Classifying Behavioral Attributes Using Conditional Random Fields

A human behavior recognition method with an application to political speech videos is presented. We focus on modeling the behavior of a subject with a conditional random field (CRF). The unary terms of the CRF employ spatiotemporal features (i.e., HOG3D, STIP and LBP). The pairwise terms are based on kinematic features such as the velocity and the acceleration of the subject. As an exact solution to the maximization of the posterior probability of the labels is generally intractable, loopy belief propagation was employed as an approximate inference method. To evaluate the performance of the model, we also introduce a novel behavior dataset, which includes low resolution video sequences depicting different people speaking in the Greek parliament. The subjects of the Parliament dataset are labeled as friendly, aggressive or neutral depending on the intensity of their political speech. The discrimination between friendly and aggressive labels is not straightforward in political speeches as the subjects perform similar movements in both cases. Experimental results show that the model can reach high accuracy in this relatively difficult dataset.