Towards a real-time high-definition depth sensor with hardware-efficient stereo matching
Ke Zhang, Guanyu Yi, Chao-Kang Liao, Christine S. Lin, Hsiu-Chi Yeh, Rudy Lauwereins, Luc Van Gool, Gauthier Lafruit · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2012
In this paper, we present a hardware design of stereo matching for depth extraction. It is capable of matching high-definition images at real-time speed, producing high-quality disparity maps. The stereo algorithm comprises three major components, i.e. census transform, dynamic programming and local-adaptive voting. The census transform makes our method robust to radiometric differences. The dynamic programming method propagates information along scanlines by imposing a smoothness criterion. The local-adaptive voting further improves the matching quality, utilizing the correlation between intensity images and disparity maps. The whole algorithm is prototyped on a FPGA platform, effectively harnessing the power of parallel computing. With a working frequency of 65MHz, our design can produce high-quality 1024×768 disparity maps at 60fps.