Most　of　the　existing　studies　on　3D　Facial　Expression　Recognition　(FER)　are　messaged-based　approaches,　which　only　detect　the　already　known　six　universal　expressions.　In　this　paper,　we　describe　the　group　of　global　and　local　features　used　to　comprehensively　characterize　facial　activities.　These　features　are　further　used　to　train　Statistical　Feature　Models　(SFMs)　associated　with　each　Action　Unit　(AU).　The　occurrence　probability　of　a　specific　AU　on　an　input　textured　3D　face　model　is　then　computed.　The　results　demonstrate　that　the　evidence　of　AUs　is　of　importance　for　applying　AU　space　to　evaluate　expressions.　©　Copyright　owned　by　the　author(s).