To improve the rehabilitation efficacy for stroke patient, a brain-computer interface (BCI) rehabilitation system based on motor imagery (MI) is investigated in this paper. Firstly, the Electroencephalogram (EEG) signals are preprocessed by Laplacian spatial filter and Butterworth band pass filter with cutoff frequency of 8Hz~24Hz to enhance the interesting components of the EEG signals. Then, the logarithmic band power (LBP) features are extracted and the motion intentions classification model is established based on linear discriminant analysis (LDA) algorithm. Finally, the correct motion intentions recognition results are converted to executive command of our designed rehabilitation hand. The motion imagery dataset from OpenVibe is used to validate our system. Results show that the rehabilitation hand is driven to effectively actuate the hand to perform the pre-set actions according to the MI recognition results.