Stereo Matching Algorithm Based on Two-Phase Adaptive Optimization of AD-Census and Gradient Fusion

Aiming at the bad performances of existing traditional local stereo matching methods in ill-posed regions, an improved AD-Census algorithm based on two-phase adaptive optimization and gradient fusion is proposed. During the cross arm construction phase, an adaptive nonlinear constraint of intensity difference between pixels is adopted to get the optimal arm length. In the cost computation phase, the absolute difference(AD) cost and Census transform(CT) cost of each pixel are weighted summed firstly. Then the result is fused with the gradient cost by adaptive weight, which is determined by the exponential function with the shortest arm length as the independent variable. Finally, the disparity map is obtained by cost aggregation, disparity selection and disparity refinement. The experimental results indicate that the average disparity error of all regions on Middlebury 2014 datasets is reduced by 31% compared with the original AD-Census algorithm, and the average disparity error of non-occlusion regions is reduced by 40%. The proposed algorithm performs better in textureless regions and disparity discontinuity regions, and it shows enough robustness for radiometric changes and noise.