Research Of Stereo Matching Technology Based On Disparity Map And Minimum Spanning Tree

Binocular stereo vision technology has always been the focus of machine vision research.Because of its low cost and simple structure,it has been widely used in visual navigation,three-dimensional reconstruction,unmanned driving,industrial detection and other fields.Stereo matching is the core technology of stereo vision.Generally,stereo matching algorithms can be divided into four steps: initial matching cost calculation,cost aggregation,disparity determination,disparity thinning and so on.This paper focuses on how to incorporate disparity information into the cost aggregation phase to obtain more reliable aggregation cost,and how to use linear SCAN-like optimization to optimize the initial aggregation cost in the structure of multi-layer support window and minimum spanning tree.Aiming at the problem that traditional stereo matching local support window is difficult to utilize the information of pixels with long spatial distance and gray distance,this paper improves the support window algorithm,uses disparity and gray information to construct double support window,forms three-layer structure of upper disparity window,middle gray window and bottom pixel,and aggregates matching cost by linear scanning optimization method,and refines disparity iteratively.The test results on the platform of Middlebury show that the proposed algorithm can obtain more accurate disparity map.To solve the problem that the traditional minimum generated set-up volume matching algorithm is difficult to obtain accurate disparity in repetitive texture regions and texture-Free regions,this paper improves the minimum generated set-up volume matching algorithm,aggregates the minimum spanning tree matching cost using linear scan optimization method,and uses adaptive penalty.Penalty value is used to improve the smoothness of the algorithm.Test results on the platform of Middlebury show that the algorithm can obtain clear disparity edges.Aiming at the problem of mismatching caused by using color information only to generate minimum spanning tree,which is easy to generate connections at disparity edges,this paper further improves the minimum spanning volume matching algorithm and generates minimum spanning tree.Parallax information is added to the process,and the minimum spanning tree segmentation algorithm is used to obtain the double-level minimum spanning tree,and the cost of Disparity Map is aggregated by the method of linear scan optimization.In this paper,the mapping relationship between left and right Disparity Map is used to mark the occluded area,and the left and right Disparity Map are refined by cross-iteration.The average mismatch rate of this algorithm is 5.19% on the platform of Middlebury.In order to verify the practicability of the proposed algorithm,this paper uses the calibrated binocular vision system to acquire multiple real scene images,uses the three algorithms to calculate disparity maps,and reconstructs three-dimensional point clouds,fits the measurement of rectangular workpiece size by fitting algorithm,and locates the workpiece position by hand-eye calibration.Experiments show that this algorithm can get clear and accurate disparity map,reconstruct smoother point cloud,and have higher matching accuracy and positioning accuracy.