A memory-efficient hardware architecture for real-time feature detection of the SIFT algorithm (abstract only)

Abstract

The SIFT (Scale Invariant Feature Transform) is a most popular image processing algorithm that has been widely used in solving image matching related problems. However, SIFT is of high computational complexity and large memory requirement that prevent it from being applied to applications that are unable to offer large on-chip memory. Based on the analysis of the memory requirement of SIFT feature detection, a novel memory access strategy is proposed to reduce the hardware memory usage. In addition, to achieve real-time performance of high resolution video streams, dedicated hardware architecture with multi-pixel based processing scheme has been developed. Compared with conventional designs, our design achieves hardware memory reduction of at least 58.8%.