On-line human action recognition based on P-LDCNF

To better describe the substructures and external dynamics of human action, a probabilistic method of human action recognition is presented that is based on an improved conditional random field (CRF) model. In our method, feature functions are redefined to describe the conversions between the substructures, and a neural network layer is used to perform a pretreatment process. A probabilistic latent-dynamic conditional neural fields (P-LDCNF) model with transition probabilities is proposed in our method, which can be used for online human action recognition. Specifically, spatial-temporal interesting points are extracted as features in the model. The efficiency of the proposed method was evaluated using unsegmented action sequences from the Weizmann and KTH datasets. The experimental results showed that P-LDCNF compares favorably with CRF, hidden conditional random field (HCRF), and latent-dynamic conditional random field (LDCRF) models for online human action recognition tasks.