Research On Parallax Estimation Algorithm Based On Binocular Vision

Binocular parallax estimation is the process of deriving depth information by comparing the left and right images and calculating the parallax value of each pixel point in the left and right images.Due to its low cost,simple structure,and abundant information acquisition,it is widely used in fields such as 3D reconstruction,autonomous driving,and virtual reality.Stereo matching is the core problem of binocular parallax estimation,the accuracy and speed of matching will directly affect the efficiency and results of parallax estimation,therefore,how to reduce the mismatch rate of traditional algorithms and how to improve the computational speed of deep learning algorithms has become the focus of research.Revolving around commonly used binocular disparity estimation algorithms,the following research work has been performed,as follows:(1)A binocular parallax estimation algorithm based on the fusion of gradient operator and improved Census transform is proposed for the problem of high false matching rate in parallax discontinuity region and weak texture region by traditional algorithms.The greyscale average of all pixel points in the initial matching window is used to instead of the central pixel,the window is adaptively adjusted by the variance of the greyscale values within the initial window,and the cost function of fusion is constructed by constraining the image edges with the absolute value of the grey-scale difference and the gradient operator to reduce the mismatching rate of the algorithm in the parallax discontinuity region and the weak texture region.Comparative experiments on the Middlebury dataset show that the algorithm in this paper reduces the average mismatch rate with the Census transform and the AD-Census transform by 7.95% and 3.00% in the parallax discontinuity,and by 5.80% and2.03% in the full region,respectively.(2)Analyzing the application of convolutional neural network in binocular parallax estimation,and to aim at the problems of its large number of parameters and slow running speed in the operation process a multi-scale feature fusion algorithm based on cavity convolution is proposed in this paper.The algorithm utilizes cavity convolution instead of conventional convolution in the feature extraction part,which can expand the perceptual field without increasing the computational effort;it uses multiscale feature information fusion to enrich contextual information.Cost Body Construction utilizes the idea of group correlation to combine with cascading cost bodies,which can reduce information loss while increasing the correlation between pixels;Experimenting on the KITTI 2015 dataset,this paper’s algorithm reduces the number of parameters by 30.3% and 44.2% relative to PSMNet and Gwc Net without increasing the mismatching rate,while the running time for single image is reduced from 0.41 and 0.32 seconds to 0.19 seconds.(3)A parallax estimation method suitable for images with different focal lengths is proposed for the situation where the focal lengths of binocular cameras are not completely equal during actual experiments.The method uses the results of binocular parallax estimation in a common region to adaptively adjust the parallax values in the monocular field of view.Experiments has been conducted on the Scene Flow dataset and the algorithm reduced the endpoint error EPE from 16.5px to 8.33 px and the mismatch rate from from64.23% to 43.96% compared to direct superposition.