Temporal order information for complex action recognition

Abstract

Exploiting simple actions to recognize complex actions instead of using complex actions as training samples can save labor expenses and time consumption. Each complex action is composed of a sequence of simple actions and different manners of combinations of simple actions can form different complex actions. Thus, in this paper, we focus on temporal order information (TOI), which can be used to improve the performance of complex action recognition. To utilize temporal information, a TOI model is proposed for complex action recognition so that users can utilize TOI for extracting features. Classifiers learned from simple actions are used to obtain classification scores that are concatenated as the complex actions final features. The proposed approach is evaluated on two complex action datasets: an Olympic Sports dataset and a YouTube Action dataset. The experiment results showed the effectiveness of the proposed method for complex action recognition.