Abstract: An auto-regressive(AR) model based approach using hybrid features of both geometric and appearance features is proposed to recognize dynamic facial expression in this paper. Six AR models are first trained for six basic expressions based on the six groups of expression sequences. Given one sequence of expression, six predicted sequences are generated using the trained AR models. The corresponding expression is inferred from the most similar predicted sequence to the given one. To incorporate structure information and provide better distinctive capability for recognition, a line segment based method is proposed to compute the similarity between the predicted and given expression sequences. Encouraging experimental results have been obtained on the extended Cohn-Kanade database.