Abstract: Brain-computer interface (BCI) refers to a communication/control channel between human and computer, which reflects human intention in the form of EEG signal then converts it into control signal. Two kinds of EEG signals of imaginary motions are classified, to which the feature extraction combining the wavelet packet decomposition (WPD) with common spatial pattern (CSP) is presented for EEG signals. Different wavelet packets are used to decompose the multi-channel EEG signals in training set, whereas CSP can extract the features of EEG signals from the subbands at different decomposed levels. Furthermore, the K-nearest neighbors (KNN) is used to classify the different features thus extracted to obtain the optimum wavelet function and relevant subband parameters which are used to classify test data. Simulation results showed that if the wavelet packet function is db4 with 4 decomposed levels and 8 feature points are selected, the highest classification accuracy can be up to 96%.