This　paper　investigated　the　feature　extraction　of　multi-channel　four-class　motor　imagery　for　electroencephalogram　(EEG).　A　new　method　which　can　adaptively　extract　features　on　the　basis　of　the　best　wavelet　package　basis　is　proposed　to　solve　the　problem　such　as　the　low　classification　accuracy　and　weak　self-adaptation.　The　traditional　distance　criterion　is　optimized　which　is　under　the　condition　that　the　criteria　is　additive　for　the　choice　of　the　best　wavelet　packet　basis.　And　the　frequency　information　is　filtered　by　OVR-CSP　algorithm　to　improve　the　separability　of　the　feature　information　in　frequency　subbands.　Simulation　results　demonstrate　that　the　proposed　approach　achieve　better　performance　than　other　common　methods.　©　2011　IEEE.