In the object recognition based on original SIFT algorithm, the calculation of the feature descriptor is complex and the time of feature point matching is long, so a fast SIFT algorithm is proposed. The fast algorithm takes cumulative gray-scale value and its difference in a concentic circular window to construct a simplified feature descriptor with 16 dimension, and it matches feature points according to a coarse-to-fine pyramid structure when recognising the object. Experiments show that, under the premise of ensuring the recognition accuracy, the computation time of the proposed fast algorithm is reduced two orders of magnitude compared with the original SIFT algorithm, so the proposed algorithm has a good real-time performance.