Discriminative Random Fields for Behavior Modeling

This paper proposed an approach of human behavior modeling based on discriminative random fields. In this model, by introducing the hidden behavior feature functions and time window parameters, the Classical CRFs models was extended to spatio-temporal fields. And feature templates were designed to capture the dynamics of human motions. Due to the conditional structure, this model can accommodate arbitrary overlapping features of the observation as well as long-term contextual dependencies among observations. Behavior recognition method was designed in the experiments. And the results proved that the proposed modeling method performed over than HMM and CRF for human behavior modeling.