Human action recognition based on latent-dynamic Conditional Random Field

Human action recognition is an important area of computer vision research and applications. In this paper, we propose a new state model-based recognition approach based on latent dynamic Conditional Random Field (LDCRF) for action recognition. Combined feature of histograms of oriented gradient (HOG) and histograms of optic flow (HOF) is extracted from each frame. Neighborhood Preserving Embedding (NPE) is employed for reducing dimensions of the combined features. LDCRF model is built based on the probe features and the most likely label can be obtained from the trained LDCRF models. Its performance is tested both on single-person action datasets and human interaction dataset. The experimental results show the effectiveness of our algorithm.