In　this　paper,　autoregressive　(AR)　model　coefficients　and　support　vector　machine　(SVM)　are　used　to　classify　the　motor　imagery　EEG　available　from　the　well-known　BCI　competition　database.　In　order　to　determine　AR　order,　we　use　paired　t-test　to　assess　the　impact　of　AR　order　on　the　classification　precision　of　motor　imagery　EEG.　The　results　show　that　there　is　a　significant　difference　in　the　classification　performance　when　the　different　AR　orders　are　used　to　model　motor　imagery　EEG.　In　this　investigation,　12-order　prevails.　We　try　using　the　method　of　continuous　re-training　the　SVM　classifier　to　improve　the　classification　precision　of　motor　imagery　EEG,　and　the　experimental　results　show　that　the　method　is　feasible　and　effective.