A robust technique for correspondence using the epipolar geometry

Abstract

The accurate matching problem is an important issue in computer vision. This paper proposes a robust techniques to find correspondence by using a classical method (correlation technique) based on the epipolar constraints and motion flows from uncalibrated stereo images. If we detect matching for only the epipolar geometry, the problem is very difficult. We first use a classical method to accurately find an initial set of matches and then exploit motion vectors in order to remove mismatches among matching candidates. The epipolar geometry can therefore be estimated accurately using well adapted criterion and computing the fundamental matrix. We also adapt to a bucketing technique to optimize the fundamental matrix. More matches are eventually founded by the F-guided matching method underlying the epipolar geometry. The experimental results show the proposed algorithm has been widely tested and works remarkably well in various circumstances with many repetitive patterns. The results also represent that the performance of the proposed algorithm is better than that of conventional methods.