Abstract: Improvements and parallelization in the aspects of creating differences of gaussian and locating sub-pixel keypoints were made against the problem that the dimension of descriptors from variant SIFT-GLOH(Gradient Location Orientation Histogram) is too high to meet the need of real time. The algorithm was implemented in multi-threads on the GPU hardware with CUDA(Compute Unified Device Architecture). On one hand, the information loss of keypoints caused by PCA was avoided. On the other hand,the speed of registration sacrificed the need of real time in engineering.The programs were compiled by C language & CUDA on VS2005 platform. The result shows that the ratio of correct matching point pairs and registration speed are both promoted greatly.

Key words: SIFT,GLOH,GPU,CUDA