Frame-rate Robust Stereo on a PCI Board

Many vision applications would benefit from framerate, dense stereo. In this paper we present a robust algorithm implemented on a single board for a PC. The algorithm used is census transform correlation, which is well-suited for hardware implementation. Unlike existing frame-rate stereo methods, census transform correlation is robust to outliers, and does not assume that corresponding points have constant brightness. We provide an empirical comparison between our method and other frame-rate stereo algorithms using real data with ground truth. The stereo computation is performed on a reconfigurable computer built from programmable logic on a standard PCI card. The resulting system computes 32 stereo disparities on 320 by 240 pixel images at 46 frames per second. This is the highest performance stereo vision system described to date.