Human behavior recognition based on fractal conditional random field

摘要

In order to meet the demand of visual behavior recognition, we introduce Fractal Conditional Random Field (FCRF) model. FCRF model has improved Latent-Dynamic Conditional Random Field (LDCRF), and proposed the concept of fractal labels that define the integrity and directionality of human behavior. FCRF model overcomes real-time issues of the Hidden Conditional Random Field (HCRF) and the problem of label bias when the behavior transform. The experimental results show that the algorithm proposed in this paper has better recognition performance than Conditional Random Field (CRF), HCRF and LDCRF.