The　feature　extraction　of　Motor　Imagery　Electroencephalography　(MI-EEG),　as　a　key　technique　of　brain　computer　interface　system,　has　attracted　increasing　attention　in　recent　years.　Because　of　the　high　temporal　resolution　of　MI-EEG,　researchers　are　usually　bedeviled　by　the　curse　of　dimensionality.　Some　manifold　learning　approaches,　such　as　Isometric　Mapping　(ISO-MAP)　and　Local　Linear　Embedding　(LLE)　etc.,　have　been　applied　to　dimension　reduction　of　MI-EEG　by　modeling　the　nonlinear　intrinsic　structure　embedded　in　the　original　high-dimensional　data.　However,　these　methods　are　difficulty　to　exactly　represent　the　nonlinear　manifold,　affecting　the　classification　accuracy.　The　Maximum　Variance　Unfolding　(MVU)　can　solve　this　problem,　but　it　is　unsuitable　for　online　application　due　to　the　computation　complexity.　In　this　paper,　a　novel　feature　extraction　approach　is　proposed　based　on　the　Landmark　version　of　Maximum　Variance　Unfolding　(L-MVU).　First,　the　MI-EEG　signals　are　preprocessed　according　to　the　event-related　desynchronization　(ERD)　and　event-related　synchronization　(ERS).　Then,　L-MVU　is　used　to　extract　the　nonlinear　features,　and　a　joint　optimization　of　parameters　is　performed　by　using　the　traversing　method.　Finally,　a　back-propagation　neural　network　is　selected　to　classify　the　features.　Based　on　a　public　dataset,　some　experiments　are　conducted,　and　the　experiment　results　show　that　L-MVU　can　preserve　more　information　and　perfectly　extract　the　nonlinear　nature　of　original　MI-EEG,　and　reduce　the　redundant　and　irrelevant　information　by　introducing　the　landmark　points　as　well,　yielding　a　higher　classification　accuracy　and　a　lower　computation　cost.　Furthermore,　the　proposed　method　has　a　better　effect　on　feature　visualization　with　an　obvious　clustering　distribution.