This paper proposes a dynamic Bayesian network (DBN) based MPEG-4 compliant 3D facial animation synthesis method driven by the (Evaluation, Activation) values in the continuous emotion space. For each emotion, a state synchronous DBN model (SS_DBN) is firstly trained using the Cohn-Kanade (CK) database with two streams of inputs: (i) the annotated (Evaluation, Activation) values, and (ii) the extracted Facial Action Parameters (FAPs) of the face image sequences. Then given an input (Evaluation, Activation) sequence, the optimal FAP sequence is estimated via the maximum likelihood estimation (MLE) criterion, and then used to construct the MPEG-4 compliant 3D facial animation. Compared with the state-of-the-art approaches where the mapping between the emotional space and the FAPs has been made empirically, in our approach the mapping is learned and optimized using DBN to fit the input (Evaluation, Activation) sequence. Emotion recognition results on the constructed facial animations, as well as subjective evaluations, show that the proposed method obtains natural facial animations representing well the dynamic process of the emotions from neutral to exaggerate.