Both improper initialization and fake Gaussian components are critical problems in GMM-based foreground detection. The former can lead to a poor local maximum, while the latter invokes unhandled disturbance. To eliminate these destructive impacts, two kinds of feedback knowledge are introduced: positive and negative prior. For appropriate initialization, high level modules provide the positive prior informations by outlining the rough foreground objects using optical flow. Moreover, the negative prior evidences in form of Dirichlet distribution are adopted to suppress the fake Gaussian components when coping with dynamic scenes. Experiments demonstrate that our method outperforms most counterparts.