In a noninvasive brain-computer interface (BCI), EEG feature extraction is a key part for improving classification accuracy and resulting information transfer rate, and it has a crucial and decisive role. In this paper, three different methods were proposed that combine spatial filtering with autoregressive model for EEG feature extraction. Six subjects participated in the BCI experiment during which they were asked to imagine movements of left hand and right hand. Each subject carried out four sessions and each session contained 120 trials. EEG data recordings were used for off-line analysis and the 10 leads around C3 and C4 were chosen for feature extraction. Autoregressive model coefficients and the parameters derived from other three methods were proposed as classification features. Fisher discriminant analysis (FDA) was used as linear classifier. The results show that classification accuracy rates obtained from the three proposed methods are far higher than those acquired from autoregressive model coefficients. At the same time the classification results of each subject are very stable, proving the effectiveness of these novel feature methods.