Design of Binocular Stereo Vision System Via CNN-based Stereo Matching Algorithm

In this paper, we design a binocular stereo vision system based on an adjustable narrow-baseline stereo camera for extracting depth information from a rectified stereo pair. The camera calibration and rectification are performed to get a rectified stereo pair serving as the input to the stereo matching algorithm. This algorithm searches the corresponding points between the left and right images and produces a disparity map that is used to obtain the depths via the triangulation principle. We focus on the first stage of the algorithm and propose a CNN-based approach to calculating the matching cost. Fast and slow networks are presented and trained on standard stereo datasets. The output of either network is regarded as the initial matching cost, followed by a series of post-processing methods for generating qualified disparity maps. The contrast tests have demonstrated that the CNN-based methods outperform census transformation on the mentioned datasets. Finally, we advance two error criteria to acquire the range of system working distance under diverse baseline lengths.