This research was aimed at the feature extraction problem in brain computer interface (BCI). The combination algorithm based on independent component analysis (ICA) and common spatial pattern (CSP) was introduced into this work for exploring frequency domain characteristics from Electroencephalography (EEG). Firstly, a pre-processing step with ICA was applied to remove artifacts, and EEG was filtered through an 8-30 Hz bandpass filter. Secondly, EEG was decomposed into spatial patterns with CSP, which were extracted from two most discriminative populations, and event related desynchronization (ERD)/event related synchronization (ERS) characteristic was extracted with power spectrum analysis. Finally, support vector machine (SVM) was used to classify motor imagery tasks, and good results were obtained. For validation, the motor imagery EEG data provided by BCI Competition 2008-Graz data set B were used, and the results showed that the combination algorithm enhanced the signal-to-noise ratio and extracted discriminative characteristics. It was an effective method for classification recognition.