A stereo matching algorithm based on Census transform and improved dynamic programming

Abstract:

A stereo matching algorithm based on Census transform and improved dynamic programming is proposed to the problems of traditional dense stereo matching methods, which have high false matching rate in the textureless areas, depth discontinuities and occlusion. The initial local matching cost is calculated by sparse Census transform correlation, and the raw cost is also optimized by a dynamic programming method by involvoing bidirectional constraints of row and column simultaneously. Meanwhile, the confidence and texture of each pixel are measured for reference image. Finally, the disparities of non-confident or textureless pixels are estimated by fitting parameters of a plane model for the corresponding segment, and the dense disparity map was obtained as well. Experiment results demonstrate that the proposed algorithm achieves high matching accuracy and robustness, especially in the textureless areas, depth discontinuities, and occlusion as well.