Exploiting Dynamic Dependencies Among Action Units for Spontaneous Facial Action Recognition
Yan Tong, Qiang Ji · 2015
This chapter proposes to systematically model the dynamic properties of facial actions including not only the temporal development of each action unit (AU), but also the dynamic dependencies among AUs in a spontaneous facial display. Specifically, a dynamic Bayesian network (DBN) is employed to explicitly model the dynamic and semantic relationships among AUs, where the dynamic nature of facial action is characterized by directed temporal links among AUs; and the semantic relationships are represented by directed static links among AUs. The DBN model is automatically constructed using both domain knowledge and the training data. Given the dynamic model, AUs are recognized through probabilistic inference over time. Experiments with real images demonstrate that by explicitly modeling the dynamic dependencies among AUs, the proposed method improves AU recognition over the existing methods, especially for spontaneous facial displays.