Brian-computer　interface　(BCI)　has　drawn　great　attention　in　sensorimotor-based　area　for　its　great　potential.　Traditional　methods　focus　on　scalp　electroencephalography　(EEG)　and　detect　the　relation　between　EEG　signals　and　motor　intention　directly,　it　might　lead　to　a　disconnection　command　to　control　end　effector.　Therefore,　it　is　necessary　to　develop　techniques　that　identify　neural　activity　to　reflect　action　with　a　helpful　instruction.　In　the　present　study,　some　research　had　developed　approaches　to　recognize　different　limb　movement　in　source　space.　EEG　source　imaging　technology　was　used　to　map　the　signals　on　the　scalp　to　source　distribution　in　the　cortex.　However,　this　method　brings　in　a　large　number　of　dipoles,　which　may　arise　overfitting　and　time-consuming　in　computation.　In　order　to　solve　the　problem,　a　novel　criterion　based　on　the　improved　Fisher　discriminant　was　applied　for　dipole　selection　in　this　research,　and　only　those　time　series　of　selected　dipoles　were　selected　for　further　analyzed.　Common　spatial　pattern　(CSP)　was　used　for　feature　extraction　and　support　vector　machine　(SVM)　for　classification.　Experimental　results　show　that　few　dipoles　in　the　cortex　could　performe　an　improvement　in　classification　accuracy　compared　with　118　channels　uniform　distributed　on　the　scalp.　©　2018　IEEE.