Fast Census Transform-based Stereo Algorithm using SSE2

In this paper, we propose a census transform-based fast stereo algorithm. Most conventional fast stereo algorithms are based on SAD (Sum of Absolute Difference) or NC (Normalized Correlation). However, SAD is often sensitive to noise or illumination changes. Image matching by census transform is known to be robust to radiometric distortion, since the differences in gain and bias between two images will not affect the ordering of pixels within a window. Thus, in this paper, we develop a fast and accurate census transform-based stereo algorithm using the moving window and parallel processing techniques. The resulting system computes 32 disparity of 320 by 240 images at 34 frames per second for the window size of 5x5.