Research On Local Stereo Matching Method Of Binocular Vision

Binocular stereo vision is one of the important research directions in the field of computer vision.It reconstructs the three-dimensional structure of the observed object in the real world by simulating the human eye's perception of stereo space.In recent years,binocular vision has been widely used in areas such as driverless driving,binocular ranging,3D reconstruction,and virtual reality.A complete binocular vision system includes four steps: camera calibration,rectification,stereo matching and 3D reconstruction.Among them,stereo matching is the most important steps in binocular vision research.The matching result directly affects the effect of three-dimensional reconstruction.Local stereo matching not only has low hardware requirements,but also has fast calculation speed and high matching accuracy.Local stereo matching is a hot topic at present.This paper studies the principles,basic steps,algorithm classification and evaluation criteria of stereo matching,and innovatively proposes a series of optimization ideas and methods to solve the problems and difficulties in the current local stereo matching algorithm:(1)To solve the problems of mismatching due to edge depth discontinuity and the ambiguity of weakly textured regions in stereo matching,a cross-scale local stereo matching method based on edge weighting is proposed.In the cost calculation phase,first,the pixel intensity and gradient are combined as the initial cost,and then the pre-processed image is detected by Canny algorithm.An edge similarity measurement method is proposed according to the number and structural information of edge points,and the points satisfying the constraint conditions are weighted by two strategies.In the cost aggregation stage,guided filtering is used for aggregation under the cross-scale aggregation model,which effectively improves the matching effect of weak texture regions.Experiments are performed on 4 sets of standard stereo image pairs and 27 sets of extended stereo image pairs on the Middlebury test platform.Experiment results show that the average mismatch rate of non-occlusion regions is 7.88% without the any refinement steps.The proposed method can effectively improve the matching accuracy.(2)Aiming at the problem that the stereo matching accuracy is susceptible to amplitude distortion and noise interference,an anti-noise matching method based on HSV color space and improved Census is proposed.In the cost calculation phase,firstly,converts the image from RGB color space to HSV color space.Using the hue channel as the matching primitive.Establishing the HAD cost calculation function,which can suppress the effects of amplitude distortion.Secondly,an improved neighborhood weighting Census method is proposed to solve the problem that the traditional Census transform relies on the central pixel excessively,thereby improving the noise immunity of the algorithm.Finally,the HAD cost and the improved Census cost are nonlinearly combined as Initial cost.In the cost aggregation stage,an outlier removal method based on the confidence interval is proposed.The cost values that are not in the confidence interval are eliminated,and the remaining values is filtered and aggregated to reduce noise interference and improve matching accuracy in noisy environments.Experiments prove that the proposed method can not only effectively suppress the influence of noise,but also achieve a more robust matching effect in scenes with inconsistent exposure and illumination conditions.There are 43 figures,13 tables,and 75 references in this paper.