Real-Time Accurate Stereo Matching Using Modified Two-Pass Aggregation and Winner-Take-All Guided Dynamic Programming

Abstract

This paper presents a real-time stereo algorithm that estimates scene depth information with high accuracy. Our algorithm consists of two novel components. First, we apply a modified two-pass aggregation to the adaptive cost aggregation process, use color similarity to calculate support weight, and introduce a credibility estimation mechanism to reduce accuracy loss during two-pass aggregation. Second, we present an amended scan-line optimization technique, which combines winner-take-all and dynamic programming. Our algorithm runs at 20 fps on 320×240 video with a disparity search range of 24. The experimental results are evaluated on the Middlebury benchmark data sets, showing that our method achieves the best reconstruction accuracy among all real-time stereo algorithms.