Simple to complex dictionary learning for action recognition

Fang Liu, Feifei Qi, Xiaofen Xing, Kailing Guo

Abstract: Complex human action recognition has gained extensive attention in recent years. However, Own to view point change and other factors, recognizing complex actions is greatly challenging. Considering each complex action can be decompose into a sequence of simple actions, and dictionary learning has gained promising performance. Thus, a simple to complex action dictionary learning model (SCA-DLM) is proposed for complex action recognition. The proposed model used a sequence of simple actions to help complex action learning. More importantly, we unify the simple action and complex action into an objective function for learning a sparse dictionary for complex actions. The objective function learns simple action dictionary shared by different actions to model action-shared features. This approach represents the complex actions using complex action dictionary and the simple action dictionary. The extensive experiments on the complex action datasets demonstrate that SCA-DLM outperforms state-of-the-art approaches.