A novel depth estimation approach based on bidirectional matching for stereo vision systems

Recently, the extraction efficiency and quality of depth data have become the major concern in stereo vision. This work presents stereo depth estimation method based on Bidirectional Matching Method (BM ${}^2$) aiming at achieving a good speed-accuracy trade-off. In this study, a novel bidirectional matching process is proposed for matching cost computation. Additionally, a robust cost function is designed for computing color, gradient and local binary pattern similarity between potential matching points. Next, an efficient adaptive spatially smooth edge-preserving filter is applied to minimize matching ambiguities. Following this, a Knock-Out Disparity Selection Technique (KO-DST) that iteratively solves a sequence of discrete binary optimization problem is used to generate disparity map. Finally, a discontinuity preserving disparity refinement is applied to correct disparity errors and refine the resulting disparity map. Rigorous experiments and analysis on Middlebury datasets validated the effectiveness of the presented idea and its superiority over some state-of-the-art algorithms. The proposed method runs approximately two times faster than the standard bidirectional matching.