(A) time efficient real-time stereo matching algorithm using spatial redundancy removal

Abstract

Stereo Matching has been one of the most active research topics in the area of computer vision. Many research works in this field have focused on either improving the accuracy of a resulting disparity map or reducing the processing time in a separate manner. However, there has been relatively little research effort in order to reduce computational complexity or processing time, while improving the accuracy of disparity estimation at the same time. In this thesis, a time efficient approach for real-time stereo matching is proposed. To this end, Dynamic Programming (DP), a popular global energy optimization algorithm, has been chosen to be a framework where the proposed approach is applied, since the DP is able to meet real-time requirements even in a computation-limited environment and thus implies its broad applicability. From extensive experiments, we show that the proposed method can significantly reduce the processing time to the extent that real-time implementation of disparity map estimation can be realized. In addition, a novel modified adaptive support-weight filter that is designed to guarantee reliable estimation of disparity map has also been successfully evaluated in comparison with conventional filters in our experiments. In conclusion, as opposed to most previous works with emphasis on hardware level optimization for the processing time reduction, the proposed approach can provide a time efficient method which is employed at an algorithm level. Further, it can preserve the reliable estimation accuracy by successfully addressing the errors (little increased by the proposed preprocessing time reduction) and even improve the overall estimation accuracy in a resulting disparity map.