A robust underwater Semi-Global Stereo Matching Method using an improved Census Transformation

Abstract Stereo matching algorithms are the core technique in binocular vision, which can acquire 3D information for underwater scenes. However, it's challenging to obtain correct matching results, as the images taken underwater are affected by light absorption, light imbalance, and scattering noise. To optimize the underwater stereo matching, an improved Census Transformation is proposed as a new stereo matching metric to estimate dense disparity maps in the underwater environments and obtain the disparity for target object. First, some improvements using Gaussian filtering are made to introduce a confidence interval to enhance the robustness of the original Census Transformation against noise and light imbalance. Then, the angle metric is calculated using support points, which are obtained from the SURF algorithm. Therefore, the proposed novel method can further improve the ability of the initial cost to cope with light imbalance and underwater noise. Experimental results show that the proposed method is superior to existing algorithms and can better meet the actual needs of underwater experiments.