The accurate classification of ERPs is very important for numerous human cognition studies and clinical evaluations. Extracting feature from ERPs is very important due to high dimension of ERPs which includes much information having nothing to do with classification. The principle and weakness of CSP were analyzed and the method to extract spatio-temporal feature by combining AR model and Whiten transformation was proposed. Cognitive experiments were designed to verify our method. Two kind of features were extracted from the data collected from the cognitive experiments separately by spatio-temporal method and CSP, the classifiers were trained both by SVM, and compared the two on effectiveness of classification were compared. The result demonstrates the spatio-temporal feature method is clearly superior to CSP in the classification of ERPs and the precision rate of classification based on spatio-temporal feature method may be over 90% if the parameters are reasonably determined.