Research on binocular stereo matching algorithm based on dynamic tilt window

Abstract

Aiming at the low precision of the traditional binocular stereo matching algorithm in calculating the matching cost of the strong texture area, a binocular stereo matching algorithm based on a dynamic tilted window is proposed. First, the absolute value of brightness or color difference is replaced by random initialization of pixels, and then the traditional Census cross-domain transformation is replaced by a dynamically tilted disparity plane. For the traditional algorithm, the matching accuracy is improved, making the window more adaptable to the actual environment. In the cost calculation step, a gray histogram is added as an indicator for judging the texture difference, which improves the matching cost in the strong texture area; on this basis, iterates from space propagation, plane propagation to view propagation. In the parallax optimization step, left-right consistency detection and parallax filling are used to further optimize the reduction of the false matching rate. The experimental data is compared with the standard images on the Middlebury dataset. The results show that the average error matching rate of the disparity map generated by the stereo matching algorithm of the dynamic tilted window of this method reaches 4.03%. Compared with the Census algorithm, the matching error rate is respectively reduced. 21.1%, effectively improving the matching accuracy; compared with other algorithms, the false matching rate for high texture areas increased by 1.2% and 3.71%.