SIFT Algorithm Optimization And Implementation Towards Real Time Enviroment

SIFT algorithm has been used in pattern recognition as the state-of-the-art in local invariant feature detection and representation field. It is robustness in many varieties such as scale, light, view and rotation. However SIFT algorithm is so complex that cannot meet the real-time requirements in some application such as target tracking and object recognition in a video. It needs to be much faster.My work is organized as follow:I have analyzed in-depth the basic theory of SIFT, and optimized the algorithm from several points to solve the problem of high time consuming.First, propose a new method based on moment to obtain the orientation of the local feature. This method uses the integral calculate instead of gradient calculate to reduce the time spending in calculating the orientation. The results indicate that the time decrease to 1/3 of the original.Second, use binary-descriptors which is introduced in BRIEF algorithm to instead of the gradient direction histogram. Binary descriptors is a bit-string descriptors that obtained by simply comparing the gray value of some pairs of points. Base on the binary descriptors we propose a novel algorithm named B-SIFT. The results demonstrate that B-SIFT has a much lower complexity in generating descriptors and spend 100 times less than SIFT without losing recognition precision.At last, analyze deeply the structure of B-SIFT for image matching. We combine hardware into software to reconstruct the B-SIFT algorithm. In this chapter we implement the scale space part and the matching part on hardware. Because these two parts are suitable to be implement on hardware. And the speed of the B-SIFT algorithm becomes faster.We test our algorithm in many scene such as rotate variety, scale variety etc. And the results indicate that compared to SIFT, B-SIFT spend much less time without losing must more recognition precise and turn out to be more stable, efficient and reliable.