Abstract: SIFT algorithm has invariance of rotation,scale and translation,so it is used widely in the field of image matching and 3D reconstruction．But the SIFT algorithm is complicate,making the processing speed slow,difficult to meet the application of high real-time requirements.On the basis of analysis of the principle of SIFT algorithm,in view of the large numbers of extracted features,high-dimensional of the feature vector,we refined the algorithm for parallel processing to take advantage of modern graphics hardware,and compared it with the CPU SIFT algorithms．Experiments demonstrate that the algorithm based on GPU SIFT significantly increases efficiency,and can reach the speed ratio of more than ten times averagely.

Key words: GPU,SIFT,CUDA,Feature matching