In order to solve the problem that the classification recognition rate of the multi-class motor imagery EEG signals is low and that there is difference among subjects in the brain-computer interface, we propose a feature extraction method based on the ERS/ERD phenomenon for the two level common spatial pattern. Firstly, we select the EEG of all channels and make wavelet packet de-noising and decomposition (WPD) of specific frequency bands. Secondly, we conduct common spatial pattern (CSP) on the signals of reconstructed decomposition coefficients to obtain spatial filtering devices for the two classes of hands (left, right) and feet (feet, tongue), and use the 2-norm screening criteria to extract N leads of the heavily weighted factor. Thirdly, the projection matrix of the optimized leads are used to filter hands (left, right) and feet (feet, tongue), and the signals which are regard as the original signals are used to conduct two layers common spatial pattern. Finally, feature vectors are categorized by the support vector machine (SVM). The highest classification accuracy rate of the simulation on three subjects from BCI2005IIIa reaches 92.55%, and the simulation results show that the proposal has a good effect on the feature extraction of EEG signals.