequence

In daily living, person often performs quite differently for finishing the same semantic task, because of her/his mood and the scene state of that time. A large amount of variations are caused by personal petty actions, aimless wandering and additive actions in special scenes, which we term irrelevant events. In this paper, we explore how activities containing sparse irrelevant events can be recognized. We introduce an irrelevant event state into hidden semi-Markov model to cover variations because of irrelevant events. The proposed model remains the partial order of sub-events of the activity of interest, and keeps its discriminability from others with the help of higher-order Markov setting. The experimental results demonstrate the efficiency of the proposed approach for human activity recognition.