Research On Refinement Method Of Typical Stereo Matching Algorithms

The research goal of computer stereoscopic vision is to obtain the information of three-dimensional scene from two-dimensional image data.The research object of binocular dense matching algorithm is an epipolar image pairs,and the horizontal pixel coordinate values of the corresponding points are the same;the research goal is to use one of the images as a reference image to find the corresponding point on the other image of each pixel in the reference image.With the development of dense matching technology,the matching algorithm generates three-dimensional information including location,distance,and sometimes can replace the human eye to complete the impossible tasks.It has been widely used in aerospace,robot navigation,medical robots,outer space Detection and other fields.The binocular intensive matching has been widely and deeply studied.Scholars have put forward many algorithms,and among those methods,CSCA(cross-scale cost aggregation),ELAS(efficient large-scale stereo)and SGM(semi-global matching)are all typical ones.CSCA constructs a framework of cross-scale cost aggregation.Under this framework,the effect of any existing cost calculation method can be improved.ELAS can generate disparity map faster than other matching algorithms.With the development of hardware,the operation efficiency of SGM is also paid more attention,and due to the high efficiency and satisfactory accuracy of SGM,it is widely used in commercial products.However,the effectiveness of these classical algorithms can be further improved:Aiming at the lack of smoothing constraint of CSCA algorithm on adjacent cost that is on the same scale,this paper proposes an improved method to fuse this constraint and construct a new objective function.Specifically,based on this assumption,neighboring pixels on the same scale are smoothed with the corresponding smoothing constraints on the cost value in the form of weighted least squares,and a new objective function is obtained.The final disparity value can be obtained by utilizing the WTA(winner-takes-all)strategy on the optimized cost values that has the minimal energy according to the objective function.In order to break the pattern of the fixed status of supporting points set,this article introduces the existing concept of confidence,and selects the point with high confidence as the new support points from the matching result.Pixels with low confidence value will be assigned an new parallax with a relative high confidence value.By calculating the confidence value and the set of support points throughout the iterative process,the calculated disparity value is highly reliable.Aiming at the wide use of SGM in aerial video,a dense matching method suitable for high-resolution aerial image is elaborately designed.Support points are points with high confidence value that provide very valuable information during matching process.In this paper,from the perspective of maximum a posteriori probability,the objective function leveraged by sparse support points is proposed.GMM(Gaussian Mixed Model)is originally used to merge the data items based on the support points into the densely matched objective function.At the same time,a confidence updating method based on the GMM model is further established.Experiments show that the proposed method improves the SGM algorithm Parallax estimation result.