People monitoring using face recognition with observation constraints

We propose a people monitoring system to recognize people by probabilistic inference methods exploiting low-level facial feature and high-level domain knowledge. In particular, the faces of people in the view of a monitoring camera are first detected and modeled. Optimal recognition of people leaving and entering a closed-room is accomplished by exploiting temporal correlation and constraints among the observed face sequence. The optimality is achieved in the sense of maximizing a joint posterior probability of multiple observations. Experimental results of real and synthetic data suggest the efficacy of the proposed system.