Human action classification based on sequential bag-of-words model

Recently, approaches utilizing spatial-temporal features have achieved great success in human action classification. However, they typically rely on bag-of-words (BoWs) model, and ignore the spatial and temporal structure information of visual words, bringing ambiguities among similar actions. In this paper, we present a novel approach called sequential BoWs for efficient human action classification. It captures temporal sequential structure by segmenting the entire action into sub-actions. Each sub-action has a tiny movement within a narrow range of action. Then the sequential BoWs are created, in which each sub-action is assigned with a certain weight and salience to highlight the distinguishing sections. It is noted that the weight and salience are figured out in advance according to the sub-action's discrimination evaluated by training data. Finally, those sub-actions are used for classification respectively, and voting for united result. Experiments are conducted on UT-interaction dataset and Rochester dataset. The results show its higher robustness and accuracy over most state-of-the-art classification approaches.