Census transform, one of representative stereo matching algorithms, has a merit that is robust to the
radial distortion and brightness changes between stereo views. In general, census transform compares an
intensity value of central pixel in a matching window with relative intensity distribution of neighbors. So
this method is sensitive to image noise. More specifically, performance of the census transform is
deteriorated as an intensity value of the central pixel is degraded by noise. This paper presents a novel
census transform that performs 5 levels decision instead of binary decision in neighbors intensity
comparison. Conventional census transform is based on central pixel only, but the proposed method divides
a matching window into several sub-windows and employs their average values in intensity comparison
process. When the intensity values of neighbors in the matching window are similar to the average values,
the census transform is difficult to obtain precise stereo matching result. The proposed method examines a
variance of intensity distribution in sub-window and computes relative weight values. Experimental results
showed that the proposed method can achieve better performance than previous census transform.