Error assessment of fundamental matrix parameters

Abstract

Stereo image matching comprises of establishing epipolar geometry based on fundamental matrix estimation. Accuracy of the epipoles is governed by the fundamental matrix. A stereo image pair may contain errors on a systematic and/or random basis which determine the accuracy of the fundamental matrix required for matching. The algorithm used to extract the image pair correspondence, and the method used to estimate the correct parameters of the matrix, controls its accuracy further. A performance analysis of widely adopted matrix estimators over the point pairs found by correspondence determiners is undertaken in this chapter. The methods are modified for the best combination of results, based on the properties of the resulting fundamental matrix. Permutations are analyzed over the possible paths for obtaining the matrix parameters with the expected characteristics, followed by error analysis. Amongst the estimators analyzed, RAndom SAmple Consensus (RANSAC) and M-estimator SAmple Consensus (MSAC) estimators were found to produce the best results over the features detected by the Harris–Stephens corner detector.