A novel fast feature based stereo matching algorithm with low invalid matching

Abstract

Computations in a feature based stereo matching which is basically used for depth extraction are generally very high. These computations essentially include feature extraction and matching which feature matching is usually higher. For a feature-based stereo matching, we accurately tune the search space based on some stereo imaging parameters like the focal length with pixels scale, the displacement of features points and maximum disparity. We show that results of previous matches can be used to narrow down the search space to find current match. We use directional derivative of disparity as a temporary concept to tune the search space accurately. Then we develop a fast feature based stereo matching algorithm based on the proposed search space tuning and non-horizontal thinned edge points as features. For reducing the error in the matching stage, we use left-right consistency checking technique for a small number of feature points. Usually this technique doubles the execution time of matching, but we will show that increasing of execution time in the proposed algorithm is negligible respect to the other similar methods as well as the invalid matching is also highly reduced. Comparing to the other similar methods, experimental results for three tested images show that not only the execution time of the matching stage of the proposed algorithm is decreased to 42%, but also the error in the matching is decreased to 90%.