Using harris corner points to reduce the complexity of a local stereo image matching algorithm

The purpose of a short-baseline stereo image matching algorithm is a calculation of the disparity map of a given rectified stereo pair. Most stereo algorithms presented in the literature achieve this goal by searching all possible disparities in the assumed search range for all pixels. The proposed algorithm consists of two stages. In the first stage detected feature points in the stereo pair images are matched. In the second, a local stereo matching algorithm is used but with a disparity search range limited to a subset of the possible disparities determined by the matched feature points. It is shown that the usage of feature points is an efficient way of limiting the number of disparities searched at the cost of only minor increase in the result disparity map's error. The selected feature points are Harris corners, which can be quite easily computed. Furthermore, the algorithm works on image blocks and each block is processed almost independently. Because ofthat the proposed framework seems suitable for real-time parallel implementation on GPU or in the FPGA technology.