In the dynamic tracking mode of a star sensor, the signal noise ratio(SNR) of the shifted star image attenuates resulting in difficulty of dim star extraction. When the number of extracted stars drops to 2, attitude error increases. Obviously for the dynamic star image the aligned pixel gray superposition method was incapable of increasing its SNR remarkably, so a novel inter-frame shifted window gray superposition method was proposed instead. This novel method raised the gray level and the SNR of a dim star by means of superposing pixel gray of inter-frame window area, in a shifted style rather than an aligned one, during which the gray data of the shifted windows of several successive frames were stored into a cache in a pipeline mode. So the peak gray value of a dim star was enhanced and turned to be over the detecting threshold, and no less than 3 stars in the FOV can be finally guaranteed to be used in attitude calculation. The simulation results manifest that the SNR of the dim star image is increased more than 1 times after applying the new algorithm with 4 frames, thus a higher precision can be achieved by the three-star improved attitude determination due to the added one star.