Sub-pixel centroid algorithm is widely used in spacecraft attitude measurement, astronomical observations, space precision image stabilization control system, space laser communication and other fields. As the detection module of spatial precision stabilization system, the fine guidance sensor detects the jitter of the guide star at a high speed of 100 frames per second, to provide compensation information for the vibration compensation mechanism. It is required that the algorithm has high accuracy as well as high speed real-time performance. In order to meet the needs of the fine guidance sensor, through simulations and experiments, firstly, analyze the influence of spot drift during exposure time on sub-pixel centroid locating accuracy. In order to obtain certain accuracy, the allowed maximum drift speed of spot must be calculated. Then the author put forward an improved real-time grayscale centroid algorithm, which uses Gaussian surface fitting, digital correlation method and grayscale centroid method in different stages of sub-pixel centroid locating. The simulation and experiment results prove that the method can receive balance between efficiency and accuracy.